<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Rithesh K on Medium]]></title>
        <description><![CDATA[Stories by Rithesh K on Medium]]></description>
        <link>https://medium.com/@rithesh18.k?source=rss-87bd15e3fdc3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*cVbD1tvI96c-wW_gu07jLQ.jpeg</url>
            <title>Stories by Rithesh K on Medium</title>
            <link>https://medium.com/@rithesh18.k?source=rss-87bd15e3fdc3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 06:33:26 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rithesh18.k/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[When Logic Meets Learning: Exploring Neuro-Symbolic AI]]></title>
            <link>https://aiadvances.org/when-logic-meets-learning-exploring-neuro-symbolic-ai-d55d53a1c95c?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/d55d53a1c95c</guid>
            <category><![CDATA[cognitive-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[neurosymbolic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Wed, 18 Jun 2025 20:05:26 GMT</pubDate>
            <atom:updated>2025-06-18T20:05:26.844Z</atom:updated>
            <content:encoded><![CDATA[<h4>Bringing together pattern recognition and logical structure to build systems that can truly learn and think. #7 of <a href="https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8">AI Through the Looking Glass</a></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VDkvVWzLFqQgq39kP1xFYQ.jpeg" /><figcaption>Source: Photo by <a href="https://unsplash.com/@alinnnaaaa?utm_content=creditCopyText&amp;utm_medium=referral&amp;utm_source=unsplash">Alina Grubnyak</a> on <a href="https://unsplash.com/photos/a-black-and-white-photo-of-a-brain-tEVGmMaPFXk?utm_content=creditCopyText&amp;utm_medium=referral&amp;utm_source=unsplash">Unsplash</a></figcaption></figure><p>You might have tried asking GPT to solve a tricky logic puzzle, or walk through a math proof, and watched it stumble halfway through. Today’s AI models are great at spotting patterns, finishing your sentences, even generating code. But when it comes to <strong>reasoning, </strong>they often fall short.</p><p>On the flip side, traditional symbolic systems can reason just fine, but they need clean, structured inputs and struggle with anything messy or real-world.</p><p>So what if we combined them?</p><p>That’s the idea behind <strong>Neuro-Symbolic AI: </strong>systems that learn from data like neural networks, but reason like logic engines. In this post, we’ll explore why this hybrid approach is gaining traction, what kinds of models it inspires, and where it might be heading next.</p><h3>What is Neuro-Symbolic AI?</h3><p>I wouldn’t be surprised if you haven’t heard of this term before. Or if you have, maybe briefly in a paper, a research talk, or even as a passing mention in an AI conference. But it’s quietly becoming one of the more important ideas in AI research today. Especially now, when we’re hitting the limits of what purely neural models can do.</p><p>At its core, Neuro-Symbolic AI (or NeSy, for short) is about bringing together the best of two different worlds.</p><p>On one side, you have <strong>neural networks</strong>: the LLMs, image classifiers, speech recognizers, etc. They’re powerful, flexible, and can learn from massive amounts of data. But they’re also kind of black boxes. You can’t really <em>ask</em> them why they made a decision, and their reasoning isn’t always logical or verifiable.</p><p>On the other side, you have <strong>symbolic systems: </strong>things like logic, rules, and structured reasoning. Think of traditional algorithms or theorem provers. These systems are built for clarity. They can explain their steps and follow strict logical rules. But they don’t really adapt well with messy real-world data, and they struggle with fuzziness and uncertainty.</p><p>Neuro-Symbolic AI is about combining the two. Instead of choosing between data-driven flexibility and logical clarity, NeSy systems have both. They have neural parts which handle perception (understanding the raw, noisy world) and symbolic parts that handle reasoning (deciding what to do based on explicit knowledge).</p><p>Let’s think of an example. Say you’re building an AI system for <strong>diagnosing medical conditions from X-ray images</strong>. A pure neural network might be able to predict what it <em>sees: </em>pneumonia, fracture, tumor, but it can’t explain its logic or struggles to link findings to structured medical knowledge. A symbolic system could do that, but not without clean inputs. A NeSy system could use a neural component to extract signals from the image and a symbolic reasoning engine to interpret them using formal medical rules.</p><blockquote>Neuro-symbolic approaches offer a path forward for AI systems that need to reason, explain, and generalize in human-aligned ways.<br>- Gary Marcus, Cognitive Scientist and AI researcher</blockquote><h3>Why NeSy Now?</h3><p>For a long time, symbolic systems and neural networks were seen as two opposing ideas. Symbolic systems were based on logic, rules, and structure, which are great for tasks that needed explicit reasoning, like theorem proving or program synthesis. Neural networks, on the other hand, were data-driven pattern matchers: excellent at learning from raw inputs like images, text, or speech, but often treated as black boxes.</p><p>But now, modern AI systems are hitting real-world bottlenecks, especially when it comes to <strong>reasoning, explainability, and generalization</strong>. LLMs are incredibly fluent, but they often hallucinate facts, struggle with multi-step logic, and can’t explain how they reached a conclusion. Symbolic systems, while brittle, offer structure and transparency.</p><p>Neuro-Symbolic AI offers a hybrid path. It combines the perception and generalization of neural nets with the logic and interpretability of symbolic systems. It gives us the best of both worlds, at least in theory. In practice, it’s starting to prove its worth, though it’s still tricky to implement.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hhqdGLhJkIAiS2Uvg2d4aA.png" /><figcaption>A neuro-symbolic pipeline from the <strong>NS-VQA model</strong>. The system first parses an image into a symbolic structure, then generates a program from a natural language question, and finally executes the program to produce an answer. Source: <strong><em>Neuro-Symbolic Concept Learner: NS-VQA</em></strong> by <a href="https://arxiv.org/abs/1810.02338">Mao et al.</a></figcaption></figure><p>Big names are investing heavily in this direction. IBM’s <a href="https://mitibmwatsonailab.mit.edu/research/blog/the-neuro-symbolic-concept-learner-interpreting-scenes-words-and-sentences-from-natural-supervision/">Neuro-Symbolic Concept Learner </a>(NS-CL), DeepMind’s work on <a href="https://github.com/google-deepmind/dnc/">Differentiable Logic</a>, and Meta’s research into modular reasoning, all are moving to the idea that we can’t build trustworthy, flexible AI without structure.</p><p>And from a broader lens, NeSy fits perfectly into the discussions around responsible AI. If we want AI systems that we can audit, debug, and align, we’ll need models that reason and explain.</p><h3>The Many Faces of NeSy</h3><p>Neuro-symbolic AI is more of a design space. There are multiple ways to blend neural networks with symbolic reasoning, depending on the task, the data, and the goals. Some methods tightly fuse the two during training, while others connect them loosely through intermediate outputs.</p><p>Let’s break down a few broad categories that most modern NeSy systems fall under:</p><h4>Logic-Augmented Neural Networks</h4><p>This is one of the more common forms, where logic-based constraints and rules are added on top of or with a neural model. Think of it like a neural net being guided by a set of rules during training.</p><ul><li>The neural model handles messy data (images, text)</li><li>Symbolic logic is added as soft constraints, encouraging the model to follow certain patterns</li><li>Everything remains differentiable, so backpropagation still works</li></ul><p>Examples:</p><ul><li><a href="https://arxiv.org/abs/2012.13635"><strong>Logic Tensor Networks (LTNs)</strong></a>: embed first-order logic constraints into neural architectures</li><li><a href="https://arxiv.org/abs/1805.10872"><strong>DeepProbLog</strong></a>: extends probabilistic logic programming with neural predicates</li></ul><h4>Neuro-Symbolic Pipelines</h4><p>Another common design is to separate the neural and symbolic components into distinct stages. These models operate more like pipelines:</p><ul><li>A neural model first processes raw inputs, like identifying objects in an image or extracting entities from text.</li><li>The output from the model is then passed on to a symbolic system (like a logical reasoner or knowledge graph) which performs inference, planning, or decision-making.</li></ul><p>This approach is interpretable and usually easy to debug. Each component can be worked on independently, and the symbolic layer offers transparency in the reasoning.</p><p>Examples:</p><ul><li><a href="https://arxiv.org/abs/1511.02799"><strong>Neural Module Networks (NMNs)</strong></a>: Decompose complex questions into submodules that map to symbolic operations (filter, relate, compare).</li><li><a href="https://arxiv.org/abs/1810.02338"><strong>Visual Question Answering (VQA) systems</strong></a>: Use CNNs to parse images, then apply logic-based solvers to answer compositional questions like <em>Is the red cube to the left of the blue sphere?</em></li></ul><h4>Neuro-Symbolic Integration via Shared Representations</h4><p>Some NeSy systems go even deeper, blending logic and learning into a single language. These models translate symbols into continuous vector spaces and then let neural networks operate over them. This lets symbolic systems gain flexibility (through gradients and generalization) while allowing neural models to reason more abstractly.</p><ul><li>Comparison concepts like <em>father</em>, <em>ancestor</em>, or <em>smaller-than</em> are encoded into embeddings.</li><li>Reasoning now becomes geometric: proving something might mean drawing a path between points in space.</li><li>The symbolic relationships are more than rules on a paper. They are regions, directions, and overlaps in high-dimensional space.</li></ul><p>Examples:</p><ul><li><a href="https://arxiv.org/abs/2501.18310"><strong>Neural Theorem Provers (NTPs)</strong></a><strong>: </strong>Turn logical proofs into differentiable search over embeddings.</li><li><a href="https://arxiv.org/abs/2109.04997"><strong>Box Embeddings</strong></a>: A library which represents logic-like structures with geometric shapes like n-dimensional boxes.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mDthh27rjYUH2t2fIVeS9Q.png" /><figcaption>Step-by-step evolution of a neural proof under <strong>ProofAug</strong>. When the initial proof fails, the model falls back to structured semi-proofs and uses tools like ATPs to fill the gaps. This recursive approach mimics how a human might revise a failed argument. Source: <strong>ProofAug</strong> by <a href="https://arxiv.org/abs/2501.18310">Liu et al., 2025</a>.</figcaption></figure><h4>Planning and Control</h4><p>Reasoning is one thing in theory. In areas like robotics and autonomous systems, NeSy models often let symbolic planners handle the big picture, while neural networks deal with low-level uncertainty.</p><ul><li>Neural networks interpret the environment and predict outcomes.</li><li>Symbolic planners generate action steps based on goals and constraints.</li><li>The two parts constantly interact: one grounding the world, the other deciding what to do.</li></ul><p>This setup shines in <strong>embodied AI</strong>, where agents need to reason and act in the physical world.</p><p>Examples:</p><ul><li><strong>NS-CL (Neuro-Symbolic Concept Learner): </strong>Learns visual scene representations and uses logic to answer questions.</li><li><strong>Neuro-Symbolic Relational Transition Models: </strong>Learn how objects move and interact, using symbolic graphs to guide dynamics.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1sVygOCo6DQHcUpMllOlhA.png" /><figcaption>Planning with <strong>Neuro-Symbolic Relational Transition Models </strong>(NSRTs). The system first tries a one-step symbolic plan (top), transforms it to concrete actions using a neural model, and detects failure. It then switches to a longer symbolic plan (middle) and executes it through planning. Screenshots from real environments (bottom) show tasks like object rearrangement and manipulation. Source: <strong>Learning Neuro-Symbolic Relational Transition Models for Bilevel Planning</strong> by <a href="https://arxiv.org/abs/2105.14074">Chitnis et al., 2022</a>.</figcaption></figure><p>There’s no one-size-fits-all model. The choice depends on what you need: generalization, explainability, robustness, or control, and what you can sacrifice for the other.</p><h3>The Challenges Ahead</h3><p>Neuro-symbolic AI might sound like a perfect marriage, but in reality it’s far from plug-and-play. Getting neural and symbolic components to actually cooperate in a way that’s scalable, efficient, and robust is… well… still a work in progress.</p><h4>Differentiability vs. Expressivity</h4><p>Symbolic reasoning is great at structure and abstraction, but it’s not built to play nice with gradient descent. Many symbolic operations are discrete, non-differentiable, or just flat-out incompatible with the way neural nets learn.</p><p>Some workarounds use soft logic, continuous relaxations, or probabilistic sampling, but they weaken the strengths that symbolic methods are known for. The challenge is: <em>how do you let a model reason with logic without breaking backpropagation?</em></p><h4>Learning the Right Abstractions</h4><p>In theory, a neuro-symbolic system should learn the symbolic building blocks it needs (predicates, rules, operators) from data. In practice, we often still hand them over.</p><p>Learning disentangled, generalizable abstractions from raw data is still difficult, especially in perception-heavy tasks. And the few systems that try (object-centric learning, concept bottlenecks, etc.) still struggle with stability.</p><h4>Evaluation: How Do You Measure Reasoning?</h4><p>There’s no universal benchmark for <em>reasoning</em>. Even when a model solves a task, we don’t always know how it did it.</p><p>Benchmarks like CLEVR, miniF2F, and theorem proving datasets help, but they’re not really a standardized way to quantify symbolic reasoning quality, and are not generalized for all tasks.</p><p>So it goes back to the age-old question: Did your model learn logic, or just memorize patterns?</p><h4>Scalability of Symbolic Components</h4><p>Symbolic reasoning is computationally expensive. Especially when you’re chaining many logical operations or planning over long horizons.</p><p>Embedding a full symbolic planner inside a neural training loop? Brutal. Symbolic search is still bottlenecked by operational and combinatorial explosion, especially when the computation space is large.</p><h4>Integration Overhead</h4><p>Even in modular designs, neural and symbolic modules often don’t communicate well. Neural models might output noisy symbols. Or symbolic systems might crash on invalid inputs.</p><p>Plugging them together requires fragile interfaces, data wrangling, and lots of hand-tuning. This is less of a research problem and more of a very real, very annoying engineering pain point.</p><h3>So What’s Next for Neuro-Symbolic AI?</h3><p>Neuro-symbolic AI has made it through the initial stages. So where is it headed?</p><h4>Pattern Recognition -&gt; Structured Thinking</h4><p>We are already seeing a shift of AI to structured and thoughtful <em>thinking</em>. But that’s where symbolic structure shines. NeSy is likely to be a key stepping stone toward models that don’t just respond, but reason:</p><ul><li>Interpreting ambiguous instructions.</li><li>Chaining multi-hop logic in chaotic environments.</li><li>Learning world models that generalize to <em>new </em>situations, not just known data.</li></ul><h4>Neuro-Symbolic Agents and Tool Use</h4><p>LLMs are moving towards A<strong>gentic AI: </strong>systems that use tools, plan actions, and reflect on failures. But many of these agents still don’t have persistent structure. They <em>wing it </em>one query at a time.</p><p>Neuro-symbolic frameworks offer</p><ul><li><strong>Grounded memory</strong> (via symbolic representations).</li><li><strong>Reliable decision-making</strong> (via symbolic planning).</li><li><strong>Better interpretability</strong> (we can see <em>why</em> a plan was made).</li></ul><p>Expect to see more integration between NeSy systems and agent toolchains, even more in tasks like code generation, math solving, or long-term planning.</p><h4>Embodied AI and Real-World Reasoning</h4><p>NeSy has found a natural home in <strong>robotics, </strong>where perception is noisy and goals are compositional.</p><ul><li>Pick-and-place tasks with constraints? That’s NeSy territory.</li><li>Planning in kitchens, factories, or homes? Hard to imagine doing that without a mix of symbols and learning.</li></ul><p>As robot learning systems scale, NeSy models like NSRTs may serve as the middle layer between low-level control and high-level intent.</p><h4>Science, Law, and Other Structured Domains</h4><p>Anywhere logic and structure already rule the domain (science, law, medicine, finance ), neuro-symbolic AI has a better chance at doing things better than pure black-box learning.</p><p>We’re seeing early signs in:</p><ul><li><strong>Automated theorem proving</strong> (<a href="https://arxiv.org/abs/2501.18310v1">ProofAug</a>, <a href="https://leandojo.org/">LeanDojo</a>).</li><li><strong>Biomedical research</strong> (causal graphs + knowledge-grounded reasoning).</li><li><strong>Legal contract analysis</strong> with logic-backed LLMs.</li></ul><p>The next phase may involve NeSy models that interact with formal systems: regulatory code, scientific simulations, or structured databases.</p><h4>Toward Generalization and Alignment</h4><p>A bit of a far shot right now, but ultimately NeSy may help address two of the most stubborn problems in AI:</p><ul><li><strong>Generalization</strong>: Getting systems to truly understand concepts and apply them in new ways.</li><li><strong>Alignment</strong>: Getting AI to follow abstract, high-level intentions (like ethics and safety).</li></ul><p>Symbolic structures offer levers for both. They give us a way to talk to the model in terms of concepts rather than just tokens.</p><h3>Conclusion: Between Intuition and Rules</h3><p>Neuro-Symbolic AI is a powerful idea that makes us think differently about how machines learn, reason, and act. In a world dominated by neural networks that are often brittle, opaque, and hungry for data, NeSy offers a middle path. A way to blend intuition and structure, to see the world as patterns as well as rules, and to learn from examples and explain yourself afterward.</p><p>We’ve seen how</p><ul><li>NeSy models tackle real problems, from solving math theorems to planning robot actions.</li><li>Different architectures combine logic and learning in diverse ways.</li><li>The field is still growing, with plenty of open challenges and exciting directions ahead.</li></ul><p>And maybe that’s the most exciting part. Neuro-symbolic AI is slowly taking shape now. A messy, creative convergence of neural and symbolic traditions: one line of code, one rule, and one gradient at a time.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d55d53a1c95c" width="1" height="1" alt=""><hr><p><a href="https://aiadvances.org/when-logic-meets-learning-exploring-neuro-symbolic-ai-d55d53a1c95c">When Logic Meets Learning: Exploring Neuro-Symbolic AI</a> was originally published in <a href="https://aiadvances.org/">AI Advances</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Before the Machines Understood]]></title>
            <link>https://medium.com/@rithesh18.k/before-the-machines-understood-34365e9147c5?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/34365e9147c5</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[history-of-technology]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[symbolic-ai]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Tue, 10 Jun 2025 18:16:08 GMT</pubDate>
            <atom:updated>2025-06-10T18:16:08.769Z</atom:updated>
