<?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 Ashish K Saxena on Medium]]></title>
        <description><![CDATA[Stories by Ashish K Saxena on Medium]]></description>
        <link>https://medium.com/@MindbytesAI?source=rss-e5a9853bb3d4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*d7jMgb3xFubm24A1tDtw5A.png</url>
            <title>Stories by Ashish K Saxena on Medium</title>
            <link>https://medium.com/@MindbytesAI?source=rss-e5a9853bb3d4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 05:23:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@MindbytesAI/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[The five controls: what an examiner can actually open]]></title>
            <link>https://medium.com/@MindbytesAI/the-five-controls-what-an-examiner-can-actually-open-5da67617ea66?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/5da67617ea66</guid>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Wed, 05 Aug 2026 00:00:21 GMT</pubDate>
            <atom:updated>2026-08-05T08:31:51.161Z</atom:updated>
            <content:encoded><![CDATA[<p>Eight principles and no checklist. Here is each of the five controls in full, with the artifact it produces, the failure it prevents, the build order nobody tells you and the drill that proves the whole thing works.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7bZRX5u4ru34azu_.jpg" /></figure><p>Five controls, five artifacts. The third one is the only one you cannot build after the fact.</p><p>The April 2026 guidance left you eight principles and no checklist. I covered that in <a href="https://www.caventia.com/blog/footnote-3-ai-agents-model-risk">Footnote 3</a>, along with five controls that turn those principles into evidence.</p><p>Five controls in five paragraphs. That was the summary. This is the version you can hand to the person who has to build them.</p><p>One rule runs through all five. A control isn’t what you intend to do. A control is a file an examiner can open, on a Tuesday, without you in the room.</p><p>If a control on your list doesn’t terminate in an artifact, it’s a policy. Policies are fine. They’re just not evidence.</p><h3>1. Validate the behavioral envelope, not the output</h3><p>Traditional model validation rests on replay. Same input, same output, six months later.</p><p>LLMs break that, and not only because of sampling temperature. Cross a model version or an infrastructure boundary and the output moves.</p><p>So stop validating the output. Validate the envelope around it.</p><p><strong>What the envelope is.</strong> A stratified set of inputs plus a set of assertions over the responses. Three kinds of assertion, and you need all three:</p><ul><li><strong>Hard constraints.</strong> The agent never quotes an APR outside the product table. Never promises a decision it isn’t authorized to make. Always emits a reason code on an adverse action. These pass or fail per case. One failure is a release blocker.</li><li><strong>Distributional constraints.</strong> Recommendation rates by segment stay within a preset tolerance of baseline. This is where fair lending exposure shows up first, and it only works if your strata are large enough to move. A segment with 30 cases swings three points on one flipped decision, which means you’re measuring noise and calling it drift.</li><li><strong>Semantic constraints.</strong> Scored against a written rubric, on a sample, by people. Slow and unglamorous, and it catches the things assertions never do.</li></ul><p><strong>Set the tolerances before the run.</strong> Write them down, get them signed, then execute. Thresholds chosen after you see the numbers aren’t a control. They’re a narrative with a spreadsheet attached.</p><p><strong>The artifact.</strong> A per-release envelope report: every assertion with its result, the model version and configuration hash it ran against, plus the diff versus the previous release. The diff is the part people skip and the part an examiner reads first.</p><p>Discharges: conceptual soundness, outcomes analysis, ongoing monitoring.</p><h3>2. Treat the prompt as a versioned model component</h3><p>Everyone gets half of this right. They version the system prompt, pin the hash to the deployment and feel covered.</p><p>Then someone refreshes the retrieval index. Behavior moves while the prompt hash sits unchanged, and the control reports green through the entire incident.</p><p>The prompt isn’t the unit of change. The resolved decision context is. Hash all of it:</p><ul><li>System prompt template and every variable that fills it</li><li>Tool definitions, including the JSON schemas and the descriptions, because the description is what drives selection</li><li>Retrieval configuration: index version, embedding model, chunking, top-k, filters</li><li>Decoding parameters: temperature, top_p, max tokens, seed</li><li>Model id and version</li><li>Post-processing and guardrail rules</li></ul><p>That composite is your configuration hash. One value. It goes in the deployment manifest, and it goes on every single decision record you write.</p><p><strong>Then put prompt edits under change control.</strong> Plenty of banks now run a prompt console that lets a product owner edit production behavior in an afternoon. That’s a model change with a friendlier interface.</p><p>Same review, same approval, same record. If your change log shows six months of model stability and your prompt console shows 40 edits, you don’t have a change log.