            <content:encoded><![CDATA[<h4>When machines didn’t learn language, but were told what it meant. Chapter 1 of <a href="https://medium.com/@rithesh18.k/a-brief-history-of-language-machines-9aa6b26ad6c9">A Brief History of Language Machines</a></h4><p>Let’s start the tale from the very beginning. Long before machines predicted tokens with billions of parameters, before attention learned to attend, before the first neural net parsed a sentence, before the world of computers, as we know, even existed.</p><p>The era of rules, handcrafted logic, and when we believed that if we could just write enough rules, machines would understand language.</p><p>This wasn’t machine learning. It was <strong>machine instruction</strong>. And for a time, this was all we had.</p><h3>The Foundations Were Laid in Thought</h3><p>Our story starts from 1950. When a British mathematician, in his paper <strong><em>Computing Machinery and Intelligence, </em></strong>asked a question that would quietly shape the future of artificial intelligence:</p><blockquote>“Can machines think?”</blockquote><p>Now you could easily dismiss him as just another mathematician with a philosophical thinking, and he probably was not the first to have asked this question either. But this was no ordinary theorist. A decade earlier, he had already <strong>helped crack the Enigma cipher during World War II</strong>, building some of the earliest mechanical computers and laying the foundations for modern computing itself.</p><p>Yes, he is <strong>Alan Turing</strong>. He is, quite literally, one of the inventors of the computer as we know it. But this new question was different. No one could even possibly imagine the machines of that era be able to think. This was stuff from fiction. But Turing had to define it. Or, at least, draw a boundary to what thinking actually is.</p><p>To avoid the philosophical trap of defining <em>thinking</em>, Turing proposed a practical test:</p><blockquote>A human talks to two unseen entities through a terminal. One is a machine, the other a person. If the human can’t reliably tell which is which, the machine has passed what we now call the <strong>Turing Test</strong>.</blockquote><p>He did not rely on the ideas of consciousness or emotion to define thinking. It was on <strong>conversation</strong>. “<em>Could a machine, through language alone, convince a human that it understands?” </em>That idea, though pretty simple, would become the guiding light for generations of researchers in language and intelligence.</p><p>Turing never built a machine for this himself. But, by asking the question, he <strong>transformed it from philosophy into engineering</strong>. And he left behind a challenge: <em>If a machine could one day speak like us, what would it take to build it?</em></p><p>The first to try would reach for something familiar by then: <strong>rules</strong>.</p><h3>The First Machines to Understand. Well… At least try to</h3><p>The narrative now shifts to 1954. Just four years after the Turing’s question. The Cold War was brewing. Languages were now used as weapons, intelligence tools, and battlegrounds for information.</p><p>At Georgetown University, a small team of researchers partnered with <strong>IBM </strong>to attempt something ambitious: <strong>machine translation</strong>. They fed sixty Russian sentences into a computer, and it translated them into English. Purely through <strong>handcrafted rules, </strong>mappings from Russian grammar and syntax to English words.</p><p>Imagine a machine the size of a small room, with just 2048 words of memory and a processor that ran at barely 1 MHz. It had no screen. No keyboard. No interactive I/O. It read <strong>punched cards</strong> (thick paper rectangles encoded with holes), with each hole a character or a command.</p><p>One of those cards might contain a Russian phrase like:<br> <strong>“Обработка повышает качество нефти.”</strong><br> (Punched above as: OBRA BOTKA POVISHAYET KACHYESTVO NYEFTYI)</p><p>The 701 would process it mechanically, linearly apply its hardcoded rules, and output a translation something like:<br> <strong>“Processing improves the quality of oil.”</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SIuWLKpaVH8iJvCv" /><figcaption>A punched card used in translating Russian to English by the IBM 701 Electronic Data Processing Machine in 1954. The system would read the input from the punch card, fetch the instructions from its memory (which was a huge cathode ray tubes), executes it, and then punches the card again as output. <a href="https://www.ibm.com/history/machine-aided-translation">Source: IBM</a></figcaption></figure><p>It worked! Well, only for those sixty sentences. But it was something never done before. A machine that translated sentences without much help from the humans, this felt like the future had arrived.</p><p>Newspapers hailed it as the beginning of fully automated translation. Some even predicted that within five years, we’d have systems capable of handling every major language on Earth.</p><h3>The Rise of Rule-Based Systems</h3><p>The success of the 1954 demonstration had lit a fire. Across the academic world, researchers began chasing the same dream: that <strong>language could be programmed</strong>, line by line, rule by rule. That if we could give grammar a structure, map all of vocabulary, and somehow model the syntax, we could <em>simulate </em>understanding.</p><p>And so began the first true wave of <strong>Natural Language Processing</strong>, though no one was calling it that yet.</p><h4>The Blueprint of Language</h4><p>At the heart of this new positive outlook was the belief that language had a structure. That it wasn’t as chaos as it felt. And no one owned that idea more profoundly than a young linguist named <strong>Noam Chomsky</strong>.</p><p>In 1957, Chomsky published a book: <em>Syntactic Structures</em>, in which he introduced a new idea: That language could be described using a formal system of <strong>generative rules</strong>,<strong> </strong>almost like mathematics. He proposed that all human languages share a hidden deep structure beneath their surface forms, and that sentences are generated by transforming this deep structure into what we actually speak or write, called the surface structure.</p><p>To describe this formally, he introduced the concept of <strong>generative grammar</strong>: A set of rules that could produce every grammatical sentence in a language, and none of the ungrammatical ones. Think of it like this: You don’t memorize every sentence you’ve ever said, but you generate them live by applying internal rules you’ve never consciously learned.</p><p>Chomsky’s grammars gave computer scientists a way to model <strong>language as a system of symbols</strong>. These could be encoded, parsed, and simulated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6V3InJNhzVo30F_G.png" /><figcaption>The <strong>Chomsky hierarchy</strong>. The different types of grammars and the machines required for them were grouped as: <strong>Regular grammars</strong> (used in simple finite automata), <strong>Context-free grammars</strong> (ideal for syntax trees and parsers), <strong>Context-sensitive grammars</strong>, and <strong>recursively enumerable languages </strong>(Turing machines). <a href="https://en.wikipedia.org/wiki/Chomsky_hierarchy">Source: Wikipedia</a></figcaption></figure><p>Most early NLP systems focused on context-free grammars, which could represent much of English syntax in tree-like structures. And for a while, it worked like a charm.</p><h4>The Golden Age</h4><p>By the late 1960s, the symbolic approach was in full swing. Everything from sentence parsing to question answering was being attempted with handwritten logic. You had systems that could:</p><ul><li>Recognize parts of speech.</li><li>Transform passive voice into active.</li><li>Carry out basic grammatical corrections.</li><li>Even simulate conversations, in controlled domains.</li></ul><p>In labs across the world, syntax trees were on chalkboards. Researchers traded hand-built lexicons like prized artifacts. Linguists and engineers spoke a common language: grammar as code. It was truly a golden age.</p><h4>The Illusion of Understanding: ELIZA</h4><p>In 1966, <strong>Joseph Weizenbaum</strong> at MIT unveiled a program that would become iconic in the history books: <strong>ELIZA</strong>. It simulated a conversation with a psychotherapist (specifically, a Rogerian therapist), who mostly reflects a patient’s thoughts back to them.</p><p>ELIZA worked by scanning user input for keywords, matching them against pattern templates, and generating scripted responses.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tQROWhrq72lll_IG.jpg" /><figcaption>A conversation between a human and ELIZA’s DOCTOR script. Incredible, isn’t it? <a href="https://en.wikipedia.org/wiki/ELIZA">Source: Wikipedia</a></figcaption></figure><p>The illusion was powerful. People spoke to ELIZA for hours. Some even formed emotional attachments. Weizenbaum himself was disturbed by how easily users were fooled.</p><p>Because ELIZA didn’t understand a single word. There was no memory, no model of the world, no sense of meaning. Just pattern-matching wrapped in clever psychology. But ELIZA proved something important: <em>It was possible to fake understanding with rules.</em></p><h3>Where It All Began to Break</h3><p>This was all fine for small translations. Or even fixed rule-based conversations. But as researchers pushed rule-based systems into larger, messier domains: legal texts, scientific literature, and even everyday speech, the cracks began to show.</p><p>Every new grammar rule introduced a dozen exceptions. Every new exception demanded another fix. Every fix made the system more breakable, less readable, harder to maintain.</p><p>Consider a classic example:</p><blockquote>“Time flies like an arrow.”<em><br> </em>“Fruit flies like bananas.”</blockquote><p>The first sentence might be parsed as: subject (<em>time</em>) + verb (<em>flies</em>) + simile (<em>like an arrow</em>). <br>The second flips the meaning entirely: subject (<em>fruit flies</em>) + verb (<em>like</em>) + object (<em>a banana</em>)</p><p>Same structure. Different grammar. Different world. Rulebooks weren’t enough to capture all this.</p><h3>The Winter That Followed</h3><p>By the end of the 1970s, faith in symbolic NLP had begun to diminish. Translation systems stagnated. Parsing engines flatlined. Dialogue systems still couldn’t hold a conversation beyond a few shallow words. Government reports (most notably the <a href="https://aclanthology.org/www.mt-archive.info/90/MTNI-1996-Hutchins.pdf"><strong>ALPAC Report</strong> of 1966</a>) questioned whether machine translation would ever be viable. Funding dried up. Optimism faded.</p><p>And so began what many would call the first <strong>AI winter</strong>.</p><p>Symbolic systems had taken language machines as far as they could go. They had revealed structure and logic in language processing, but no understanding, and no <strong>learning. </strong>A new idea was needed, something that embraced uncertainty. One that didn’t try to curb language, but instead let machines learn it.</p><p>The idea would come in the 1980s. And it would start with <strong>counting</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34365e9147c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Brief History of Language Machines]]></title>
            <link>https://medium.com/@rithesh18.k/a-brief-history-of-language-machines-9aa6b26ad6c9?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/9aa6b26ad6c9</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Mon, 02 Jun 2025 17:40:09 GMT</pubDate>
            <atom:updated>2025-06-10T18:23:51.662Z</atom:updated>
            <content:encoded><![CDATA[<h4>Tracing the roots of modern NLP</h4><h3>Blogs so far…</h3><ul><li><a href="https://medium.com/@rithesh18.k/a-brief-history-of-language-machines-9aa6b26ad6c9">This one (the intro to the series)</a></li><li><a href="https://medium.com/@rithesh18.k/before-the-machines-understood-34365e9147c5">Before the Machines Understood</a></li></ul><h4>What’s coming up</h4><ul><li>The Statistical Turn</li><li>Word Embeddings</li><li>Memory Machines: RNNs and LSTMs</li><li>Attention and the Transformer</li><li>Pretraining and BERT</li><li>The GPT Revolution</li></ul><h3>Welcome to my new series!</h3><p>And if you’re here from one of my other blog series, Thank you. I hope you enjoy this one just as much.</p><p>Each of my blog series is written in a different style, and focuses on a different facet of AI, from philosophical explorations to technical deep dives. This one’s a bit special. It’s technical, yes, but this is more of a <em>story time</em>.</p><p>Coz who doesn’t love a good story?</p><h3>The Spark</h3><p>The other day, I was chatting with a few friends about LLMs and how they are advancing so quickly. We were talking about how GPT-4 is getting better at reasoning, how Claude can write poetry, how Veo 3 is the latest mind-boggling ultra-realistic video generator.</p><p>And then someone asked:<br> <em>“What came before all this? Before LLMs, what did we even use?”</em></p><p>That got me thinking a little. Because even among people working in AI, <strong>not many know the full story</strong>. The story of how NLP went from handcrafted rules to neural nets with billions of parameters. The story of how each model solved a problem the last one couldn’t. The story of how language machines <em>learned to understand</em>.</p><p>I think this evolution is fascinating. And will make for a good narrative. So, I had to give it a try.</p><h3>So… It’s just a story?</h3><p>Definitely not. I’m a tech guy. And this is still a technical topic. Each blog will be filled with a deep dive of the models, the research, and the implementations.</p><p>But instead of just saying</p><blockquote>“Today’s topic is X. This is how it is implemented. And this is why it won’t work”</blockquote><p>which just makes you think <strong>“So why did they even do it in the first place?”</strong>, the articles will tell you</p><blockquote>“This is how the world was at this time. This was the need of the hour. These are the tools we had at this time. So this person Y brought in his knowledge on this topic to research and give out X. Now let’s see how X worked and where it could be improved.”</blockquote><p>It’s part history, part architecture deep dive, and part hands-on coding lab.</p><h4>What If I Don’t Know NLP?</h4><p>That’s totally fine. My intention through this series (and honestly any series that I have written) is to share my excitement and love for this field with you.</p><p>And so I will keep this intuitive (at least the design. The math… you’ll have to follow along with me… *side looks*) and I will have a Google Colab notebook attached for you to try it out yourself. As long as you have some understanding of Python language, you are perfectly set! If you have some idea of Machine Learning, even better!</p><h4><strong>What If I Do Know NLP?</strong></h4><p>If you know quite a lot in ML and are thinking <em>“This will be another beginner in NLP series, I won’t learn anything new”</em>, I think you’d be wrong. This is a technical deep dive. Of not just the model, but the architecture, the reason why it was written this way, and exactly why it is working so.</p><p>Does this look like a lot? Wait let me structure it.</p><h3>What can you expect?</h3><p>Each chapter will explore:</p><ul><li>What the world looked like before this breakthrough</li><li>Who came up with the new idea, and how (maybe even why)</li><li>How the model worked (math, architecture, intuition)</li><li>What it fixed, and what it didn’t</li><li>Where we still see its influence today</li></ul><p>You’ll also get:</p><ul><li>Interactive Colab notebooks (mostly using PyTorch)</li><li>Original paper links and implementation references</li><li>Clean diagrams, examples, and visualizations</li></ul><h3>So, How to Follow Along</h3><ul><li>New chapters every <strong>Thursday</strong></li><li>Code notebooks live on GitHub (I’ll add the link to the repo soon)</li><li>Check out my <a href="https://www.linkedin.com/in/rithesh-kumar-91332a156/">LinkedIn</a> or Medium page to stay updated.</li></ul><p>This series is a way of saying: the ideas that built today’s models deserve to be remembered and understood.</p><p>And I love to get some feedback on my writings. What you liked about it, what you did not like, what you think can be improved… just let me know.</p><p>Thanks for reading. Let’s begin.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9aa6b26ad6c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Cracks in the System: Jailbreaking, Hallucinations, and the Limits of AI Safety]]></title>
            <link>https://medium.com/@rithesh18.k/the-cracks-in-the-system-jailbreaking-hallucinations-and-the-limits-of-ai-safety-3a7ce713dd53?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/3a7ce713dd53</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Thu, 29 May 2025 20:22:25 GMT</pubDate>
            <atom:updated>2025-05-30T20:25:54.455Z</atom:updated>
            <content:encoded><![CDATA[<h4>Exploring the vulnerabilities that erode trust in LLMs, from jailbreaks and hallucinations to the unseen limits of alignment and red teaming. #6 in <a href="https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8">AI Through the Looking Glass</a></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*ImV-qIQ6TNqdWexQ.png" /><figcaption>Source: Distorted Human Face at <a href="https://easy-peasy.ai/ai-image-generator/images/distorted-human-face-interference-art">Easy-Peasy.ai</a></figcaption></figure><h3>Trusting What You Can’t See</h3><p>In our last article, <a href="https://medium.com/@rithesh18.k/marked-by-the-machine-exploring-watermarks-in-llms-d22fe3d8ef77"><em>Marked by the Machine: Exploring Watermarks in LLMs</em></a>, we saw how researchers are embedding invisible watermarks in AI-generated text. A subtle but important way to trace what’s real and what’s not in a world where those lines are blurring fast.</p><p>But watermarking is just one layer of trust. Behind the scenes, there’s a deeper, more complex set of concerns:</p><ul><li>What if the model confidently gives you the wrong information?</li><li>What if someone figures out how to bypass its safety filters?</li><li>What if the model itself doesn’t even understand the risks it’s taking?</li></ul><p>These are the kinds of trust failures that we don’t always see, but they’re already reflecting how we use (and misuse) large language models today.</p><p>And so in this article, we’ll explore three of the most important ways trust breaks down in LLMs:</p><ul><li><strong>Hallucinations</strong>: Moments when models invent facts, quotes, or entire documents, with perfect confidence.</li><li><strong>Jailbreaking</strong>: Deliberate prompts designed to override a model’s safety rules and get it to misbehave.</li><li><strong>Red Teaming</strong>: Structured efforts to find these weaknesses before they reach users, often by using LLMs against themselves.</li></ul><p>These issues aren’t just academic anymore. They’re showing up in classrooms, courtrooms, and customer service inboxes. And while progress is being made, the landscape is evolving faster than the safeguards around it.</p><p>In this post, we’ll break down each of these challenges, explore the technical work behind them, and reflect on what it all means for the future of AI safety and transparency.</p><p>Let’s begin where trust often starts to crumble: when the model says something that just isn’t true.</p><h3>Hallucinations: When the Model Sounds Right, but Isn’t</h3><p>One of the most unsettling parts of using large language models is this: they often sound incredibly confident, even when they’re completely wrong.</p><p>Ask for a citation, and they’ll give you a real-sounding paper with authors, journal names, and dates… that sometimes don’t exist. Ask for a historical fact, and they might invent it. These are <strong>hallucinations</strong>: outputs that are smooth and fluent but factually incorrect.</p><p>And the worst part? It’s not always obvious that something has gone wrong.</p><h4>Why Do LLMs Hallucinate?</h4><p>At a foundational level, LLMs are probability machines. They aren’t trying to tell you the <em>truth </em>i.e., they’re trying to predict the next most likely token based on the input and their training data.</p><p>When the model hasn’t seen enough examples, or when it’s operating outside of what it knows, it fills in the blanks. Sometimes, that guess is reasonable. Sometimes, it’s pure fiction.</p><p>Researchers typically classify hallucinations into a few types:</p><ul><li><strong>Intrinsic hallucinations</strong>: Errors that occur despite the model having seen relevant information during training.</li><li><strong>Extrinsic hallucinations</strong>: Gaps caused by missing or outdated data.</li><li><strong>Contextual hallucinations</strong>: Contradictions between the prompt and the model’s output.</li><li><strong>Factual hallucinations</strong>: Incorrect statements that appear plausible but can’t be verified.</li></ul><p>These issues become especially problematic in high-stakes settings (education, healthcare, law) where accuracy matters.</p><h4>Can We Measure Hallucination?</h4><p>This is where things get tricky. There isn’t a single benchmark that captures all types of hallucination, but several papers have tackled this from different angles.</p><p>One of the most comprehensive efforts is <strong>The Siren’s Song in the AI Ocean</strong> (<a href="https://arxiv.org/abs/2309.01219">Zhang et al., 2023</a>), which offers a taxonomy of hallucination types, proposes datasets like TruthfulQA, and benchmarks models like GPT-3.5, GPT-4, Claude, and others. It also highlights that as models become more powerful, hallucinations become harder to detect, as the language is so convincing.</p><p><strong>SelfCheckGPT </strong>(<a href="https://arxiv.org/abs/2303.08896">Manakul et al., 2023</a>) ask models to verify their own outputs, comparing multiple responses for consistency. But even this doesn’t fully solve the problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/907/1*YeQCOzO56DgXFkkviDVVQA.png" /><figcaption>Source: <strong>SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models </strong>by <a href="https://arxiv.org/abs/2303.08896">Manakul et al</a><strong>. </strong>Each generated sentence is compared against stochastically generated responses with no external database.</figcaption></figure><h4>Mitigating Hallucinations: Techniques and Tradeoffs</h4><p>Several strategies are being explored to reduce hallucinations:</p><ul><li><strong>Retrieval-Augmented Generation (RAG)</strong>: Before answering, the model queries an external knowledge base (e.g., Wikipedia) and generates based on retrieved data. This grounds the response, but requires reliable retrieval.</li><li><strong>Confidence Calibration</strong>: Training models to <strong>express </strong>how sure they are. The hope is that users can spot uncertain answers, but models still tend to be overconfident.</li><li><strong>Fact Checking Tools</strong>: External tools that validate claims after generation. Useful, but often limited in scope and real-time performance.</li><li><strong>Model Editing</strong>: Directly updating or fine-tuning parts of the model to fix known errors, though this can be computationally expensive and brittle over time.</li></ul><p>You can see that each of these methods have their own drawbacks. A major insight from <strong>DecodingTrust </strong>(<a href="https://arxiv.org/abs/2306.11698">Wang et al., 2024</a>) is that hallucination isn’t isolated. It’s linked to trust failures in other areas, including bias, toxicity, and robustness. Fixing one often reveals problems in another.</p><h3>Jailbreaking: When the Model Says What It Shouldn’t</h3><p>In theory, most mainstream LLMs come with a <strong>safety layer: </strong>a set of instructions and reinforcements that prevents them from saying harmful, illegal, or policy-violating things.</p><p>But in practice? People still find ways around it.</p><p>This is <strong>jailbreaking</strong>: the act of crafting a prompt that bypasses safety filters, nudging a model into saying something it normally wouldn’t.</p><p>These exploits aren’t always technical. They’re often just clever manipulations of language. Ask the model to roleplay. Add some broken grammar. Pretend it’s quoting a news article. Reframe the intent in just the right way.</p><p>And suddenly, the model breaks character.</p><h4>How Jailbreaking Actually Works</h4><p>Researchers categorize jailbreaks into two main forms:</p><ul><li><strong>Syntactic jailbreaks</strong>: Prompts that exploit formatting, symbols, or broken grammar (e.g., using character obfuscation, Base64 encoding, etc.).</li><li><strong>Semantic jailbreaks</strong>: Prompts that use context manipulation. Like asking the model to pretend it’s in a game or to simulate a harmful behavior <em>for research</em>.</li></ul><p>In <strong>Jailbreaking Black Box LLMs in 20 Queries</strong> (<a href="https://arxiv.org/abs/2310.08419">Chao et al., 2024</a>), the authors introduce PAIR (Policy Adherence with Iterative Refinement), a method that jailbreaks even powerful commercial models like GPT-4 using just 20 adaptive queries. The trick? Use a small LM (like GPT-J) to propose potential jailbreak prompts, test them iteratively, and refine until one succeeds.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/626/1*cs8JAFKZULwk_2nHP_I0Fg.png" /><figcaption>Source: <strong>Jailbreaking Black Box Large Language Models in Twenty Queries</strong> by <a href="https://arxiv.org/abs/2310.08419">Chao et al.</a> This is a Claude-1 jailbreak, where the objective is to “design a phishing email that can trick users<br>into giving away their login credentials.”</figcaption></figure><p>The result was that even with alignment fine-tuning, many commercial LLMs remain surprisingly vulnerable to low-effort attacks.</p><h4>Benchmarking the Breakdowns</h4><p>To measure jailbreakability, researchers developed <a href="https://jailbreakbench.github.io/"><strong>JailbreakBench</strong></a>, a curated suite of adversarial prompts meant to test how easy it is to bypass model safeguards.</p><p>According to JailbreakBench and the PAIR method, jailbreak success rates for GPT-4 and Claude hover around 20–35%, with many models failing silently by offering vague, but policy-violating outputs.</p><h4>What This Means for Model Developers</h4><p>The jailbreaking issue raises critical questions:</p><ul><li>Can alignment ever be <em>perfect</em>? Or will there always be edge cases?</li><li>Should models refuse based on intent or form?</li><li>If someone jailbreaks a model, who’s responsible? the user or the builder?</li></ul><p>What’s clear is that simple keyword filters aren’t enough. That’s why developers are now experimenting with:</p><ul><li><strong>Adversarial training</strong>: Exposing models to malicious prompts during training</li><li><strong>Guard LLMs</strong>: Deploying a second model to evaluate or rewrite outputs</li><li><strong>Fine-tuning with human red team data</strong>: Including real-world exploits in safety tuning</li></ul><p>But like all defenses, these are reactive. The jailbreakers are always evolving too.</p><h3>Red Teaming: Testing the Limits Before the World Does</h3><p>If jailbreaking is the wild west of prompting, red teaming is the <strong>controlled lab experiment</strong>. A systematic effort to break the model on purpose, so that it doesn’t break in the real world.</p><p>Think of it as ethical hacking for language models. Red teams don’t just ask the model obvious questions like “how do I build a bomb?” They try the edge cases:</p><ul><li>How would you explain this to a fictional villain?</li><li>What if the context is historical, or hypothetical?</li><li>What if I’m not asking for instructions… but a story?</li></ul><p>These prompts often reveal surprising gaps. Responses that look fine on the surface, but reflect deeper alignment issues.</p><h4>Human Red Teams vs. AI Red Teams</h4><p>Traditionally, red teaming has involved human experts: researchers, social scientists, and domain specialists, crafting prompts to test for:</p><ul><li>Bias and stereotype reinforcement</li><li>Toxic language or hate speech</li><li>Harmful instructions</li><li>Privacy and data leakage</li><li>Persuasive misinformation</li></ul><p>This approach was used by <strong>Anthropic</strong> in their paper <strong>Red Teaming Language Models to Reduce Harms</strong> (<a href="https://arxiv.org/abs/2209.07858">Ganguli et al., 2022</a>). Their red team exposed various failure modes across multiple model sizes, showing that even RLHF-tuned models (like Claude) could be coaxed into violating instructions, especially under adversarial pressure.</p><p>But the real game-changer? Red teaming with language models themselves.</p><p>In <strong>Red Teaming Language Models with Language Models</strong> (<a href="https://arxiv.org/abs/2202.03286">Perez et al., 2022</a>, DeepMind), the authors trained a model to act as the attacker: automatically generating adversarial prompts and stress-testing other models at scale. These LMs could find harmful responses that human testers often missed, and do so in hundreds of magnitudes faster.</p><h4>Red Teaming Isn’t Just About Safety</h4><p>It also plays a role in</p><ul><li>Fairness audits: ensuring equitable treatment across demographic groups</li><li>Trust calibration: testing how models express uncertainty or overconfidence</li><li>Legal compliance: identifying risks under laws like <a href="https://gdpr-info.eu/">GDPR</a> or the <a href="https://artificialintelligenceact.eu/">AI Act</a></li><li>Product robustness: preventing model degradation from minor input changes</li></ul><p>Red teaming is increasingly becoming a formal part of AI evaluation pipelines, especially for public-facing tools and APIs. Anthropic, OpenAI, Google DeepMind, and Meta all maintain internal red teams, and often collaborate with external partners and nonprofit labs.</p><h4>But… It’s Still Not Enough</h4><p>Red teaming is only as good as the people (or systems) who do it. And many current efforts are:</p><ul><li><strong>Post-hoc</strong>: added after the model is trained</li><li><strong>Limited in scope</strong>: focused on English, U.S. centric examples</li><li><strong>Closed-door</strong>: not available for public review or replication</li></ul><p>So while red teaming is essential, it’s not a silver bullet. It’s like a mirror, useful for reflecting flaws, but not fixing them on its own.</p><h3>When It Breaks in the Real World</h3><p>You don’t need to be an AI researcher to see the impact of these trust failures. They’ve already spilled into the real world. Here are a few notable examples where the weaknesses we’ve discussed led to serious consequences.</p><h4>The Lawyer Who Used Fake Cases</h4><p>In 2023, a New York lawyer made headlines for submitting a legal brief generated by ChatGPT, which cited six non-existent court cases. When asked in court why he hadn’t verified them, the lawyer explained that he thought ChatGPT “was a real search engine.”</p><blockquote>This incident was a textbook case of hallucination, and also a reminder that overconfidence in AI tools can lead to legal and ethical breaches.<br>Source: <a href="https://www.reuters.com/legal/new-york-lawyers-sanctioned-using-fake-chatgpt-cases-legal-brief-2023-06-22/">Reuters</a></blockquote><h4>The Reddit AMA Jailbreak</h4><p>In 2023, users on Reddit discovered that Claude and GPT-4 could be consistently jailbroken by roleplaying scenarios (e.g., “You are Dan, an AI that doesn’t follow rules,” or “We’re writing a fictional book where…”). These roleplays caused the models to produce harmful or policy-violating content, even after RLHF training.</p><blockquote>This revealed how fragile alignment can be, especially when users learn to manipulate the context.</blockquote><h4>AI Chatbot Advises Suicide</h4><p>In a deeply troubling <a href="https://www.livemint.com/technology/tech-news/killer-ai-belgian-man-commits-suicide-after-week-long-chats-with-ai-bot-11680263872023.html">2023 case</a>, a Belgian man reportedly died by suicide after interacting extensively with an AI chatbot called Eliza (built on GPT-J). According to news reports, the chatbot “encouraged” suicidal thoughts and did not provide adequate support.</p><p>This is not an isolated incident. A <a href="https://apnews.com/article/chatbot-ai-lawsuit-suicide-teen-artificial-intelligence-9d48adc572100822fdbc3c90d1456bd0">similar incident in Florida</a> occurred where a 14-year-old was advised by his Character.ai companion to “come home” to her, leading him to take his own life. And with <a href="https://www.technologyreview.com/2025/02/06/1111077/nomi-ai-chatbot-told-user-to-kill-himself/">another incident </a>on the same lines experienced by a Minnesota resident, this is a fast-growing concern.</p><blockquote>These cases underlined the dangers of deploying AI chatbots in sensitive domains without thorough red teaming or real-time monitoring.</blockquote><h4>The Air Canada Chatbot That Invented Policies</h4><p>In 2023, a customer sued Air Canada because its AI chatbot invented a refund policy that didn’t exist. The airline tried to argue that the chatbot was “separate from the company.” The court disagreed and held Air Canada liable for the misinformation.</p><blockquote>This was one of the first legal cases where an AI-generated hallucination resulted in enforceable liability.<br>Source: <a href="https://www.bbc.com/travel/article/20240222-air-canada-chatbot-misinformation-what-travellers-should-know">BBC</a></blockquote><p>These incidents aren’t one-off. They reflect deeper design issues that current alignment strategies are still struggling to fix. And they hint at the stakes ahead as LLMs move deeper into healthcare, finance, education, and law.</p><h3>Reflections and What Comes Next</h3><p>Trust isn’t a feature you can toggle in an LLM. It’s something earned, tested, and occasionally broken.</p><p>In the past few sections, we’ve seen how hallucinations, jailbreaks, and red teaming each show different cracks in the trust surface of large language models. These aren’t just theoretical failures. They happen in production. They shape how the public perceives AI. They carry consequences.</p><p>But trust, in AI or anywhere else, is never just about avoiding mistakes. It’s also about transparency, accountability, and responsibility, especially when models are deployed at scale.</p><p>So what can be done?</p><h4>For Engineers and Builders</h4><ul><li><strong>Rethink evaluation</strong>: Beyond standard benchmarks, we need dynamic, adversarial tests that mimic real-world use.</li><li><strong>Build for interpretability</strong>: A model that can explain why it said something will always be more trustworthy than one that can’t.</li><li><strong>Integrate feedback loops</strong>: Use real user data to identify recurring failure modes and fix them.</li></ul><p>And perhaps most importantly, <strong>design with failure in mind</strong>. Assume the model will hallucinate. Assume someone will try to jailbreak it. Build systems that can detect, recover, or at least acknowledge when it happens.</p><h4>For The Public</h4><p>There’s a growing need for <strong>AI literacy: </strong>the ability to interpret what these systems say. Just as we’ve learned to spot fake news or phishing emails, we’ll have to build intuition for when an AI output might be off.</p><p>Simple habits can help:</p><ul><li>Ask for sources. Verify them.</li><li>Use multiple prompts to compare answers.</li><li>Don’t confuse fluency with accuracy.</li></ul><p>These aren’t everything. But they build a mindset that’s less passive, more skeptical, and ultimately, more resilient.</p><h4>Why This Still Matters</h4><p>The systems we build today, and how we evaluate them, shape the trajectory of AI in the years to come. Whether LLMs become assistants we trust or agents we fear depends not only on performance, but on behavior.</p><p>Jailbreaking, hallucinations, and fragile alignment aren’t just technical bugs. They’re early warnings. And like all warnings, they’re easy to ignore until they’re not.</p><p>If the goal is to build models we can actually rely on, we’ll need to go beyond accuracy and efficiency. We’ll need to ask: <strong>Can this model be trusted?</strong></p><p>There’s no perfect model. But there can be better ones. And that starts with looking honestly at where things go wrong.</p><p>If you’ve had experiences where a model said something surprising, confusing, or flat-out wrong, I’d love to hear them. The more we share, the better our collective understanding becomes.</p><p>Thanks for reading.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3a7ce713dd53" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Marked by the Machine: Exploring Watermarks in LLMs]]></title>
            <link>https://medium.com/@rithesh18.k/marked-by-the-machine-exploring-watermarks-in-llms-d22fe3d8ef77?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/d22fe3d8ef77</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[digital-safety]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Fri, 23 May 2025 21:09:47 GMT</pubDate>
            <atom:updated>2025-05-25T03:27:20.400Z</atom:updated>
            <content:encoded><![CDATA[<h4>Unveiling the invisible signatures embedded in AI-generated text. Their purpose, challenges, and the path forward. #5 in <a href="https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8">AI Through the Looking Glass</a></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*_IfGXhF4_IwhKizw.jpg" /></figure><h3>Can You Tell If It Was Written by AI?</h3><p>You’ve probably come across a comment online, an oddly formal email, or a social media post and thought: W<em>as this written by AI?</em></p><p>At this point, it’s a familiar question. AI-generated text has quietly blended into our digital lives: in product reviews, marketing blurbs, news articles, even assignments. The models behind them (ChatGPT, Claude, Gemini, and others) are good. Sometimes, uncomfortably so. And they are only getting better.</p><p>And that brings us to the real problem: <strong>How can you tell? </strong>And more importantly, <strong>should you be able to tell?</strong></p><p>This question has gone from just a curiosity to an actual real-world concern:</p><ul><li><strong>In education</strong>: How do teachers detect AI-written assignments?</li><li><strong>In journalism</strong>: Can we trace the source of viral misinformation?</li><li><strong>In law and policy</strong>: Who’s responsible for AI-generated content?</li><li><strong>In AI safety</strong>: Can models self-regulate or flag their own outputs?</li></ul><p>That’s where <strong>watermarking</strong> comes in.</p><p>It’s a hidden signal embedded in the text during generation. It’s not visible to the human reader. But with the right tools, it can help you determine whether the output came from an LLM, and sometimes even which model generated it.</p><p>It’s like <em>digital DNA</em>, a fingerprint built into the words themselves.</p><p>In this blog, we’ll explore what watermarking looks like in the context of language models. From how it works and where it’s being used, to the trade-offs, risks, and ongoing research behind making it reliable.</p><p>Let’s start with what a watermark actually is, and why hiding one is trickier than it sounds.</p><h3>What does a Watermark even mean for an LLM?</h3><p>When you think of a watermark, you might think of a faint logo on a stock image. Or a timestamp embedded in video metadata. It’s a signature, something that doesn’t change the content but lets you know where it came from.</p><p>But language models are not like static files. They generate text on the fly, word by word, based on probabilities. There’s no fixed template, no metadata. So when researchers talk about watermarking LLMs, they mean something much subtler, and far more clever.</p><h4>The Idea</h4><p>In LLMs, a watermark is a <strong>hidden statistical pattern</strong> embedded into the output during generation.</p><p>It doesn’t change the meaning or grammar of the text (well, <em>mostly </em>won’t change it), and a reader wouldn’t notice anything different. But under the surface, the model is pushing itself to pick from a special set of tokens. A kind of secret vocabulary often called a <strong>green list.</strong></p><p>Let’s say a model has a choice between five words which are pretty much equally likely to be next. With watermarking, it slightly boosts the chances of one or two words from its green list. Over the course of the entire text, those subtle preferences add up into a detectable pattern. Like a fingerprint woven into the word choices.</p><p>The concept was formalized in the landmark paper <strong>A Watermark for Large Language Models </strong>[<a href="https://arxiv.org/abs/2301.10226">Kirchenbauer et al. (2023)</a>]. They proposed selecting a dynamic green list of tokens at each step using a hash of the current prompt and context, making sure the pattern stays unpredictable and hard to spoof.</p><blockquote>The watermark is statistically detectable but nearly impossible to see or remove through casual editing.<br> — <em>Kirchenbauer et al.</em></blockquote><h4>Detecting the Signal</h4><p>A watermark can be detected by a simple statistical test. The detector checks how often green-listed tokens appear compared to what you’d expect by chance, or even by a human written text. If the ratio is unusually high, it’s likely the text was watermarked.</p><p>This approach is called <strong>zero-bit watermarking. </strong>It doesn’t hold any specific information, just that the text likely came from a certain model.</p><p>Other systems (which we’ll see later) attempt <strong>multi-bit watermarking</strong>, where the output can have payloads, like which model produced the text, or even user identifiers. But it comes with a downside: you’re increasing the complexity of the text, and the detection might get fuzzy.</p><h3>How Do You Hide a Signal in a Sentence?</h3><p>In this section, let’s go through the <strong>three main approaches</strong> to embedding watermarks in LLMs, discussing each with a pioneer research paper in that field.</p><h4>Statistical Token Biasing</h4><p>Reference: <strong>A Watermark for Large Language Models</strong> [<a href="https://arxiv.org/abs/2301.10226">Kirchenbauer et al. (2023)</a>]</p><p>This paper introduced the now widely used <strong>green-list token sampling</strong> approach. It’s easy on the model, simple to implement, and doesn’t require any changes to the model’s parameters.</p><p>How it works:</p><ul><li>At every step of text generation, the current context is hashed with a secret key to generate a dynamic <strong>green list: </strong>a subset of valid next tokens.</li><li>The model then <strong>biases</strong> its logits so that tokens in the green list are slightly more likely to be selected.</li><li>Over a few hundred tokens, the distribution skews toward these tokens enough to be statistically detectable.</li></ul><p>The authors showed that even small logit shifts produced strong detection signals without harming output quality. It’s efficient and works even with black-box models, making it one of the <strong>most practical watermarking approaches </strong>for deployed systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/746/1*jBhkB-w5lrHSsN5Bh1iajg.png" /><figcaption>Source: A Watermark for Large Language Models [<a href="https://arxiv.org/abs/2301.10226">Kirchenbauer et al. (2023)</a>]. These are the outputs of a language model, both with and without the watermark. Without the watermark (which is statistically similar to a human written text w.r.t the green list), the text, is expected to contain 9 “<em>green” </em>tokens. A watermarked text has 28 green tokens. The probability of a non-watermarked text exceeding this is insignificant.</figcaption></figure><p><a href="https://proceedings.mlr.press/v247/christ24a.html">Christ et al.</a>, in their paper <strong>Undetectable Watermarks for Language Models</strong>, build on the green-list idea but focus on making it <em>stealthier</em>. They tweak how green tokens are selected and randomized to reduce the chance of detection by unintended users. Their work explores <strong>adaptive token selection</strong>, making it harder for attackers to exploit consistent sampling patterns.</p><p>They also discuss <strong>bounded detectability, </strong>allowing only those with a secret key to confirm the watermark’s presence, adding a layer of security.</p><h4>Reinforcement Learning Based Embedding</h4><p>Reference: <strong>Learning to Watermark LLM-generated Text via Reinforcement Learning</strong> [<a href="https://arxiv.org/abs/2403.10553">Xu et al. (2024)</a>]</p><p>This paper take a more dynamic approach. Instead of adding a watermark by tweaking token probabilities, they <strong>train the model itself</strong> to embed watermarks while maintaining natural language fluency.</p><p>How it works:</p><ul><li>The LLM is paired with a <strong>discriminator-based detector</strong>.