</p><p><strong>The artifact.</strong> A deployment manifest per release listing every field above, and a decision log where each row carries the hash that produced it. “What changed between March and June” stops being a memory exercise and becomes a diff.</p><p>Discharges: documentation, governance and controls.</p><h3>3. Make the tool-call trace an audit artifact</h3><p>You can’t enumerate every path an agent will take before it takes it. That was true when SR 11–7 stood and it’s truer now. Give up on prediction and get serious about reconstruction.</p><p><strong>Capture per step:</strong> timestamp, decision id, configuration hash, step index, tool name, arguments, result, latency, token counts, stop reason. Then the final action and what downstream system consumed it.</p><p>Three details separate a real trace store from an expensive log pile.</p><p><strong>Retention.</strong> Most observability vendors default to about 30 days. Regulation B puts consumer credit application records at 25 months, and if your trace is part of how the decision got made, you’ll have a hard time arguing it isn’t part of the record.</p><p>Litigation reaches further still. If the trace expires before the exposure does, the control was decorative.</p><p>So price the storage against the retention you actually owe, and do it before you instrument. Retrofitting retention onto data you already deleted isn’t a project.</p><p><strong>Tamper evidence.</strong> A log an insider can edit isn’t evidence, and everyone in the exam room knows it. Hash-chain the records or write to append-only storage.</p><p>The point isn’t that you distrust your engineers. The point is that you can demonstrate integrity without asking anyone to take your word.</p><p><strong>Redaction at instrumentation time.</strong> Decide field by field what gets captured, hashed or dropped, before the first trace is written. Teams that skip this discover in month four that their audit trail is full of data they aren’t allowed to retain, and then they delete the trail.</p><p><strong>The artifact.</strong> One decision, one page, reconstructible on demand: what came in, which tools fired in what order with what arguments, what came back, what the agent did, under which configuration.</p><p>Discharges: ongoing monitoring, documentation and the access that makes effective challenge possible at all.</p><h3>4. Document your model and the vendor model separately</h3><p>You’ll never author a model card for a frontier model. Stop treating that as a gap in your documentation and start treating it as a boundary in your documentation.</p><p><strong>Layer one, the vendor model.</strong> What you can actually obtain: the published model card and evals, the terms covering training use and data retention, the deprecation policy, region and subprocessor list, incident history. Owner: vendor management. Refresh trigger: any provider version change.</p><p><strong>Layer two, your model.</strong> The prompts, the tool inventory, the retrieval corpus and how it gets updated, the post-processing, the guardrails, the human review points and an explicit statement of what the system is authorized to decide and what it isn’t.</p><p>This layer is fully documentable. It’s also where your proprietary work actually lives.</p><p><strong>Classify every tool by blast radius.</strong> Read-only, writes to a system of record, moves money or speaks to a customer. Sort your tool inventory by that column and you’re looking at your incident list in advance. The failures that reach a regulator tend to be write-tool failures, not reasoning failures.</p><p><strong>The seam is the risk.</strong> The interesting exposure isn’t inside either layer. It’s in what you assume the vendor model does and never tested.</p><p>Write those assumptions down as assumptions. That page ages better than anything else in the pack.</p><p>Two documents, two owners, one inventory entry, one named accountable individual. The name matters more than it used to. The replacement guidance never mentions your board of directors, so the accountability you get is the accountability you assign.</p><p>Discharges: model inventory, documentation, vendor oversight, aggregate risk.</p><h3>5. Pin validation evidence to a model version</h3><p>“The agent has been validated” is a claim, not a statement. Validated against what, and until when?</p><p><strong>Pin the full coordinate.</strong> Model id, version and endpoint or region. Alias targets that resolve to whatever shipped last night are the specific enemy here. Your pin should be a value someone could type back in a year and get the same thing.</p><p><strong>Negotiate while you still have leverage</strong>, which is before the renewal and long before the incident: advance notice of deprecation, notice of behavior-affecting changes, a defined window to stay on a pinned version, access to eval artifacts.</p><p><strong>Write the re-validation trigger list and make it boring.</strong> Any one of these fires it:</p><ul><li>Provider version change</li><li>Configuration hash change</li><li>Retrieval corpus refresh past a defined threshold</li><li>Tool added or its scope widened</li><li>New product or population</li><li>Drift alarm firing</li><li>Calendar floor expiring, so nothing sits unvalidated forever on the grounds that nothing happened</li></ul><p><strong>Track the expiry.