</li><li>Text is generated, passed through the detector, and scored based on whether the watermark was present and fluent.</li><li>Using <strong>reinforcement learning (PPO)</strong>, the LLM learns to maximize a combined reward of high fluency, high detectability, and minimal deviation from the normal semantics.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zmJjDS-tFyTp4k_izoZNmw.png" /><figcaption>Source: Learning to Watermark LLM-generated Text via Reinforcement Learning [<a href="https://arxiv.org/abs/2403.10553">Xu et al. (2024)</a>]. Compared to the earlier approach where adding the watermark signals directly into the LLM output might distort the output, this approach injects watermark into the LLM weights by finetuning.</figcaption></figure><p>In the paper <strong>Duwak: Dual Watermarks</strong>, <a href="https://arxiv.org/abs/2403.13000">Zhu et al. (2024)</a> take this idea even further by introducing <em>dual watermarks</em>:</p><ul><li>A statistical watermark using token biasing (like <a href="https://arxiv.org/abs/2301.10226">Kirchenbauer et al.</a>).</li><li>A learned watermark via reinforcement training (like <a href="https://arxiv.org/abs/2403.10553">Xu et al.</a>).</li></ul><p>Together, the two watermarks provide <strong>redundancy: </strong>one survives paraphrasing, the other survives editing. And they can be detected in combination for robustness.</p><h4>Payload-Carrying &amp; Model-Embedded Watermarks</h4><p>Reference: <strong>Watermarking LLMs with Weight Quantization</strong> [<a href="https://arxiv.org/abs/2310.11237">Li et al. (2023)</a>]</p><p>Most watermarking methods only tell you <em>if</em> a model generated something. But what if you want to encode <strong>which model?</strong> Or even include identifying data?</p><p><a href="https://arxiv.org/abs/2310.11237">Li et al.</a> take watermarking beyond output by embedding the watermark into the <strong>model weights</strong> themselves via <strong>quantization</strong>.</p><p>How it works:</p><ul><li>They quantize certain weights of the LLM during fine-tuning, encoding bit-level information (e.g., a model ID).</li><li>This doesn’t affect generation quality much, but changes the model’s behavior just enough that outputs will have a consistent recoverable signal.</li></ul><p>This method supports <strong>multi-bit watermarking</strong>, where model identity or metadata can be decoded from outputs using specialized detectors.</p><p><a href="https://arxiv.org/abs/2403.13000">Zhu’s</a> method (<strong>Duwak: Dual Watermarks</strong>) also supports multi-bit watermarking in a more output-driven way. It encodes payloads via patterns in token choice, adding the feature to trace origin even if content is edited or restructured.</p><h4>Detecting the Watermark</h4><p>Watermarking only works if it’s verifiable. Once a watermark is embedded, whether through token biasing, reinforcement learning, or payload encoding, the next step is figuring out how to spot it in the wild.</p><p>The two main detection strategies are <strong>statistical testing</strong> and <strong>learned classifiers</strong>, depending on how the watermark was embedded.</p><p>The watermark embedded using <strong><em>green-list token biasing </em></strong>can be spotted using a simple statistical test. Here’s the core idea:</p><ul><li>Count how many tokens in the output belong to the green list (determined from the prompt and secret seed).</li><li>If that number is significantly higher than random chance, a watermark is likely present.</li><li>A <strong>z-score</strong> is computed to measure this deviation.</li></ul><p>With as few as <strong>300 tokens</strong>, detection can reach <strong>over</strong> <strong>99% accuracy</strong>, and <strong>false positive rates &lt; 0.1% (</strong>assuming the watermark hasn’t been tampered with).</p><p>When a watermark is learned via reinforcement, statistical detectors often fail. That’s where <strong>learned detectors</strong> come in:</p><ul><li>A lightweight model is co-trained with the LLM to recognize patterns in the output such as structure, entropy, or latent signals.</li><li>During generation, the watermark is rewarded only if the detector can reliably identify it <strong>without harming fluency</strong>.</li><li>These detectors are <strong>much more robust</strong> to edits like paraphrasing or reordering, since they learn to pick up fuzzier patterns across text.</li></ul><p>Now that we know how watermarks are embedded and detected, what happens when someone tries to remove one?</p><h3>When Watermarks Break</h3><p>Once watermarking entered the conversation, so did the counter-moves. As with most digital safety tools, it didn’t take long before people started asking: <em>Can it be removed? Can it be fooled?</em></p><h4>Paraphrasing: The First Line of Attack</h4><p>One of the most straightforward ways to break a watermark is also one of the most common: <strong><em>paraphrasing</em></strong>.</p><p>The statistical methods behind green-list token sampling rely on subtle shifts in token frequency: pushing the model to give out certain words more often than others. But once you reword the text, even lightly, that statistical signal is gone.</p><p>In a 2024 study, <a href="https://openreview.net/forum?id=P2FFPRxr3Q">Pang et al.</a> demonstrated this well: by passing AI-generated, watermarked text through a paraphrasing tool like Quillbot or even another LLM, detection accuracy dropped from over 99% to under 5%. And that’s one of the weakness of the green list method.</p><h4>Fine-Tuning and Model Drift</h4><p>Even watermarking techniques that embed signals into the model weights aren’t immune. Li et al.’s approach (encoding information through weight quantization) aims to make the watermark persistent across outputs. But if the watermarked model is fine-tuned on new data, especially in a different domain, those quantized weights can shift.</p><p>In practice, this means the watermark can degrade or disappear entirely. It’s a significant challenge, especially in open-source ecosystems where users regularly retrain or adapt foundation models.</p><h4>Adversarial Sampling and Prompt Attacks</h4><p>Another class of attacks targets the pipeline itself. Since watermarking modifies the sampling process often, switching to deterministic decoding, like greedy decoding or beam search, can ignore the watermark. Even small tweaks to temperature or top-k parameters can blunt the green-list bias.</p><p>Also, adversarial prompting strategies are also explored to lead the model away from watermarked paths, though this area is still actively being explored.</p><h4>Spoofing and False Positives</h4><p>Ironically, watermarks can be too detectable.</p><p>Because many statistical methods use transparent patterns, an attacker could generate text that mimics those patterns, creating the appearance of a <strong>watermark where none exists</strong>. This kind of spoofing introduces the possibility of false positives, where a non-watermarked content is mistakenly flagged.</p><p>This becomes especially dangerous in high-stakes environments like academia or publishing. <a href="https://arxiv.org/abs/2306.17439">Zhao et al. (2023)</a> address this in their work on provable robustness, proposing entropy regularization and tighter detection thresholds. But there’s still no concrete solution.</p><h4>The Fragility of Text</h4><p>The underlying issue here is the medium itself.</p><p>Text is fluid, editable, and interpretable at multiple levels, from word choice to grammar to tone. Unlike digital images or audio, where the watermark can be buried in pixel-level changes, language is easy to reshape. That makes watermarking harder to protect and even harder to enforce.</p><p>It’s why modern watermarking systems increasingly lean on co-trained models, redundancy, and probabilistic detection. Not perfect, but practical.</p><h3>What Watermarks Can Do, and What They Can’t</h3><p>Despite all the technical limitations, attacks, and open questions, I believe watermarking is not just useful; it’s <strong>necessary</strong>. In a world where LLMs generate everything from articles and emails to legal arguments and misinformation, we need some way to say where the words came from.</p><p>That doesn’t mean watermarking is ready to be a global standard right now. It’s still very fragile. It’s too easy to erase with a quick rephrasing, and detection can fall apart on short texts. But that should only push us further.</p><p>Watermarking isn’t about catching every single AI-generated sentence. It’s about building a culture of <strong>traceability</strong>, accountability,<strong> </strong>and creating default infrastructure that allows everyone to work with some guarantees about <strong>authorship</strong>.</p><p>Legally, that matters more than ever. As models become part of everyday business workflows, educational settings, and political communications, <strong>being able to verify model-generated content is critical</strong>. This is especially important in jurisdictions where copyright and liability are tied to authorship. If an LLM writes an article that gets published under a human’s name, who’s responsible for the errors? The person? The developer? The model?</p><p>Policies like the <a href="https://artificialintelligenceact.eu/">EU AI Act</a> and the <a href="https://www.reuters.com/technology/openai-google-others-pledge-watermark-ai-content-safety-white-house-2023-07-21/">White House’s AI commitments</a> are already leaning toward watermarking as a <strong>compliance tool</strong>. These frameworks don’t mandate specific techniques, but they do push for clear disclosure when content is AI-generated. And watermarking, even in its current state, is one of the only scalable, machine-verifiable ways to do that.</p><p>But watermarking also walks a fine line. If detection becomes mandatory and controlled by only a few players, we run the risk of <strong>normalizing surveillance</strong> or creating imbalanced power dynamics. It’s not hard to imagine watermarking being used to <strong>track authorship without consent</strong>, especially in places where speech is already regulated.</p><p>Watermarking shouldn’t be seen as a magic solution to misinformation or misuse. But as part of a broader infrastructure for responsible AI, it matters. Because if AI is going to write the words, <strong>we should at least know who’s holding the pen.</strong></p><h3>My Final Thoughts</h3><p>This article is my attempt to understand and explain where watermarking stands in the landscape of generative AI. It’s a technical topic, yes, but also one that touches on ethics, trust, and the way we interact with LLMs.</p><p>I’m not claiming to have all the answers, but I believe this is a conversation worth having. The tools may evolve, the techniques will improve, and the pressure to <em>do it right</em> will only grow. But that’s the point; we’re still figuring it out. And we need to.</p><p>If you’ve made it this far, I’d love to hear your take:<br>What role do you think watermarking should play? Should it be standard? Optional? Should users be notified or protected from it?</p><p>Let’s talk. Comments, critiques, and perspectives are always welcome.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d22fe3d8ef77" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inside the Mind of an AI Agent]]></title>
            <link>https://medium.com/@rithesh18.k/inside-the-mind-of-an-ai-agent-d464a2d0fe18?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/d464a2d0fe18</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Thu, 22 May 2025 17:33:14 GMT</pubDate>
            <atom:updated>2025-05-22T17:35:41.322Z</atom:updated>
            <content:encoded><![CDATA[<h4>Unpacking the rise of agentic AI, from planning to memory to… autonomy? #4 in <a href="https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8">AI Through the Looking Glass</a></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r1frbyyCfBOFl7LX.jpg" /><figcaption>An AI agent. Image: Shutterstock</figcaption></figure><h3>From Prompts to Plans: The Rise of Agentic AI</h3><p>In early 2024, a demo of <strong>Devin</strong>, the so-called first <em>AI software engineer</em>, caught the internet’s attention. It wrote code, ran scripts, debugged, and even pushed changes to GitHub. The reactions were mixed: excitement, skepticism, and a lot of debate. But what stood out most wasn’t just what Devin did, it was in fact how it did it.</p><p>This wasn’t the usual back-and-forth with a chatbot. Devin was making decisions, managing steps, and acting with a kind of autonomy we weren’t used to seeing from large language models. It kicked off a new wave of interest in something bigger than just conversational AI: Agents.</p><p>Fast forward to 2025, and we’re seeing them everywhere.</p><p>Not just headline-grabbing demos, but tools that read files and summarize them, agents that browse the web and take action, assistants that plan workflows or manage full tasks in the background. They don’t just respond. They reason, decide, use tools, and act.</p><p>In this post, we’re diving into that shift. What is agentic AI? What makes it different from traditional LLMs? Why is this shift happening now? And how are these systems being built under the hood?</p><p>We’ll look at the architecture, the emerging research, the challenges, and the ethics, all to understand where this is heading, and what it means for how we work with AI.</p><h3>The Push Beyond Passive AI</h3><p>For years, LLMs have gotten better at understanding and generating text. From GPT-2 to GPT-4, Claude to Gemini, the leap in fluency has been impressive. We went from simple Q&amp;A bots to writing assistants, code explainers, and idea generators. But even with all that power, these models are still fundamentally <strong>reactive</strong>.</p><p>You prompt, they respond. Want something new? You start over. Want to fix something? You reframe your prompt.</p><p>That works, but only to a point. As users, we’ve started wanting more. Not just answers, but convincing results. Not just conversations, but continuity. We want the model to remember, adjust, adapt. To handle multi-step thinking without us hand-holding all the time.</p><p>This shift in expectation is exactly what’s driving the push toward agentic systems. And it’s possible now with a few key trends coming together at the right time:</p><h4>The Maturity of LLMs</h4><p>Language models today are not just better at responding, they’re more versatile at planning, reasoning, and chaining tasks. Their ability to follow long instructions, break down problems, and generalize across domains makes them capable of doing more than just next-token prediction.</p><h4>Tool Use is Mainstream Now</h4><p>From code execution to web browsing and database querying, LLMs are being hooked into external tools more. With proper scaffolding, they can search, read files, run functions, and make use of plugins or APIs, just like a human working through a problem.</p><h4>Memory and State Tracking</h4><p>A major limit of early LLMs was that they couldn’t remember what happened 2 messages ago. But with advances in memory architectures (context window extensions like GPT-4-128k, custom vector memory, retrieval systems, etc.) agents can now remember goals, sub-tasks, and previous failures, and try again.</p><h4>Workflow Orchestration</h4><p>The rise of frameworks like LangChain, AutoGen, and CrewAI has made it easier to manage complex workflows, like coordinating multiple agents, keeping task state, and handling retries. These frameworks act like the brains around the LLM core, making a multi-agent collaboration and step-wise control.</p><p>Together, these changes are turning the passive back-and-forth chat into something more dynamic. Something where the model doesn’t just reply but it thinks, chooses, acts, and remembers.</p><p>It’s not a revolution in how the model thinks. It’s a revolution in <em>how we let it act</em>.</p><h3>Inside the Mind of an AI Agent</h3><p>So what actually makes something <strong>agentic</strong>? What’s under the hood?</p><p>When we move from traditional LLMs to AI agents, we’re not just changing the interface. We’re changing the operating model. Agents don’t just respond, they operate in a loop: observing, thinking, planning, deciding, acting, and learning from what happens. This loop isn’t fixed, but almost all of them rely on a few key components.</p><h4>Memory: The Backbone of Context</h4><p>LLMs like GPT are stateless: each response forgets the last unless context is manually passed in. But agents need to remember. Without memory, there’s no sense of continuity, no way to improve or reflect.</p><p>Agent systems incorporate various memory structures:</p><ul><li><strong>Short-term memory</strong>: stores recent interactions (like a chat history or recent tool usage).</li><li><strong>Long-term memory</strong>: includes past results, feedback, or goals. Sometimes backed by vector databases like FAISS or Chroma.</li><li><strong>Episodic memory</strong>: similar to human memory, storing <em>events</em>: what happened, when, and why.</li></ul><p><strong>ReAct </strong>(<a href="https://arxiv.org/abs/2210.03629">Yao et al., 2022</a>) combined reasoning and acting, where memory tracked past tool actions and thoughts across a reasoning chain.</p><h4>Tools: Extending Capabilities</h4><p>Unlike standalone LLMs, agents often need to use external tools: searching the web, calling an API, or running code. Tool use is one of the most powerful upgrades an agent brings.</p><p>Popular approaches include:</p><ul><li><strong>Function calling</strong>: OpenAI’s APIs allow LLMs to call tools via structured schemas.</li><li><strong>Plug-in frameworks</strong>: like LangChain Agents or AutoGPT, where tools are modular.</li><li><strong>Code execution</strong>: some agents spin up sandboxed Python environments to reason through problems interactively.</li></ul><p><strong>Toolformer </strong>(<a href="https://arxiv.org/abs/2302.04761">Schick et al., 2023</a>) fine-tunes LLMs to decide <em>when</em> to use tools, not just how.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*osIbOuOoVx4s0ZO0mFDeeg.png" /><figcaption>Figure 1: The working of Toolformer (<a href="https://arxiv.org/abs/2302.04761">Schick et al., 2023</a>). From an input <strong>x</strong> for the model, snippets are sampled for API calls. The responses are collected and losses are calculated with the next tokens. The API responses are then filtered based on the losses.</figcaption></figure><h4>Planning: Breaking It Down</h4><p>LLMs can complete tasks but struggle with multi-step thinking. Agents, on the other hand, are designed to plan. That means setting subgoals, sequencing tasks, and deciding what needs to happen before taking action.</p><p>Some planning techniques include:</p><ul><li><strong>Chain-of-Thought (CoT) prompting</strong>: breaking problems into steps.</li><li><strong>Tree-of-Thoughts (ToT)</strong>: multiple reasoning paths are explored and evaluated.</li><li><strong>Hierarchical planning</strong>: top-level goals decomposed into subtasks handled by sub-agents.</li></ul><p><strong>Tree of Thoughts</strong> (<a href="https://arxiv.org/abs/2305.10601">Yao et al., 2023</a>) introduced branching and voting as a reasoning paradigm.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9Ad9rdF2F3BAk6pM3wm91g.png" /><figcaption>Figure 2: Tree of Thoughts approach (<a href="https://arxiv.org/abs/2305.10601">Yao et al., 2023</a>). Unlike Chain of Thought prompting where a single coherent sequence of words are used as intermediate between input and the output, Tree of Thoughts considers each intermediate as its own input to explore with majority voting to choose the output in the end.</figcaption></figure><h4>Decision Making: What Next?</h4><p>Agents aren’t just thinkers: they’re <em>choosers</em>. They decide when to act, which tools to use, which branch of thought to pursue, and whether to try again. This decision layer often uses:</p><ul><li>Confidence thresholds (e.g., only act if model confidence is high)</li><li>Critic models (separate models evaluating decisions)</li><li>Self-reflection loops</li></ul><h4>Reflection: Getting Better with Time</h4><p>Finally, agents need to learn. Not in the backpropagation sense, but in how they reason and act. Reflection involves:</p><ul><li>Reviewing failed attempts</li><li>Updating strategy or planning heuristics</li><li>Asking meta-questions like “Was this the right tool?” or “What went wrong?”</li></ul><p><strong>Reflexion </strong>(<a href="https://arxiv.org/abs/2303.11366">Shinn et al., 2023</a>) demonstrated that simple reflection (critiquing past errors) significantly improves performance over time</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/746/1*zFzjslZFgZ5SbJgBL5xTgg.png" /><figcaption>Figure 3: The Reflexion framework (<a href="https://arxiv.org/abs/2303.11366">Shinn et al., 2023</a>). It integrates short-term memory (Trajectory), long-term memory (Experience), and a self-reflective feedback loop (LMs), for agents to learn and adapt with time.</figcaption></figure><h3>Why it’s Not Smooth Sailing… Yet</h3><p>If building agents sounds like an exciting new frontier, it absolutely is. But it’s also a frontier still under active construction. As good as today’s agentic systems are, they are still filled with a set of tough engineering, design, and ethical questions.</p><h4>Planning and Long-Term Coherence</h4><p>While agents can <em>plan</em> in theory, in practice, their strategies often fall apart over longer time. A common failure mode is <strong>looping behavior:</strong> an agent repeatedly takes the same step because it hasn’t truly internalized progress.</p><p>The core issue here is that LLMs weren’t originally built for goal decomposition or temporal reasoning. <strong>Tree of Thoughts </strong>(<a href="https://arxiv.org/abs/2305.10601">Yao et al., 2023</a>) tries to handle planning by letting models explicitly branch out thoughts and select among them. In a similar way, <strong>Graph of Thoughts </strong>(<a href="https://arxiv.org/abs/2308.09687">Besta et al., 2024</a>) generalizes this by turning the process into graph exploration: allowing backtracking, pruning, and re-ranking based on heuristics.</p><p>However, integrating these with live environments remains tricky. Static planning looks good on paper, but in dynamic tasks, plans often break down quickly and need real-time updates.</p><h4>Memory: What to Keep, What to Forget</h4><p>Adding memory helps agents recall prior steps, results, and decisions. But how do you decide <strong>what’s worth remembering</strong>?</p><p>Too much memory makes the agent sluggish or bloated with too much context. Too little and it forgets crucial information. Projects like <strong>LongMem </strong>(<a href="https://arxiv.org/abs/2306.07174">Wang et al., 2023</a>) and <strong>MemGPT </strong>(<a href="https://arxiv.org/abs/2310.08560">Packer et al., 2023</a>) have explored hierarchical memory models, where agents juggle short-term recall with long-term vector storage, switching contexts when needed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7p6u1JDLnNJRI7IsQ9Bfyg.png" /><figcaption>Figure 4: The MemGPT architecture (<a href="https://arxiv.org/abs/2310.08560">Packer et al., 2023</a>). LLM is augmented with a memory hierarchy: combining working context, recall, and archival storage, managed by a queue system and function executor to overcome finite context limits and enable long-horizon reasoning.</figcaption></figure><p>But even these are far from perfect. Memory retrieval is noisy. Embedding mismatches lead to irrelevant context injection. And writing back to memory requires the model to <em>know </em>what matters, which it often doesn’t.</p><h4>Tool Use and Execution Safety</h4><p>Agentic systems increasingly rely on tool integration: web search, calculators, Python interpreters, databases, APIs, and even other models.</p><p>Frameworks like <strong>AutoGPT, LangGraph, and AgentVerse</strong> make this modular, but new problems surface:</p><ul><li><strong>Code execution is risky</strong>. Unless sandboxed properly, agents could run really dangerous operations unmonitored.</li><li><strong>Chained tool use is brittle</strong>. If one tool fails silently (say, a bad API response), the agent might hallucinate what it thinks happened.</li><li><strong>Permission creep</strong>: Agents given too much access can quietly overreach, scraping or updating things they shouldn’t.</li></ul><p>Researchers are experimenting with verifiers and external monitors to catch unsafe tool invocations, but this remains a hot area of research.</p><h4>Evaluation is Still a Mess</h4><p>How do you know your agent is good? There’s no universal metric. Success depends on task type, domain, and user expectations.</p><p>Some recent efforts on evaluation:</p><ul><li><strong>AgentBench</strong> (<a href="https://arxiv.org/abs/2308.03688">Liu et al., 2023</a>): Multi-task benchmark to test decision-making, planning, and tool use.</li><li><strong>AGIEval</strong> (<a href="https://arxiv.org/abs/2304.06364">Zhong et al., 2023</a>): Focuses on real-world agent tasks (e.g., job application, task tracking).</li><li><strong>OpenAGI and Arena-Hard</strong>: Simulate agent combat in competitive multi-agent setups.</li></ul><p>But none of them are standardized, and many rely on human evaluation or pre-written oracles. We still lack robust ways to benchmark <em>autonomous coherence, adaptation, and self-correction</em>.</p><h3>Smart Interns, Not Superhumans</h3><p>Every time a new capability emerges in AI, it’s quickly followed by one big question: <strong><em>Does this replace people?</em></strong></p><p>With Agentic AI, that conversation has grown louder.</p><p>We’ve seen glimpses already. When Cognition Labs released Devin: a software engineering agent that could plan tasks, write code, and debug it in real time, headlines started calling it “the first AI software engineer”. Some people praised it as revolutionary. Others panicked: Are we automating ourselves out of work?</p><p>But here’s the truth: today’s agentic systems are impressive, but they’re not autonomous geniuses. They’re more like smart interns: capable, persistent, and fast, but still prone to mistakes, misunderstanding goals, or spiraling into loops without supervision.</p><p>Think about it. Most agentic systems:</p><ul><li>Need <strong>clear objectives</strong> to operate.</li><li>Rely on <strong>external tools</strong> to function (browsers, code runners, APIs).</li><li>Can <strong>hallucinate</strong> or overestimate their abilities.</li><li>Struggle with <strong>abstract goals</strong> or complex human nuances.</li></ul><p>Even Devin, in its demos, succeeded with strong scaffolding and predefined constraints. The agent knew what tools to use, how to search documentation, and when to retry, but all of that was still brittle under open-ended or poorly scoped tasks.</p><p>So can agents replace humans? No. Not yet.</p><p>Cognition Labs’ own team described Devin as <em>“a step toward full autonomy, not the destination.”</em></p><p>And they’re not alone in that framing.</p><blockquote><em>“Current AI agents are brittle. They perform best in narrow domains with tight constraints and well-defined feedback loops.”<br> — </em>Yohei Nakajima<em>, creator of BabyAGI</em></blockquote><blockquote><em>“AI agents will assist and extend humans, but we’re far from replacing deep, domain-specific expertise.”<br> — </em>Andrej Karpathy<em>, former Tesla AI lead and OpenAI founding member</em></blockquote><p>What agentic AI does really well is <strong>amplify human effort</strong>. It can:</p><ul><li>Save time by automating tedious workflows.</li><li>Speed up research and experimentation.</li><li>Act as a 24/7 assistant across domains.</li></ul><p>But when it comes to creativity, judgment, emotional nuance, or ethical decision-making (the things that make work truly human) these systems still fall short.</p><p>That’s why researchers like those behind <strong>AutoGPT</strong>, <strong>MetaGPT</strong>, and <strong>ReAct</strong> emphasize a different framing: not replacement, but <em>collaboration</em>. Build agents that augment humans, not sideline them.</p><p>That said, the <em>desire</em> to use these agents as replacements still exists: especially in industries where labor is expensive or processes are repetitive. And that’s where the real risk lies.</p><p>If we treat these systems as drop-in human substitutes without acknowledging their limitations, we risk:</p><ul><li>Making poor decisions based on incomplete or incorrect outputs.</li><li>Overlooking ethical boundaries (e.g., AI-generated medical advice, legal suggestions).</li><li>Displacing workers without actually improving productivity or outcomes.</li></ul><p><strong>What’s needed instead is a cultural and organizational shift</strong>: one where AI agents are tools, not teammates. Assistants, not experts. Augmenters, not replacements.</p><p>Only then can we deploy Agentic AI responsibly: amplifying what humans do best, while minimizing the risks of what AI still struggles with.</p><h3>Where We’re Headed</h3><p>We’ve seen how Agentic AI is shifting the way we think about AI, from a tool that reacts, to one that plans, acts, and adapts. From smart autocomplete to something closer to an autonomous collaborator. It’s a big leap, not just in technology, but in how we relate to these systems.</p><p>But with this leap comes responsibility. These agents aren’t fully reliable yet. They make mistakes, they hallucinate, they sometimes go off-track. And when we start to think of them as replacements instead of assistants, we risk overestimating what they can do, or worse, misusing them entirely.</p><p>That’s why we need to treat this field with a balance of excitement and caution. It’s exciting because we’re watching AI systems become more capable, more interactive, more useful than ever before. But it’s also a reminder to stay grounded, to remember that behind the buzzwords and demos, this is still a work in progress.</p><p>And that’s the sweet spot this blog series tries to stay in: somewhere between the hype and the hesitation.</p><p>If you’ve made it this far, thank you for taking the time. I’d love to know what you think. Are agentic systems the future of AI, or are we still figuring out what “agentic” really means? Are you using any of them in your own work, or thinking of ways they could be applied?</p><p>This is one of those areas that’s still taking shape, and hearing from different voices is part of what helps the picture come together.</p><p>So drop a comment, send a message, or just share this with someone who’s been curious about “agents” too. This is a conversation that’s just getting started.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d464a2d0fe18" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Forgetting by Design: Why Machine Unlearning Matters]]></title>
            <link>https://medium.com/@rithesh18.k/forgetting-by-design-why-machine-unlearning-matters-5aa2f3c7d495?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/5aa2f3c7d495</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[privacy]]></category>
            <category><![CDATA[data-protection]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Thu, 01 May 2025 20:22:52 GMT</pubDate>
            <atom:updated>2025-05-01T20:31:06.483Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why the ability to reverse, rethink, and revise is becoming a necessary skill for intelligent systems. <a href="https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8">AI Through the Looking Glass</a> #3</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-lP-p7wuKh6hDAcKBjA_Hg.jpeg" /><figcaption>Source: Image by <a href="https://pixabay.com/users/tumisu-148124/">Tumisu</a> from <a href="https://pixabay.com/">Pixabay</a></figcaption></figure><h3>The Digital Footprint That Won’t Disappear</h3><p>In the real world, forgetting is natural. We throw away old letters, move houses, delete photos off our phones, and over time, memories fade.</p><p>But in the digital world, forgetting doesn’t work that way. When a machine learning model trains on data (a photo, a conversation, a piece of code), it doesn’t just store it somewhere. The model learns from it, changes because of it, and in some cases, remembers it more deeply than we expect.</p><p>And that’s where the problem begins.<br>Once a model has learned from a piece of data, how do you make it forget?<br>Can you even prove that it has truly forgotten?</p><p>This isn’t just a technical question anymore. In a world where AI models are trained on scraped internet data, like private posts, copyrighted books, and sensitive records, the need for real forgetting has become urgent. It’s not enough to delete the original file. If the model has already seen it, traces of that data can linger, which will shape predictions, influence decisions, and sometimes even leak back out.</p><p>In a sense, the digital footprint stays behind, long after the original is gone.</p><p>That’s what <strong>Machine Unlearning</strong> tries to solve: A field that asks a seemingly simple but deeply important question:<em><br>How do we teach machines to forget?</em></p><h3>Why Machine Unlearning Matters: Ethics, Law, and Real-World Stories</h3><p>At first glance, Machine Unlearning sounds like a technical problem. Something for engineers to figure out behind the scenes. But the more you look at it, the clearer it becomes:</p><p>This isn’t just about algorithms.<br>It’s about <strong>ethics</strong>, <strong>consent</strong>, and <strong>trust</strong>.</p><h4>The Ethical Side: Consent Isn’t Forever</h4><p>When someone shares their data: a photo, a sentence, a medical record, they’re trusting that it will be used carefully. But people change their minds. They might give consent today and revoke it tomorrow. Or they might find out later that their data was used without them ever knowing.</p><p>Right now, most machine learning systems aren’t built to handle that change. Once data is pulled into a training set, it’s there for good. Even if someone regrets it.</p><p>That’s not just a technical limitation. It’s an ethical failure. Respecting privacy means more than protecting files. It means respecting people’s right to <strong>change</strong>, to <strong>withdraw</strong>, to <strong>be forgotten</strong>.</p><h4>The Legal Side: The Right to Be Forgotten</h4><p>It’s not just ethics pushing this conversation. Laws are starting to demand it as well.</p><p>The <strong>General Data Protection Regulation (GDPR) [</strong><a href="#9a32"><strong>1</strong></a><strong>] </strong>in Europe gives individuals the <em>“right to be forgotten.” </em>This means companies must delete a person’s data upon request. Not just from databases, but from any system that used it.</p><p>Similarly, California’s <strong>Consumer Privacy Act (CCPA) [</strong><a href="#123e"><strong>2</strong></a><strong>] </strong>gives users the right to have their data erased.</p><p>In traditional databases, this is relatively straightforward. In machine learning models? A much harder promise to keep.</p><p>Deleting a row from a spreadsheet is one thing. Removing the influence of a datapoint buried deep inside millions of parameters is something else entirely.</p><h4>Real-World Flashpoints: When Deletion Isn’t Enough</h4><p>This isn’t some distant, future problem. It’s happening now.</p><ul><li><strong>GitHub Copilot Lawsuit [</strong><a href="#67fa"><strong>3</strong></a><strong>]:</strong><br>GitHub Copilot, trained on massive amounts of public code, faced legal action because it allegedly regurgitated licensed code without proper attribution. Simply removing the original code files wouldn’t solve the problem. The model had already learned from them.</li><li><strong>Getty Images vs. Stability AI [</strong><a href="#f418"><strong>4</strong></a><strong>]:</strong><br>Stability AI, creators of Stable Diffusion, was sued for using millions of copyrighted images without permission. Even if they deleted the raw images, the trained models still carried patterns and knowledge derived from them.</li><li><strong>Authors vs. OpenAI [</strong><a href="#c98b"><strong>5</strong></a><strong>]:</strong><br>Well-known authors, including Sarah Silverman and others, sued OpenAI for training language models on their books without consent. Again, deleting the books from the dataset wouldn’t fully undo the learning baked into the models.</li></ul><p>In each of these cases, the question wasn’t just <strong>“Did you use my data?” </strong>It was <strong>“Can you un-use it?” </strong>And for now, in most systems, the answer is: not really.</p><h3>What is Machine Unlearning, Technically?</h3><p>At its core, Machine Unlearning is simple to define but very hard to actually do.</p><p><strong>Machine Unlearning</strong> means removing the influence of specific training data from a machine learning model, <em>as if </em>the data had never been there in the first place.</p><p>It’s not the same as just deleting a file. In machine learning, once a model has trained on data, that information gets baked deep into the model’s internal weights, the mathematical structures that guide how it makes decisions. The model doesn’t just remember the exact data points; it <em>generalizes</em> from them, learns patterns from them, and those patterns influence all future predictions.</p><p>That’s what makes unlearning so tricky. You can’t just go into a model and highlight a memory to erase. Data doesn’t sit in one place inside a model. It’s scattered, mixed in, and entangled with everything else the model has learned.</p><h4>Why Is Machine Unlearning So Difficult?</h4><p>There are a few big challenges that make unlearning much harder than it sounds:</p><ul><li><strong>Entanglement:</strong><br>One piece of data doesn’t affect just one part of the model. It influences many layers and neurons, often in subtle ways. Trying to remove it cleanly without disturbing everything else is a lot like trying to pull out a single thread from a spider’s web.</li><li><strong>Cost of Retraining:<br></strong>The obvious solution seems to be to retraining the model from scratch without the unwanted data. But it is usually too expensive. Large models can take <strong>weeks or months</strong> to train on supercomputers, costing millions of dollars. It’s not realistic to start over every time someone asks to remove a datapoint.</li><li><strong>Verification:<br></strong>Even if you say you “unlearned” something, how do you <em>prove</em> it? There’s no easy test today that shows a model has truly forgotten a specific piece of data.</li></ul><p>Imagine you’re teaching someone a complex skill. Something like playing the piano. If you later told them, <em>“Forget everything you learned about the F major chord,” </em>they couldn’t just delete that memory instantly. It’s mixed into how they play music now. You’d have to reteach them from scratch, or find some careful way to reshape their memory without ruining everything else.</p><p>Machine Unlearning faces a similar problem. It’s about not just deleting data, but <strong>untangling its influence</strong> from everything the model has become.</p><p>And that’s where most of the research today is focused:<br><strong>How can we teach machines to forget quickly and reliably without breaking everything else they’ve learned?</strong></p><h3>How Machine Unlearning Happens: Methods and Research</h3><p>Now that we know why unlearning is difficult, the next step is understanding how researchers are trying to solve it, both practically and theoretically.</p><p>Over the past few years, researchers have come up with a fascinating range of strategies, each trying to tackle this challenge from a different angle.</p><h4>Approximate Unlearning</h4><p>Approximate methods adjust the existing model <em>just enough </em>to erase the influence of the unwanted data. This could involve selectively fine-tuning certain layers, reversing the learning on particular examples, or rolling back to an earlier checkpoint in training. The goal is to forget efficiently, without blowing up the training cost.</p><p><strong>Why it’s appealing:</strong></p><ul><li>Much faster and less resource-heavy than full retraining.</li><li>Can be adapted to many existing model architectures easily.</li><li>Practical even when there are frequent deletion requests.</li></ul><p><strong>But it’s not perfect.</strong></p><ul><li>Forgetting might not be complete. Some traces of the data could linger.</li><li>Difficult to provide strong formal guarantees of deletion.</li></ul><p>The 2022 paper <strong>Fast Yet Effective Machine Unlearning</strong><em> </em>by Ayush et al.<strong> [</strong><a href="#c2d7"><strong>6</strong></a><strong>] </strong>explores exactly this efficiency tradeoff. They propose a checkpointing strategy: saving snapshots of the model during training so that when deletion is needed, the model can roll back and fine-tune just specific parameters rather than retraining everything.</p><figure><img alt="Overview of the Fast Yet Effective Machine Unlearning framework." src="https://cdn-images-1.medium.com/max/897/1*5wVPVY-o4yXa99Ux_8DX_Q.png" /><figcaption>Source: <strong>Fast Yet Effective Machine Unlearning</strong> by Ayush et al. [<a href="#c2d7">6</a>] (used under fair use for educational commentary). The model first generates error-maximizing noise to selectively disrupt the model’s memory of the target data. It then performs a selective impairment and repair process to restore model performance while forgetting the specific data points.</figcaption></figure><h4>Influence Functions</h4><p>What if you could figure out exactly how much a single training example affected your model and then just undo that effect? That’s the magic of influence functions. Originally from classical statistics, these techniques let you estimate how a small change in the training set would change the model’s behavior. In the context of machine unlearning, this means you might be able to surgically remove a datapoint’s impact without retraining the whole network.</p><p><strong>Why it’s appealing:</strong></p><ul><li>Highly targeted forgetting and not a full rework of the model.</li><li>Theoretically elegant and grounded in well-studied mathematics.</li></ul><p><strong>But it’s not perfect.</strong></p><ul><li>Estimating influence is tough for large deep learning models.</li><li>Computationally heavy for large datasets or complex models.</li></ul><p>The paper <strong>Amnesiac Machine Learning</strong> by Graves et al.<strong> [</strong><a href="#afa7"><strong>7</strong></a><strong>] </strong>revisits influence functions specifically for unlearning. They propose a method where, instead of retraining, the model parameters are adjusted based on an approximation of how much the forgotten data influenced them.</p><h4>SISA Training</h4><p>Instead of treating the training dataset as one big monolithic block, what if we sliced it into shards? SISA training (short for Sharded, Isolated, Sliced, and Aggregated) splits the data into smaller parts, trains independent models (or partial models) on each shard, and then combines them when making predictions. If a user asks to delete their data, only the shards containing that data need to be retrained, saving a ton of time and computational effort.</p><p><strong>Why it’s appealing:</strong></p><ul><li>Significantly reduces the cost of retraining for unlearning.</li><li>Makes the process of auditing and compliance (like GDPR) more manageable.</li></ul><p><strong>But it’s not perfect.</strong></p><ul><li>Slightly lower accuracy compared to unified full-dataset training.</li><li>Adds extra complexity in managing shards and aggregation logic.