</strong> Providers retire versions on their own schedule, which means your pin has an end date. On that date you get pushed onto an unvalidated version whether or not your validators have capacity. Put the date in the risk register the day you pin, not the week it lands.</p><p><strong>And give the challenger real power.</strong> Effective challenge means someone can fail the thing. Independent reporting line, access to the trace store, budget and the authority to block a release.</p><p>Without that last one it’s review theater, and it reads that way from across the table.</p><p><strong>The artifact.</strong> A one-page validation certificate: what was validated, against which configuration hash and model version, on what data, by whom, what the challenge found, what limitations remain and the expiry date.</p><p>Discharges: validation, effective challenge, governance and controls, vendor oversight, aggregate risk.</p><h3>How the five map to the eight</h3><p>Every principle terminates in a file. That’s the whole design.</p><h3>Build order, which isn’t the numbered order</h3><p>Footnote 3 numbered these one through five and said I’d build them in that order. If you’re starting from zero, don’t. The reason is worth a paragraph.</p><p><strong>Controls 2 and 3 first.</strong> The configuration hash and the trace are instrumentation. They only describe decisions made after you turn them on.</p><p>Every week you wait is a week of production history you can never reconstruct, whatever you spend later.</p><p><strong>Controls 1, 4 and 5 after.</strong> Envelope tests, two-layer documentation and version pinning can all be built retroactively against a system already running. Late is expensive. It isn’t impossible.</p><p>So the sequencing rule is simple. Anything that captures the present goes first. Anything that can look backward can wait.</p><h3>Scale it, but don’t skip it</h3><p>Model materiality survived into the new guidance, so use it. A low-exposure internal drafting assistant doesn’t need the same envelope depth, the same challenge budget or the same retention posture as an agent touching consumer credit decisions.</p><p>Scale four of the five. Don’t scale control 3 to zero.</p><p>The trace is the one control whose absence you can’t remediate after the fact, and the low-materiality system that quietly acquires a write-scoped tool is exactly how a small exposure becomes a large one.</p><h3>The drill that tells you if this is real</h3><p>Pick a decision your agent made six months ago. Pick it at random, not from the demo set.</p><p>Hand it to someone who didn’t build the system. Ask them to reconstruct what happened and why, using only what’s in the record. Start a timer.</p><p>If they come back in an hour with the input, the trace, the configuration hash, the validation certificate that covered that hash and the documentation for both layers, your five controls are real.</p><p>If it takes three weeks and four engineers, you have five intentions.</p><p>The gap between those two outcomes is the entire discipline now. The agencies stopped telling you what enough looks like. The record you kept is the only remaining answer.</p><p>The full framework, with templates and an implementation roadmap, is in the <a href="https://www.caventia.com/sr117">model risk whitepaper</a>. The FDA 510(k) equivalent for clinical AI is in the <a href="https://www.caventia.com/fda">FDA whitepaper</a>. Both are free.</p><p><em>Originally published at </em><a href="https://www.caventia.com/blog/five-controls-examiner-can-open"><em>https://www.caventia.com</em></a><em> on August 5, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5da67617ea66" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Footnote 3: the sentence that took AI agents out of the model risk rulebook]]></title>
            <link>https://medium.com/@MindbytesAI/footnote-3-the-sentence-that-took-ai-agents-out-of-the-model-risk-rulebook-7728a9cfec1c?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/7728a9cfec1c</guid>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Thu, 30 Jul 2026 00:00:00 GMT</pubDate>
            <atom:updated>2026-07-30T20:52:24.166Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5-KPwXr4ELAF0FgY" /></figure><h3>Footnote 3: the sentence that took AI agents out of the model risk rulebook</h3><p>The April 2026 guidance rescinded SR 11–7 and then excluded generative and agentic AI from its own scope. Out of scope is not off the hook. Here is what survived, what quietly vanished, and the five controls that turn eight principles into evidence.</p><p>On April 17, 2026, the Federal Reserve, the OCC and the FDIC jointly rescinded SR 11–7.</p><p>They rescinded OCC 2011–12, FIL-22–2017 and FIL-27–2021 along with it. Fifteen years of model risk doctrine, replaced by a 12-page principles-based document titled “Supervisory Guidance on Model Risk Management”.</p><p>If you run model risk at a bank, you know this already. What you may not have read closely is footnote 3 on page 3. It is the most consequential sentence in the document and it sits in small type.</p><p>Here it is, verbatim:</p><blockquote><em>“Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance. Nonetheless, a banking organization’s risk management and governance practices should guide the determination of appropriate governance and controls for any tools, processes, or systems not covered in this document. However, the principles described in this guidance apply to traditional statistical and quantitative models and non-generative, non-agentic AI models.”