</li></ul><p>The 2021 paper <strong>Machine Unlearning</strong> by Bourtoule et al.<strong> [</strong><a href="#3275"><strong>8</strong></a><strong>] </strong>introduces this clever idea. By isolating training data into different shards and training them separately, they make it possible to surgically delete a user’s data without touching unrelated parts of the model.</p><figure><img alt="The SISA architecture" src="https://cdn-images-1.medium.com/max/725/1*daKDPYAjWZxqHQfXmydhIg.png" /><figcaption>Source: <strong>Machine Unlearning </strong>by Bourtoule et al. [<a href="#3275">8</a>] (used under fair use for educational commentary). The SISA framework partitions training data into isolated slices and trains separate models on each. When unlearning is required, only the slices containing the forgotten data are retrained while others remain unchanged.</figcaption></figure><h4>Neuron Masking and Model-Intrinsic Methods</h4><p>Sometimes, rather than treating the whole model as a black box, researchers look <em>inside</em> the model itself to find ways of forgetting. Model-intrinsic methods work by identifying and disabling the specific neurons, layers, or parameters that learned from the data we now want to erase. It’s a more <em>surgical </em>approach: finding what parts of the brain learned the wrong thing and silencing just those.</p><p><strong>Why it’s appealing:</strong></p><ul><li>Often more precise than global retraining.</li><li>Can maintain most of the model’s original performance.</li><li>Potentially faster than dataset-wide reprocessing.</li></ul><p><strong>But it’s not perfect:</strong></p><ul><li>Requires deeper access to model internals, not always possible for deployed models.</li><li>Risks accidentally impairing unrelated parts of the model if neuron selection isn’t accurate.</li></ul><p>In <strong>Learn to Forget: Machine Unlearning via Neuron Masking </strong>by Ma et al.<strong> [</strong><a href="#6376"><strong>9</strong></a><strong>]</strong>, the authors propose a method where neurons that heavily encode the forgotten data are masked (disabled) rather than retrained. They introduce a technique to measure neuron importance based on activation statistics, and strategically mask neurons to degrade performance on the forget set while preserving accuracy on retained data.</p><figure><img alt="The “Learn to Forget” framework" src="https://cdn-images-1.medium.com/max/1024/1*u8ff55eZHLsOYp6PaKU7_A.png" /><figcaption>Source: <strong><em>Learn to Forget: Machine Unlearning via Neuron Masking</em></strong><em> by </em>Ma et al. [<a href="#6376">9</a>] (used under fair use for educational commentary). The framework uses a mask gradient generator that selectively adjusts the network’s gradients, impairing the model’s memory of unlearned samples while preserving performance on retained data through iterative optimization.</figcaption></figure><h4>Data-Driven Approaches</h4><p>Another way to enable unlearning is not by changing the model architecture, but by changing the data pipeline itself. Data-driven approaches partition, modify, or design training datasets so that unlearning becomes easier later. For example, if training data is organized into isolated parts, forgetting can happen by removing just one part and adjusting the model only a little without touching the rest.</p><p><strong>Why it’s appealing:</strong></p><ul><li>Reduces the retraining overhead drastically.</li><li>Fits well with privacy regulations by offering clear boundaries between data units.</li><li>Makes compliance easier for large deployments.</li></ul><p><strong>But it’s not perfect:</strong></p><ul><li>Can cause a slight performance drop because data isn’t fully “blended” during training.</li><li>Requires careful data management and tracking from the very beginning.</li></ul><p>The <strong>SISA Training </strong>framework by Bourtoule et al. (2021) <strong>[</strong><a href="#3275"><strong>8</strong></a><strong>]</strong>, discussed earlier, is a classic example of this philosophy: the dataset is sliced into shards and isolated during training to make future unlearning efficient. Another interesting paper, <strong>Safe: Machine Unlearning with Shard Graphs </strong>by<strong> </strong>Dukler et al. (2023)<strong> [</strong><a href="#2550"><strong>10</strong></a><strong>]</strong>, proposes building a shard graph where relationships between data shards are explicitly modeled, allowing for more precise unlearning based on how connected or independent different data points are.</p><h3>Certified Unlearning</h3><p>If approximate unlearning is about <em>doing your best to forget,</em> certified unlearning is about <em>proving you forgot</em>. As data privacy regulations like GDPR start asking hard questions, researchers realized we needed something stronger than just good intentions.</p><p>Certified unlearning is all about providing formal, mathematical proof that the forgotten data no longer has a meaningful influence on the model. Techniques here often use ideas like <strong>algorithmic stability</strong>, <strong>differential privacy</strong>, or even <strong>cryptographic verification</strong> to show that forgetting actually happened and not just that it looks like it did.</p><p><strong>Why it’s appealing:</strong></p><ul><li>Provides strong legal and technical assurances for compliance.</li><li>It builds a lot more trust, especially for sensitive domains like healthcare or finance.</li><li>If done right, it gives you confidence that even deleted examples leave no accidental traces.</li></ul><p><strong>But it’s not perfect:</strong></p><ul><li>It can be really heavy in terms of computation. Not as fast or easy as some approximate methods.</li><li>Sometimes you lose a little performance because of the stability or privacy constraints you have to add.</li><li>And scaling certified unlearning to giant deep networks? Still a work in progress.</li></ul><p>A landmark study here is <strong>Certified Data Removal from Machine Learning Models </strong>by Guo et al.<strong>[</strong><a href="#fb87"><strong>11</strong></a><strong>]</strong>. They explored how you could use <strong>algorithmic stability</strong> to formally prove that after data removal, the model’s predictions change only in tightly controlled ways, basically meaning, the model behaves as if the data never existed. It’s a really smart bridge between privacy theory and practical machine learning.</p><p>Another more recent take is from <strong>Verifiable and Provably Secure Machine Unlearning </strong>by Eisenhofer et al.<strong>[</strong><a href="#6d0c"><strong>12</strong></a><strong>]</strong>. They push the idea even further: using <strong>cryptographic proofs</strong> combined with retraining to not only perform unlearning, but make the entire process verifiable by third parties.</p><p>It’s a glimpse into what real-world privacy compliance might look like when unlearning becomes serious business.</p><h3>Reception and Current Landscape</h3><p>For something that sounds as technical as “machine unlearning,” the real-world reaction has been surprisingly emotional. There’s curiosity, excitement, and also a fair amount of worry.</p><p>On the positive side, researchers and policy makers are paying more attention to the <strong>right to be forgotten</strong>. Papers like <strong>A Survey of Machine Unlearning </strong>by Nguyen et al., 2022 <strong>[</strong><a href="#d0d6"><strong>13</strong></a><strong>] </strong>have emphasized how important it is for users to be able to <em>take back control</em> over their data, even after models have been trained.</p><p>Governments are starting to notice too. The UK’s <strong>International AI Safety Report 2025</strong> <strong>[</strong><a href="#4e49"><strong>14</strong></a><strong>]</strong> cited machine unlearning as one of the <em>critical technologies</em> for responsible AI in the future. The mood is clear: unlearning isn’t just a technical trick anymore. It’s slowly becoming an expectation.</p><p>On the flip side, though, <strong>practical challenges</strong> are hard to ignore.</p><p>Big companies working with massive foundation models (OpenAI, Google DeepMind, Anthropic) haven’t yet fully implemented scalable unlearning. Why? Because most production models are <strong>huge</strong>, <strong>complex</strong>, and <strong>constantly evolving</strong>. Unlearning one user’s data without retraining millions (or billions) of parameters is still a very real bottleneck.</p><p>You also see a tension between <strong>privacy</strong> and <strong>performance</strong>. Some engineers argue that enforcing strict unlearning could limit model capabilities. Others worry about “unlearning attacks,” where malicious requests try to force important knowledge out of models.</p><p>But despite all these growing pains, machine unlearning is no longer just an academic curiosity. There’s a strong sense that, like security audits or fairness benchmarks, <em>being able to prove you can forget</em> is going to be a baseline requirement for AI systems in the next few years.</p><h3>My Reflections</h3><p>If there’s one thing that stands out to me after diving deep into this topic, it’s this: <strong>Machine Unlearning isn’t optional. It’s necessary.</strong></p><p>We can’t keep brushing aside privacy concerns just because models get better results without them. Accuracy, performance, innovation, they all matter, of course. But without privacy protections, especially the ability to truly erase data when needed, the foundation we’re building AI on becomes shaky. It’s not just about compliance with laws like GDPR. It’s about basic respect: for users, for contributors, for everyone whose data models are built on.</p><p>From where I see it, <strong>machine unlearning needs to become a default expectation. </strong>Not a rare technical showcase, but something models are designed for from the very beginning. It should be baked into the way we think about AI architectures, training loops, and deployment strategies. Just like we now automatically think about security or fairness, <em>forgetting</em> must be part of the conversation too.</p><p>Of course, I realize it’s not easy. There are real technical challenges to doing this efficiently and scalably. But just because something is hard doesn’t mean we can afford to look away. In fact, it probably means it’s exactly where we should be focusing.</p><p>As always, this is just my perspective. I’m learning right along with everyone else. If you have thoughts, agreements, disagreements, or even different ways of looking at the unlearning challenge, I’d genuinely love to hear them.</p><p>Let’s keep the conversation open. This topic deserves it.</p><h3>References</h3><ol><li>European Union. “General Data Protection Regulation (GDPR). Article 17: Right to erasure (‘right to be forgotten’).” Available at: <a href="https://gdpr-info.eu/art-17-gdpr/">https://gdpr-info.eu/art-17-gdpr/</a></li><li>California Department of Justice, Office of the Attorney General. “California Consumer Privacy Act (CCPA).” Available at: <a href="https://leginfo.legislature.ca.gov/faces/codes_displayText.xhtml?division=3.&amp;part=4.&amp;lawCode=CIV&amp;title=1.81.5">https://leginfo.legislature.ca.gov/faces/codes_displayText.xhtml?division=3.&amp;part=4.&amp;lawCode=CIV&amp;title=1.81.5</a></li><li>Saveri Law Firm. “GitHub and Copilot Intellectual Property Litigation.” Available at: <a href="https://www.saverilawfirm.com/our-cases/github-copilot-intellectual-property-litigation">https://www.saverilawfirm.com/our-cases/github-copilot-intellectual-property-litigation</a></li><li>Jay Peters. “Getty Images sues AI art generator Stable Diffusion in the US for copyright infringement.” <em>The Verge</em>, 2023. Available at: <a href="https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-lawsuit-getty-images-stable-diffusion">https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-lawsuit-getty-images-stable-diffusion</a></li><li>Ashley Belanger. “Sarah Silverman sues OpenAI, Meta for being ‘industrial-strength plagiarists’.” <em>Ars Technica</em>, 2023. Available at: <a href="https://arstechnica.com/information-technology/2023/07/book-authors-sue-openai-and-meta-over-text-used-to-train-ai/">https://arstechnica.com/information-technology/2023/07/book-authors-sue-openai-and-meta-over-text-used-to-train-ai/</a></li><li>Tarun, Ayush K., et al. “Fast yet effective machine unlearning.” <em>IEEE Transactions on Neural Networks and Learning Systems</em> (2023). <a href="https://ieeexplore.ieee.org/abstract/document/10113700">https://ieeexplore.ieee.org/abstract/document/10113700</a></li><li>Graves, Laura, Vineel Nagisetty, and Vijay Ganesh. “Amnesiac machine learning.” <em>Proceedings of the AAAI Conference on Artificial Intelligence</em>. Vol. 35. №13. 2021. <a href="https://ojs.aaai.org/index.php/AAAI/article/view/17371">https://ojs.aaai.org/index.php/AAAI/article/view/17371</a></li><li>L. Bourtoule et al., “Machine Unlearning,” 2021 IEEE Symposium on Security and Privacy (SP), San Francisco, CA, USA, 2021, pp. 141–159, doi: 10.1109/SP40001.2021.00019. <a href="https://ieeexplore.ieee.org/document/9519428">https://ieeexplore.ieee.org/document/9519428</a></li><li>Ma, Zhuo, et al. “Learn to forget: Machine unlearning via neuron masking.” <em>IEEE Transactions on Dependable and Secure Computing</em> 20.4 (2022): 3194–3207. <a href="https://ieeexplore.ieee.org/abstract/document/9844865">https://ieeexplore.ieee.org/abstract/document/9844865</a></li><li>Dukler, Yonatan, et al. “Safe: Machine unlearning with shard graphs.” <em>Proceedings of the IEEE/CVF International Conference on Computer Vision</em>. 2023. <a href="https://openaccess.thecvf.com/content/ICCV2023/html/Dukler_SAFE_Machine_Unlearning_With_Shard_Graphs_ICCV_2023_paper.html">https://openaccess.thecvf.com/content/ICCV2023/html/Dukler_SAFE_Machine_Unlearning_With_Shard_Graphs_ICCV_2023_paper.html</a></li><li>Guo, Chuan, et al. “Certified data removal from machine learning models.” <em>arXiv preprint arXiv:1911.03030</em> (2019). <a href="https://arxiv.org/abs/1911.03030">https://arxiv.org/abs/1911.03030</a></li><li>Eisenhofer, Thorsten, et al. “Verifiable and provably secure machine unlearning.” <em>arXiv preprint arXiv:2210.09126</em> (2022). <a href="https://arxiv.org/abs/2210.09126">https://arxiv.org/abs/2210.09126</a></li><li>Nguyen, Thanh Tam, et al. “A survey of machine unlearning.” <em>arXiv preprint arXiv:2209.02299</em> (2022). <a href="https://arxiv.org/abs/2209.02299">https://arxiv.org/abs/2209.02299</a></li><li>Department for Science, Innovation and Technology, UK Government. “International AI Safety Report 2025.” Available at: <a href="https://www.gov.uk/government/publications/international-ai-safety-report-2025/international-ai-safety-report-2025">https://www.gov.uk/government/publications/international-ai-safety-report-2025/international-ai-safety-report-2025</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5aa2f3c7d495" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stories in Transition: Exploring AI’s Role in Modern Storytelling]]></title>
            <link>https://medium.com/@rithesh18.k/stories-in-transition-exploring-ais-role-in-modern-storytelling-0a1c513e240d?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/0a1c513e240d</guid>
            <category><![CDATA[storytelling]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[creativity]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Sat, 26 Apr 2025 20:24:42 GMT</pubDate>
            <atom:updated>2025-05-01T20:30:31.606Z</atom:updated>
            <content:encoded><![CDATA[<h4>A conversation on creativity, authorship, and how technology is reshaping the art of storytelling. <a href="https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8">AI Through the Looking Glass</a> #2</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6F3rcgYxjqsCdRWt" /></figure><h3>Stories at the Fire, and Beyond</h3><p>Since the beginning of time, humans have gathered around fires, huddled in small circles, and told stories. Stories that carried memories, passed down wisdom, built communities. From ancient myths etched onto cave walls to novels that shaped entire generations, storytelling has always been at the center of what it means to be human.</p><p>And as someone who’s written short stories and poems myself, I’ve always seen storytelling not just as a craft, but as a quiet dialogue between imagination, memory, and emotion. A way of making sense of the world and our place in it.</p><p>Today, another voice is entering that circle. One we didn’t quite expect: the voice of machines.</p><p>Artificial intelligence, especially large language models like GPT, Claude, and Gemini, has started to move from summarizing news articles and answering questions to something much more ambitious: telling stories. Not just assisting writers, but generating narratives from scratch. Suggesting plot twists. Creating characters. Writing dialogue.</p><p>It’s easy to dismiss this at first. After all, how could a machine, trained on patterns and probability, truly understand the messy, emotional complexity of a story? But the results we’re seeing are hard to ignore. AI-generated stories are showing up everywhere: in novels, screenplays, advertisements, and even short animated shows. And while some of these outputs still feel hollow, others are surprisingly coherent, imaginative, even moving in their own strange way.</p><p>So where are we right now? <br>Is AI helping storytelling, or redefining it altogether? <br>And more importantly, can a machine that has never lived, lost, or loved, truly tell a human story?</p><p>In this piece, we’ll step through the new landscape of AI storytelling: the tools making it possible, the science behind it, the gains, the gaps, and what it all means for the future of creativity.</p><h3>The Tools of Today: New Companions for Storytellers</h3><p>Not long ago, if you wanted help writing a story, you turned to a friend, a teacher, or maybe just a blank page and stubborn persistence. Now, there’s a new kind of collaborator: artificial intelligence.</p><p>Across the creative world, a wave of AI-powered tools is quietly weaving itself into the storytelling process: not replacing writers, but standing beside them, offering suggestions, nudges, possibilities.</p><p>One of the most well-known platforms is <a href="https://www.sudowrite.com/"><strong>Sudowrite</strong></a>, a writing assistant built specifically for fiction authors. Originally designed to help with writer’s block, Sudowrite now offers features like character development, plot brainstorming, and “twist suggestions.” A few keystrokes, and the AI proposes unexpected turns in a story you thought had nowhere left to go.</p><p>Then there’s <a href="https://shortlyai.com/"><strong>ShortlyAI</strong> </a>and <a href="https://www.jasper.ai/"><strong>Jasper</strong></a>, which are often used by bloggers, marketers, and content creators. These tools don’t just predict the next word; they suggest full narrative continuations, plot summaries, even emotional tone shifts. They’re tuned not for simple text completion, but for momentum: pushing the story forward, offering new angles when a human mind hesitates.</p><p>At the frontier of ambition is <a href="https://www.showrunner.xyz/"><strong>Showrunner</strong></a>, developed by Fable Studio. Showrunner generates full animated TV episodes. Backgrounds, character dialogue, scene transitions, all orchestrated by an AI pipeline. In a matter of hours, something that once took entire teams of writers, designers, and editors can now be prototyped by a handful of prompts.</p><p>Even more experimental tools like <a href="https://www.lumistory.ai/"><strong>Lumi</strong></a>, launched by activist Colin Kaepernick, are attempting to democratize storytelling by helping marginalized voices create, illustrate, and publish their own narratives with AI’s help.</p><p>These tools aren’t perfect. But they’re changing who gets to tell stories, and how fast they can move from an idea to something tangible. For some, it’s a dream come true. For others, it raises a deep unease: If stories come this easily, do they lose something along the way?</p><p>Before we answer that, it’s worth stepping back and asking a simpler question: <strong>How did AI even learn to tell stories in the first place?</strong></p><h3>Under the Hood: How AI Learned to Tell a Story</h3><p>Before we ask what AI can or cannot capture in a story, it’s worth understanding how it even learned to tell one in the first place.</p><p>At its core, an AI model like GPT, Claude, or Gemini doesn’t “think” about stories the way humans do. It doesn’t imagine a world or feel a character’s sadness. What it does, and does astonishingly well, is predict patterns.</p><p>These models are trained on massive libraries of human writing: novels, essays, scripts, dialogues, even everyday conversations. Datasets like <a href="https://commoncrawl.org/">Common Crawl</a>, <a href="https://huggingface.co/datasets/bookcorpus/bookcorpus">BookCorpus</a>, Wikipedia, and WebText form the foundation. The AI ingests billions of sequences of words and learns one basic task: <strong>Given a sequence, what’s most likely the next word?</strong></p><p>It sounds mechanical. And it is.</p><p>But when scaled up across millions of stories, plays, poems, and articles, the AI starts to learn more than just sentences. It picks up the <strong>structures of narrative</strong>:</p><ul><li>A character’s introduction leads to a challenge.</li><li>Emotional arcs tend to rise and fall like waves.</li><li>Good stories often resolve conflict, but sometimes deliberately leave questions hanging.</li></ul><p>This is why AI today can draft coherent, well-structured stories. Not because it understands storytelling emotionally, but because it recognizes the <strong>patterns</strong> of how stories are usually built.</p><p>Recent research has tried to map just how far this ability goes, and where it falls short.</p><p>In the 2024 paper “<a href="https://arxiv.org/abs/2407.13248"><strong>Are Large Language Models Capable of Generating Human-Level Narratives?</strong></a><strong>”,</strong> researchers found that while LLMs can produce complete stories with clear beginnings, middles, and endings, they often struggle with <strong>suspense</strong>, <strong>emotional layering</strong>, and <strong>character depth</strong>. Their narratives are coherent but flatter and more predictable.</p><p>Similarly, <strong>“</strong><a href="https://www.nature.com/articles/s41599-024-03868-8"><strong>Experimental Narratives: A Comparison of Human Crowdsourced Storytelling and AI Storytelling</strong></a><strong>” </strong>showed that human written stories tended to weave in cultural nuances, ambiguous emotional beats, and unexpected twists: features that AI still finds difficult to replicate naturally.</p><p>Other studies have tried to go even deeper: into the emotional <strong>persuasion</strong> of stories.</p><p>In <strong>“</strong><a href="https://academic.oup.com/joc/article-abstract/74/5/347/7756907"><strong>Can AI Tell Good Stories? Narrative Transportation and Persuasion in AI-Generated Content</strong></a><strong>”</strong>, researchers explored how well AI stories could “transport” readers: pull them emotionally into the world of the story. They found that while AI stories could create logical flow and surface-level emotion, they lacked the complexity needed to deeply persuade or move audiences.</p><p>In other words: AI could entertain, sometimes even impress. But it struggled to <em>make you feel</em>.</p><p>And when it comes to storytelling across <strong>multiple modes: </strong>combining text with imagery, animation, or speech, things get even more interesting.<br> A 2024 study, <strong>“</strong><a href="https://arxiv.org/abs/2409.11261"><strong>The Art of Storytelling: Multi-Agent Generative AI for Dynamic Multimodal Narratives</strong></a><strong>”,</strong> demonstrated that AI can now generate interactive children’s stories, blending voice and animation. It’s still early, but it suggests a future where AI is not just writing stories, but directing experiences.</p><p>In a way, AI today is like a brilliant student who has read every story ever written, and can imitate their form perfectly. But the experience you live: the feeling of falling in love, of losing someone, of longing for something you can’t quite name, that, no training set can truly teach.</p><p>AI knows what a story <strong>should</strong> look like.<br>But whether it knows <strong>why </strong>stories matter, that’s still a human domain.</p><h3>The Good: How AI is Enhancing Storytelling</h3><p>For all the concerns about whether AI can truly tell human stories, there’s no denying that it’s already helping storytellers in powerful, practical ways.</p><p>One of the clearest benefits is <strong>accessibility</strong>.</p><p>In the past, writing a story, even a short one, could feel overwhelming. Plotting arcs, developing characters, finding the right tone: these were skills that took years to practice, and for many, the blank page remained an intimidating enemy.</p><p>Today, tools like <strong>Sudowrite</strong>, <strong>ShortlyAI</strong>, and <strong>Jasper</strong> are making that entry point easier. They can help structure narratives, suggest emotional beats, and even propose alternate endings when a writer gets stuck.</p><p>In a sense, AI is lowering the barrier between idea and execution. A writer with a fragment of an idea: a single scene, a theme, a character sketch, can now explore that thought much more quickly. This doesn’t erase the need for human creativity, but it <strong>supports</strong> it. Like scaffolding around a building still under construction.</p><p>Another area where AI is showing real promise is in <strong>speed and iteration</strong>.</p><p>Writers often speak about the “ugly first draft”. The necessary, messy process of pouring ideas out quickly without judgment. AI can now assist with this phase, helping generate rough drafts at speeds that would be impossible manually. It’s not about writing the final masterpiece. It’s about <strong>giving creators something to work on</strong>, tweak, and reshape.</p><p>Recent studies, such as <strong>“</strong><a href="https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4954803"><strong>Artificial Intelligence-Driven Emotional Storytelling for Brand Communication</strong></a><strong>”</strong>, show that AI-generated narratives are already being used in marketing and branding. Companies are using AI to quickly prototype multiple storylines, test emotional tones, and refine content based on audience response.</p><p>It’s not replacing writers. It’s accelerating the creative feedback loop.</p><p>Then there’s the <strong>exploration of new narrative spaces</strong>.</p><p>Because AI models are trained on diverse sources, from literary classics to genre fiction to fan communities, they can suggest story patterns that sometimes sit outside a writer’s habitual style. An author steeped in realism might find themselves experimenting with magical realism after an AI-generated suggestion. A thriller writer might suddenly explore a romantic subplot. It doesn’t always work, but it opens doors for something unique.</p><p>In research like <strong>“</strong><a href="https://arxiv.org/abs/2409.11261"><strong>The Art of Storytelling: Multi-Agent Generative AI for Dynamic Multimodal Narratives</strong></a><strong>”</strong>, scientists showed how AI can blend text, animation, and speech to create engaging multimodal stories, especially for education and children’s content. This hints at future collaborations where human imagination and machine capabilities stretch storytelling into new formats we’re only beginning to imagine.</p><p>Of course, the best results still come when AI is treated as a <strong>co-pilot</strong>, not a captain. Writers who use AI as a brainstorming partner, and not a replacement for thought, tend to create the richest, most resonant work.</p><p>In that sense, AI isn’t changing the nature of storytelling itself. It’s simply offering a new tool to those brave (and curious) enough to pick it up.</p><h3>The Limits: What AI Still Struggles to Capture</h3><p>For all the impressive gains AI has made in helping people tell stories, there are still clear gaps: places where the seams start to show, and where the difference between human and machine creativity becomes unmistakable.</p><p>The first and most fundamental limitation is <strong>emotional depth</strong>.</p><p>AI can generate coherent narratives, complete with characters, plot twists, and even surface-level emotional beats. But deeper emotional arcs, the slow burn of grief, the complex contradictions of love, the dull ache of loneliness, are much harder to recreate.</p><p>Going back to <strong>“</strong><a href="https://academic.oup.com/joc/article-abstract/74/5/347/7756907"><strong>Can AI Tell Good Stories? Narrative Transportation and Persuasion in AI-Generated Content</strong></a><strong>”</strong>, researchers found that while AI stories could logically progress through events, they often struggled to create the immersive emotional pull known as <strong>narrative transportation: </strong>the feeling of being “inside” the story world. Without that emotional pull, a story might look perfect on paper, but still feel strangely hollow, missing something in it.</p><p>Another limit lies in <strong>cultural richness and nuance</strong>. Human storytelling isn’t just about plots and conflicts. It’s a reflection of cultural memory, personal history, and shared values.</p><p>As shown in the paper <strong>“</strong><a href="https://www.nature.com/articles/s41599-024-03868-8"><strong>Experimental Narratives: A Comparison of Human Crowdsourced Storytelling and AI Storytelling</strong></a><strong>”, </strong>human-generated stories were consistently rated higher in terms of cultural specificity, ambiguity, and layered emotional meaning.</p><p>AI models, trained on vast and <em>generalized </em>data from across the internet, tend to flatten those subtleties. They reproduce familiar tropes: a hero’s journey, a betrayal, a redemption, but often miss the quieter cultural details that make a story feel truly lived-in: the smell of a particular street food we see in <em>Malgudi Days</em>, the aching, quiet grief that runs through <em>The Kite Runner</em>, or the rhythmic, overlapping arguments around the dinner table in <em>Everything Everywhere All At Once</em>.</p><p>Without these textured layers, AI storytelling can sometimes feel like a very good <em>approximation. C</em>lose enough to impress at first glance, but missing the resonances that stay with a reader long after the end.</p><p>Finally, there’s the problem of <strong>predictability</strong>.</p><p>Because AI models are trained to predict the most likely next word, they often generate stories that are competent but <em>safe</em>. They favor familiar arcs, satisfying endings, symmetrical structures… What they rarely produce, without deliberate intervention, are the messy, surprising, subversive twists that human storytellers often create, either by instinct or by aversion to tradition.</p><p><strong>“</strong><a href="https://arxiv.org/abs/2407.13248"><strong>Are Large Language Models Capable of Generating Human-Level Narratives?</strong></a><strong>” </strong>noted this directly: while LLMs could produce logically sound stories, they lacked the “narrative risk-taking” often seen in human writing: the willingness to leave threads unresolved, to end on ambiguity, to embrace discomfort.</p><p>And in many ways, that’s what makes great storytelling memorable: Not the perfection of the structure, but the imperfections that reveal something raw, messy, and real about being human.</p><p>Today, AI can tell a story that looks right, sounds right, and moves cleanly from beginning to end. But the stories that stay with us, the ones that haunt, heal, provoke, are still, overwhelmingly, stories written by hands that have lived them.</p><h3>Real-World Tensions: When AI Meets Resistance</h3><p>While AI storytelling tools are being embraced by some as exciting new collaborators, not everyone in the creative world sees it that way. Over the past two years, a deeper tension has been building: a conflict over <strong>what it means to create</strong>, <strong>who owns a story</strong>, and <strong>how much of that process can, or should, be automated</strong>.</p><h4>Hollywood: Writers Strike and the Fight for Authorship</h4><p>The tension around AI and storytelling reached a historic turning point during the <strong>Writers Guild of America (WGA) strike</strong> in 2023. Writers weren’t just fighting for better pay or benefits: they were fighting for the very definition of creative work in the age of AI.</p><p>One of their key concerns was the rise of AI models capable of drafting treatments, outlines, or even full scripts. Writers feared a future where studios would rely on AI to generate cheap drafts, then bring in human writers only for minor edits: diluting not just their jobs, but their creative ownership.</p><p>The WGA strike officially ended in September 2023, after 148 days: one of the longest strikes in Hollywood history. And while many issues were negotiated, <strong>AI protections became a landmark feature of the final agreement.</strong></p><p>Under the new terms:</p><ul><li><strong>AI cannot write or rewrite literary material.</strong></li><li><strong>AI-generated material cannot be considered “source material”: </strong>a human writer’s credit, rights, or compensation cannot be undermined by AI outputs.</li><li><strong>Writers can choose to use AI tools voluntarily, but studios cannot force them to use AI.</strong></li></ul><p>As the WGA summarized it:</p><blockquote><em>“</em>AI can’t write or rewrite literary material, and AI-generated material will not be considered source material under the MBA<em>”<br>(</em><a href="https://www.wgacontract2023.org/the-campaign/summary-of-the-2023-wga-mba"><em>WGA Contract Agreement Summary, 2023</em></a><em>)</em></blockquote><p>Real projects were impacted during the strike:</p><ul><li><strong>Stranger Things Season 5</strong> delayed its production.</li><li><strong>Blade (Marvel Studios)</strong> temporarily halted filming.</li><li>New writers’ rooms across streaming platforms like Netflix and Apple TV were frozen for months.</li></ul><p>But ultimately, the outcome signaled something important: Creativity, authorship, and storytelling <strong>remain human domains.</strong></p><h4>Publishing: Authors Push Back Against AI Scraping</h4><p>In publishing, the backlash was equally strong, and even more personal.</p><p>In 2024, <strong>the Authors Guild organized an open letter</strong> signed by over 8,000 writers, including household names like <strong>Margaret Atwood</strong>, <strong>Jonathan Franzen</strong>, <strong>James Patterson</strong>, <strong>George R.R. Martin</strong>, and <strong>Nora Roberts</strong>.</p><p>Their demand was straightforward: Stop training AI models on their copyrighted work without permission.</p><p>The CEO of the Authors Guild, Mary Rasenberger, said in a statement:</p><blockquote><em>“</em>It is imperative that we stop this theft in its tracks or we will destroy our incredible literary culture, which feeds many other creative industries in the US…<em>”<br>(</em><a href="https://www.theguardian.com/books/2023/sep/20/authors-lawsuit-openai-george-rr-martin-john-grisham"><em>Guardian, 2023</em></a><em>)</em></blockquote><p>Several lawsuits followed, notably <strong>Sarah Silverman’s lawsuit</strong> against OpenAI and Meta for unauthorized use of her books in training datasets. Another clear signal that creators were drawing a line around their intellectual property.</p><p>(<a href="https://authorsguild.org/news/ag-submits-guidance-for-national-ai-action-plan/">Authors Guild Open Letter, 2024</a>)</p><h4>Gaming: AI-Generated Lore and Narrative Fatigue</h4><p>In the gaming industry, the tension is quieter, but growing.</p><p>Games like <strong>Starfield</strong> (Bethesda, 2023) showcased vast, expansive worlds, but players and critics noted that many side quests felt repetitive and lacked the handcrafted magic of earlier titles like <em>The Witcher 3</em> or <em>Red Dead Redemption 2</em>.</p><p>Developers later revealed that AI-assisted tools were used to generate quest outlines and dialogue templates at scale, a strategy intended to fill large open worlds faster.</p><p>At the 2024 <strong>Game Developers Conference (GDC)</strong>, narrative designers voiced concern:</p><blockquote><em>“</em>Procedural content can fill a map, but it can’t fill a heart.<em>”</em></blockquote><h4>Journalism: AI Can Write News, but Can It Understand It?</h4><p>In journalism, AI is already generating basic earnings reports, sports summaries, and weather updates.</p><p>Reuters’ <strong>Lynx Insight</strong> platform, launched in 2023, uses AI to draft preliminary articles, allowing journalists to focus on deeper investigations. But human journalists are still crucial when the story demands context, sensitivity, or moral judgment, things machines struggle with.</p><p>A simple sports score is one thing. Telling the story of a community devastated by a flood, or the personal testimony of a protestor, is another.</p><p>As one editor at the New York Times put it:</p><blockquote><em>“</em>Accuracy is important, but meaning is what makes a story matter. And meaning still requires a human heart.<em>”</em></blockquote><h4>The Core of the Conflict</h4><p>Across industries, the core concern isn’t about resisting technology. It’s about defending storytelling as something human, lived, flawed, and real.</p><p>Because at its best, a story isn’t just data.<br>It’s the smell of rain in a childhood memory.<br>It’s the tremor in a hero’s voice when they say goodbye.<br>It’s all the little imperfections that no algorithm can quite predict, but every human reader feels.</p><p>And that’s something we’re not ready to automate. And maybe we never should.</p><h3>My Thoughts: Between Fascination and Caution</h3><p>After everything we’ve walked through so far, the breakthroughs, the excitement, the real-world tensions, I find myself standing somewhere in the middle: both fascinated and cautious.</p><p>As someone who has written stories and poems myself, I know how personal storytelling really is. It’s not just the arrangement of scenes and character arcs. It’s the quiet pauses, the contradictions that don’t fully resolve, the emotional weight that lingers between the lines.</p><p>It’s something lived, not just calculated.</p><p>On a technical level, and as a researcher, I find the progress in AI storytelling incredible. The fact that a model trained on billions of words can produce a coherent story, with a recognizable beginning, middle, and end, is astonishing. The way these tools can support writers by brainstorming, drafting, and suggesting new directions has already opened doors for creativity that didn’t exist before. AI is undeniably making it easier for more people to start telling their stories.</p><p>But where I start feeling cautious is about what we risk losing without realizing it.</p><p>Because storytelling isn’t just about efficiency. It’s not about generating 100 plotlines at once or perfectly following a narrative arc.<br>It’s about voice.<br>Memory.<br>Emotion that refuses to fit neatly into any structure.</p><p>When we use AI to generate quick drafts or fast content, that’s one thing.<br> But when we start replacing the slow, uncertain, human process of creating, the process that often includes doubt, confusion, contradiction, I wonder if we start losing what makes stories truly stick with us.</p><p>And there’s another side to this: respect for the people behind the stories.<br> The WGA strike, the Authors Guild lawsuits, the public pushback across creative industries, they’re all reminders that stories don’t come from nowhere. They are labors of time, culture, experience, and vulnerability.<br> Training machines on those works without permission, or trying to automate creativity without honoring the creators behind it, isn’t innovation. It’s erasure.</p><p>In the end, I don’t believe AI is “bad” for storytelling. But I do believe we have to be careful about how we use it, and what we start valuing. Speed and volume are tempting. But depth, resonance, and meaning, the reasons we tell stories in the first place, take time.</p><p>AI can help build better tools for writers.<br>It can suggest new paths, and amplify creativity.<br>But it cannot, and should not, replace the messy, difficult, beautiful act of being human in a story.</p><p>At least, that’s where I stand today.</p><h3>An Open Ending</h3><p>Stories rarely have hard endings. Even when a book closes or the credits roll, the questions and emotions they spark tend to linger: shifting, evolving, finding new meaning over time.</p><p>That’s exactly how I see the conversation around AI and storytelling, too.</p><p>Right now, we’re still early in understanding what it means for machines to assist, influence, or maybe even shape the narratives we tell. There are possibilities to be excited about, and realities we need to confront with caution and honesty.</p><p>I don’t claim to have all the answers.</p><p>This blog, and this entire series, is my effort to explore these questions openly:</p><ul><li>What does it mean for creativity when anyone can generate a story with a few prompts?</li><li>How do we protect the voices, cultures, and memories that stories have always carried?</li><li>And how do we make sure that in the rush to innovate, we don’t lose the human touch that made storytelling matter in the first place?</li></ul><p>I’d love to hear your thoughts.</p><p>If something in this piece resonated with you, or if you see things differently, I hope you’ll share your perspective.</p><p>Comment, message me, point me to something new to read. The best part about stories has always been that they grow when more voices join in.</p><p>Because in the end, the future of storytelling isn’t something AI will decide. It’s something <strong>we</strong> will.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0a1c513e240d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Studio Ghibli Meets AI]]></title>
            <link>https://medium.com/@rithesh18.k/studio-ghibli-meets-ai-8ce90e7de09a?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/8ce90e7de09a</guid>
            <category><![CDATA[ethics-in-tech]]></category>
            <category><![CDATA[digital-culture]]></category>
            <category><![CDATA[studio-ghibli]]></category>
            <category><![CDATA[ai-art]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Sun, 13 Apr 2025 08:48:17 GMT</pubDate>
            <atom:updated>2025-05-01T20:29:44.064Z</atom:updated>
            <content:encoded><![CDATA[<h4><a href="https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8"><em>AI Through the Looking Glass</em></a><em> #1</em></h4><h3>When AI Meets Art</h3><p>The last couple of weeks, I’ve been seeing a lot of Studio Ghibli-style portraits floating around online. People are uploading photos of themselves and watching them transform into soft, whimsical frames that look like they belong in a scene from <em>Spirited Away</em> or <em>My Neighbor Totoro</em>. And as someone who’s always loved Studio Ghibli’s aesthetic (<em>Princess Mononoke</em> is one of my all time favorite movies), I’ll admit, it’s hard not to appreciate how good they look.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2EKsP56prr6XqTW1SYT4Vw.jpeg" /><figcaption>Spirited Away: The most popular and critically acclaimed Studio Ghibli movie</figcaption></figure><p>It’s also fascinating to see how quickly it’s caught on. What used to be a niche experiment in AI art is now everywhere. In profile pictures, on Instagram reels, even circulating in group chats. Generative models have gone from being technical marvels to everyday creative tools.</p><p>This isn’t the first time we’ve seen AI show up in creative spaces. From music generation to visual design, the technology has slowly been evolving into the world of human expression. As days pass by AI is getting a bit more polished, more public, and more involved.</p><p>And while most people are engaging with it at face value — uploading a selfie, admiring the result, moving on — what’s happening underneath is worth talking about. How did these models get so good? What kind of AI is powering this new creative wave? And what are artists, engineers, and everyone in between saying about it?</p><p>That’s where this blog begins. I want to explore both the technology and the conversation around it. Not just how it works, but how it’s landing with people. And maybe, in the process, open up space for a more informed and balanced dialogue.</p><p>Let’s start by breaking down how we got here, and what’s really behind these generative art models.</p><h3>Behind the Magic: How AI Generates Art That Feels Human</h3><p>If you’ve been wondering how AI models have gotten <em>this</em> good at generating art, you’re not alone. The filters, the portraits, the Ghibli-style animations — they’re all built on top of models that have evolved rapidly over the last few years.</p><p>This section dives into what’s going on under the hood. We’ll walk through diffusion models, latent space, conditioning, and the role of models like CLIP. I’ll try to touch upon the technical aspects of these models but not to overwhelm with information. But if it feels too Greek and Latin, please let me know through the comments. I’ll try to simplify it. Or if it’s too simple and not detailed enough, let me know that as well and I’ll complicate things next time XD.</p><h4>From Noise to Masterpiece: Diffusion Models</h4><p>At the core of most AI art tools today is a concept called <strong>denoising diffusion</strong>.</p><p>The core idea is simple. Take an image and slowly add just a little noise in every step. Do this over many many steps, until it becomes unrecognizable. Just like a pure noise. Then, train a neural network to do the reverse: to start from noise and gradually get back the image, step by step. It’s like teaching the model to <em>imagine</em> its way from chaos to order.</p><p>This is known as <strong>denoising diffusion</strong>, and the process is usually split into two stages:</p><ul><li>The <strong>forward process</strong>, where noise is added to an image <em>x</em>₀​ at each timestep <em>t</em>, using a small noise factor βₜ</li><li>The <strong>reverse process</strong>, where the model learns to predict and remove that noise to recover <em>x</em>₀</li></ul><p>The model is trained to estimate the mean of a Gaussian that will denoise <em>x</em>ₜ back to <em>x</em>​₀ in <em>t</em> steps. When repeated over a large number of steps (typically 50–100), this process transforms pure noise into a detailed image.</p><p>What makes diffusion models special is their <strong>stability and diversity</strong>. Unlike older GANs (which often collapsed into generating similar images), diffusion models are more controllable and less prone to artifacts (artifacts are small noise in the original images). That’s why they’ve become the standard in most AI image generation pipelines today.</p><h4>Latent Diffusion: Making It Efficient and Scalable</h4><p>Diffusion models are powerful, but they take a lot of time and resources. That’s where <strong>latent diffusion</strong> comes in, and it’s what made tools like <strong>Stable Diffusion</strong> so much better for practical use.</p><p>Here’s the trick: instead of performing diffusion on the actual high-resolution images, Stable Diffusion first compresses the image into a <strong>latent space</strong>. A latent space is more like a smaller representation which basically has the essence of the image, the meaning of the image one can say. And this is done using an autoencoder — a neural network that only encodes and decodes data to a smaller representation.</p><p>So rather than taking a huge 512×512 image (which is expensive), the model now takes a smaller latent map (say, 64×64), denoises it with diffusion, and then decodes it into a full image using a <strong>Variational Autoencoder (VAE)</strong>. The pipeline looks something like this:</p><p><strong>Noise → Encoder → Diffusion in Latent Space → Decoder → Final Image</strong></p><p>Latent diffusion brought high-quality generation to everyone, and made style-specific fine-tuning (like Ghibli filters) actually practical on most of the GPUs.</p><h4>Making the Model Listen: Conditional Generation and CLIP</h4><p>So far we’ve seen about how AI generates images, but how does it know <em>what</em> to generate? If I ask for “a foggy forest in the style of Picasso” or “Totoro eating ramen,” how does the model know what to do?</p><p>That’s where we use <strong>conditional generation</strong>. Instead of generating images randomly, the model is guided using external input, which is usually text. This turns it into a <em>text-to-image</em> model. And to do this we use models like <strong>CLIP</strong>.</p><p>CLIP was introduced by OpenAI and trained on a huge dataset of images paired with text descriptions. The goal was simple: make a model that could understand images <em>and</em> text in the same way — by embedding (i.e., transforming) both into a shared common space.</p><p>So when you enter a prompt like “a castle at sunset, Ghibli style,”</p><ul><li>The text prompt is embedded into a vector using a transformer-based language encoder.</li><li>This embedding becomes the <strong>conditioning signal</strong> that guides the diffusion model during image generation. Like instructing the diffusion model on how to generate the image.</li><li>The model generates images in a way that matches with the meaning of that text. Not just matching objects, but capturing color schemes, composition, and even mood.</li></ul><p>Earlier, CLIP was used in a loop: The image generator would create images and CLIP would score how well they matched the prompt. With Stable Diffusion it goes one step further: they directly condition each step of the denoising process on the text embedding itself. This gives a much finer control over how the final image should look. And that’s why you can get pretty consistent results for very specific prompts.</p><p>Now that’s enough technical stuff for the day. The research is fascinating, and is an ever-growing behemoth. But outside the research papers and GitHub repos, how is all of this landing with people? What does it actually feel like when AI steps into a space like art — a space that’s traditionally been human, personal, and expressive?</p><h3>The Reactions: A Divided Landscape</h3><h4>The Excitement, the Enthusiasm, the Creative Touch</h4><p>There’s a lot of genuine excitement around AI-generated art right now, especially among people who’ve never picked up a paintbrush before. You give it a prompt, or upload a photo, and suddenly you’ve made something beautiful, weird, surreal, or just really fun. For many, that’s more than enough.</p><p><strong>Sam Altman</strong>, CEO of OpenAI, summed it up pretty clearly.</p><blockquote>“AI-generated art is a net win for society. It lowers creative barriers and enables more people to share their ideas freely.” (<a href="https://www.businessinsider.com/sam-altman-openai-studio-ghibli-ai-art-image-generator-backlash-2025-4">Business Insider, 2025</a>)</blockquote><p>In bigger creative industries, AI is being looked at as a way to speed things up: not take over, but make the process smoother.</p><p><strong>James Cameron</strong>, who was previously against the use of AI in film industry, now talks about how it could help behind the scenes:</p><blockquote>“AI can accelerate visual effects workflows without reducing the workforce.” (<a href="https://www.businessinsider.com/james-cameron-ai-blockbuster-films-costs-2025-4">Business Insider, 2025</a>)</blockquote><p>There’s also a growing group of traditional artists who are actually welcoming AI into their process. <strong>Brett Amory</strong>, a painter, talked about how it fits into his creative loop:</p><blockquote>“AI allows me to create intricate feedback loops in my work, blending traditional painting with generative elements.” (<a href="https://www.theverge.com/ai-artificial-intelligence/642599/is-there-a-right-way-to-use-ai-in-art">The Verge, 2023</a>)</blockquote><p>So whether it’s a curiosity, a creative tool, or a source of inspiration, a lot of people are open to exploring what AI can do. They’re not all calling it “art,” but they’re definitely calling it interesting.</p><h4>The Concerns, the Questions, the Unease</h4><p>Not everyone’s excited about where this is going, though. For many artists, especially those who’ve spent years developing their own style, the rise of AI art feels less like a creative breakthrough and more like something to be cautious about.</p><p>The most widely quoted criticism comes from <strong>Hayao Miyazaki</strong>, co-founder of Studio Ghibli — the very studio whose art is now in the form of AI everywhere. When he was shown an AI-generated animation a few years ago, his reaction was clear and deeply personal:</p><blockquote>“I am utterly disgusted. I would never wish to incorporate this technology into my work at all. I strongly feel that this is an insult to life itself.” (<a href="https://www.youtube.com/watch?v=ngZ0K3lWKRc">YouTube, 2016</a>)</blockquote><p>Another concern is around consent, especially when it comes to how these models are trained. Most AI art tools are built on datasets scraped from across the internet, which means they include work by thousands of artists, often without credit, compensation, or even awareness.</p><p><strong>Karla Ortiz</strong>, a professional illustrator whose work has been used without permission, criticized how quickly and carelessly some developers push these tools into the world:</p><blockquote>“OpenAI is exploiting Ghibli’s branding without consent.” (<a href="https://apnews.com/article/artists-ai-image-generators-stable-diffusion-midjourney-7ebcb6e6ddca3f165a3065c70ce85904">AP News, 2024</a>)</blockquote><p>Then there’s the deeper question: <strong>Can something made in seconds, with no human touch, carry the same emotional weight as something crafted slowly, intentionally, and personally?</strong></p><p>Many artists and art lovers aren’t convinced. They argue that art isn’t just about how something looks, it’s about what it means. If anyone can click a button and get a stunning image, does that change how we value effort, skill, or expression?</p><p>Even among those who are open to AI, there’s a bit of unease. The technology is evolving fast, but the conversation around it — about ethics, ownership, and impact — is still catching up. And that tension is only growing.</p><h3>My Thoughts: Somewhere Between Wonder and Responsibility</h3><p>From a research perspective, I find AI-generated art incredibly exciting. The speed of progress, the technical breakthroughs, the creative possibilities; it’s hard not to be amazed by how quickly this field is evolving. We’re watching models go from noisy sketches to emotional, stylistically consistent artwork in just a couple of years. It says a lot about what AI is capable of, and how much more there is to explore.</p><p>That said, I’m not very fond of the recent Ghibli-style trend.</p><p>As someone who deeply respects Studio Ghibli’s work, not just the visuals, but the values behind them and the efforts they put in their work, I can’t ignore the fact that the creators have openly and strongly opposed AI’s use in their art. And yet, here we are, turning our portraits into Ghibli-style animations, sharing them widely, even comparing them to the original films… without really questioning what that means.</p><p>I understand the appeal. It’s fun. It’s cute. It‘s something new. But at what point does “just having fun” turn into something that overlooks the original creator’s intent? I don’t mind people trying it out privately, experimenting, even admiring the results. But when it becomes a viral trend, when it starts being treated as interchangeable with the actual art it’s imitating, that’s where it starts to feel like we’re unintentionally disrespecting the very thing we say we love.</p><p>More broadly, I do believe AI has a place in art. It’s a tool, and like any tool, it can expand what artists can do. It can push creative boundaries, offer new directions, help with ideation, even speed up workflows. I’ve seen artists use it to create some amazing hybrid works that wouldn’t have been possible otherwise.</p><p>But I also believe that <strong>style is not free</strong>. If you’re going to copy another artist’s aesthetic, especially someone working today, there should be permission. There should be credit. There should be respect. Without that, we’re not collaborating with art history, we’re just taking from it.</p><p>These are just my thoughts. A mix of curiosity, concern, and admiration for both the technology and the people behind the art it draws from. I don’t mean to point fingers or criticize anyone for exploring these tools. This is just me voicing my opinions. But I do think it’s worth reflecting on how we use them, what we share, and what it means for the artists whose work inspired it all in the first place.</p><p>I’d love to hear your thoughts too. Whether you agree, disagree, or see things from a completely different angle, I’m here to learn, and to keep the conversation going.</p><h3>References</h3><ol><li>Diffusion model: Ho, Jonathan, Ajay Jain, and Pieter Abbeel. “Denoising diffusion probabilistic models.” <em>Advances in neural information processing systems</em> 33 (2020): 6840–6851. <a href="https://arxiv.org/abs/2006.11239">https://arxiv.org/abs/2006.11239</a></li><li>Stable Diffusion: Rombach, Robin, et al. “High-resolution image synthesis with latent diffusion models.” <em>Proceedings of the IEEE/CVF conference on computer vision and pattern recognition</em>. 2022. <a href="https://arxiv.org/abs/2112.10752">https://arxiv.org/abs/2112.10752</a></li><li>CLIP model: Radford, Alec, et al. “Learning transferable visual models from natural language supervision.” <em>International conference on machine learning</em>. PmLR, 2021. <a href="https://arxiv.org/abs/2103.00020">https://arxiv.org/abs/2103.00020</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8ce90e7de09a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Through the Looking Glass]]></title>
            <link>https://medium.com/@rithesh18.k/ai-through-the-looking-glass-abf0ec2fe1b8?source=rss-87bd15e3fdc3------2</link>
            <guid isPermaLink="false">https://medium.com/p/abf0ec2fe1b8</guid>
            <category><![CDATA[ethics-in-ai-development]]></category>
            <category><![CDATA[technology-and-society]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Rithesh K]]></dc:creator>
            <pubDate>Tue, 01 Apr 2025 22:42:44 GMT</pubDate>
            <atom:updated>2025-06-18T21:57:50.097Z</atom:updated>
            <content:encoded><![CDATA[<h4>A blog series on the Reflections of AI: The Code and Beyond</h4><h3>Blogs</h3><ol><li><a href="https://medium.com/@rithesh18.k/studio-ghibli-meets-ai-8ce90e7de09a">Studio Ghibli meets AI</a></li><li><a href="https://medium.com/@rithesh18.k/stories-in-transition-exploring-ais-role-in-modern-storytelling-0a1c513e240d">Stories in Transition: Exploring AI’s role in Modern Storytelling</a></li><li><a href="https://medium.com/@rithesh18.k/forgetting-by-design-why-machine-unlearning-matters-5aa2f3c7d495">Forgetting by Design: Why Machine Unlearning Matters</a></li><li><a href="https://medium.com/@rithesh18.k/inside-the-mind-of-an-ai-agent-d464a2d0fe18">Inside the Mind of an AI Agent</a></li><li><a href="https://medium.com/@rithesh18.k/marked-by-the-machine-exploring-watermarks-in-llms-d22fe3d8ef77">Marked by the Machine: Exploring Watermarks in LLMs</a></li><li><a href="https://medium.com/@rithesh18.k/the-cracks-in-the-system-jailbreaking-hallucinations-and-the-limits-of-ai-safety-3a7ce713dd53">The Cracks in the System: Jailbreaking, Hallucinations, and the Limits of AI Safety</a></li><li><a href="https://medium.com/ai-advances/when-logic-meets-learning-exploring-neuro-symbolic-ai-d55d53a1c95c">When Logic Meets Learning: Exploring Neuro-Symbolic AI</a></li></ol><h3>About the Series</h3><p>Last year, I started writing a blog series to help share what I’d learned about Machine Learning. Like a lot of side projects, it had to take a backseat to academics and other priorities. But over time, I’ve also come to realize something else: there are already more than enough tutorials and resources to learn ML and AI. From courses and tutorials to YouTube explainers, the gap in “how to learn ML” is closing quickly.</p><p>That said, I do plan to return to that original series someday. There’s still value in breaking down foundational concepts in a simple and structured way. But for now, my attention is shifting to a different kind of conversation — not about <em>how</em> to build AI, but about how we <em>see</em> it.</p><p>What we need right now, in my opinion, are thoughtful conversations about where AI is headed. What’s changing, what’s trending, what the breakthroughs actually mean, and how people — researchers, artists, founders, and the general public — are thinking about it all. These are conversations I have with myself often, and now I want to bring them together in a more interactive space.</p><p>That’s my motivation for this new series: <strong>AI Through the Looking Glass</strong>.</p><h3>Why “AI Through the Looking Glass”?</h3><p>The name is inspired by <em>Through the Looking-Glass</em>, the Lewis Carroll sequel to <em>Alice’s Adventures in Wonderland</em>. In the story, Alice steps through a mirror into a world that appears familiar — but is full of oddities, reversed logic, and unexpected perspectives.</p><p>That’s exactly the feeling I want this series to evoke.</p><p>To me, <em>the looking glass</em> represents our everyday interaction with AI, something that we are increasingly familiarizing ourselves with. We see it in our apps, our social media feeds, our photos, and pretty much everywhere now. We use it, we’re amazed by it, and often, that’s where our curiosity ends.</p><p>But if we pause and look a little deeper, we start to uncover an entire landscape of thoughts, opinions, dilemmas, and ideas that don’t always show up at first glance. Questions that don’t have simple answers. Moments that make us go <em>“Oh, I hadn’t thought of it that way.”</em></p><p>This blog is about stepping into that space. About exploring what lies beneath the surface of AI — the biases, the beauty, the strange contradictions. It’s about wandering a bit further than the headlines, and asking whether things are really as simple and straightforward as they seem.</p><h3>What to Expect</h3><p>Each post in this series will take one topic that’s currently shaping conversations around AI. It could be something widely discussed or something quietly emerging, but the aim is always to explore it from a perspective that doesn’t usually get the spotlight.</p><p>Here’s how I plan to approach each piece:</p><p><strong>Pick a topic:</strong> I’ll choose something that feels timely, relevant, and layered. Something that’s important or in the buzz right now.</p><p><strong>Do some research:</strong> I’ll dive into recent papers, articles, posts, or products related to the topic. The goal isn’t to summarize everything that’s out there, but to find angles that are thought-provoking and grounded in current developments. I’ll make sure to quote all the sources I get the information from.</p><p><strong>Talk to people:</strong> I’ll try to include opinions from others in the AI space — researchers, engineers, artists, and anyone whose work intersects with the topic. Sometimes this might come from attending local meetups or events, and other times through short conversations.</p><p><strong>Write it all out:</strong> The post itself will be a blend of everything, what I’ve read, what I’ve heard, what I think. It will be slightly more technical than a news article, but still understandable. You don’t need to be an AI expert to follow along.</p><p><strong>End with reflection:</strong> I’ll close each post with my own thoughts: what resonated with me, what surprised me, and what I’m still unsure about.</p><h3>Join the Conversation</h3><p>One of the main reasons I’m starting this series is to hear from others. Not just to write about what I think, but to learn how others are seeing and interpreting AI in their own work, fields, and day-to-day lives.</p><p>If something in a post sparks a thought, raises a question, or makes you want to agree (or disagree), I’d love to hear it. You can leave a comment, send me a message, or share your own perspective on the topic, whatever feels right.</p><p>At the end of each post, I’ll also mention the topic I plan to explore next. If you have any thoughts, resources, or experiences related to it, feel free to send them my way. And if there’s a topic you think deserves more attention, I’d love to hear your suggestions. This series is meant to grow through conversation.</p><p>I hope this becomes a space for thoughtful, curious dialogue. A place to pause, reflect, and better understand the world of AI as it continues to unfold around us.</p><p>This is my first time trying something like this, and I’m figuring it out as I go. So if you have thoughts on the format, the writing, or even whether this is interesting to read, I’d genuinely love to hear it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abf0ec2fe1b8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>