</em></blockquote><p>Read it twice. The agencies looked at generative and agentic AI, decided the category was moving too fast to write rules against, and handed the problem back to you.</p><h3>Out of scope is not off the hook</h3><p>The temptation is obvious. Your gen AI agents are excluded from the guidance, so the model risk team stops asking about them and the deployment queue speeds up.</p><p>That reading survives about one page. Footnote 1, page 2, is the counterweight:</p><blockquote><em>“supervisory action may result for any violations of law or unsafe or unsound practices stemming from insufficient management of model risk.”</em></blockquote><p>So the guidance is non-prescriptive. The replacement text is explicit that “non-compliance with this guidance will not result in supervisory criticism against a banking organization”. But ECOA did not get rescinded. UDAAP did not get rescinded. The safety and soundness standard did not get rescinded.</p><p>What changed is that you no longer get a checklist telling you what “enough” looks like. You still get held to the outcome.</p><p>That is a harder position, not an easier one. A prescriptive rule is also a defense. When an examiner challenges your agent and you can point to the paragraph you satisfied, the argument ends. Now there is no paragraph. There is only the record you kept.</p><h3>What quietly disappeared</h3><p>Put the old and new documents side by side and the absences are louder than the additions.</p><p>Gone: the three lines of defense terminology. Gone: any explicit mention of ECOA and fair lending. Gone: the board reporting prescription. Gone: quantitative tier thresholds. Gone: the model risk appetite framework.</p><p>None of those obligations disappeared from your business. Only the instructions for meeting them did.</p><p>There is one more line worth flagging. The new guidance defines a model in a way that excludes “deterministic rule-based processes and software where there are no statistical, economic, or financial theories underpinning their design or use”.</p><p>Think about what that does to an agentic workflow. The LLM call may be a model. The tool-calling harness around it, the retrieval step, the routing logic, the post-processing rules? An aggressive reading puts most of that outside the model definition entirely, which means outside your model inventory, which means outside validation.</p><p>Your agent’s decision does not care where the definition ends. If the harness picks the wrong tool, the customer still gets the wrong answer.</p><h3>What survived, and why it is now your only map</h3><p>The principles that made it into the replacement guidance are the ones worth building against, because they are what an examiner will reason from when there is nothing more specific to cite.</p><ul><li>Model inventory</li><li>Documentation</li><li>Validation, split into conceptual soundness, outcomes analysis and ongoing monitoring</li><li>Effective challenge</li><li>Governance and controls</li><li>Vendor oversight</li><li>Aggregate risk</li><li>Model materiality, scaled by exposure and purpose</li></ul><p>Eight principles. That is your map now. Every one of them applies cleanly to an AI agent, and not one of them tells you what evidence to produce.</p><h3>The five controls that turn principles into evidence</h3><p>Here is what I would build, in order, for any gen AI or agentic system you would be embarrassed to explain in an exam.</p><p><strong>1. Validate the behavioral envelope, not the output.</strong> Traditional models replay exactly. LLMs do not, even at temperature zero, once you cross a model version or an infrastructure deployment. Define the set of outputs the agent should produce across a representative input distribution and test against that envelope. This is how you satisfy conceptual soundness and outcomes analysis when reproducibility is off the table.</p><p><strong>2. Treat the prompt as a versioned model component.</strong> Your system prompt is a feature. Hash it, version-control it, and pin a specific hash to every production deployment. When someone asks what changed between March and June, you answer with a diff instead of a memory.</p><p><strong>3. Make the tool-call trace an audit artifact.</strong> You cannot enumerate every execution path an agent will take. Stop trying. Capture every tool call with its arguments and results in production, so validation shifts from predicting behavior to reconstructing it. When the question is “how did this agent reach this decision”, you produce the trace.</p><p><strong>4. Document your model and the vendor model separately.</strong> You cannot produce a model card for GPT-4o or Claude. You can produce a complete one for the prompts, tools, retrieval architecture and post-processing logic, which is also where your actual proprietary work lives. Two documents, two owners, one inventory entry. That is vendor oversight in practice.</p><p><strong>5. Pin validation evidence to a model version.</strong> “The agent has been validated” is not a statement until you say validated against what. Require re-validation before provider auto-upgrades, and negotiate behavior-change notification into the contract while you still have leverage.</p><p>Read those back against the eight principles. Envelope tests cover conceptual soundness. Prompt hashes and traces cover documentation and ongoing monitoring. Two-layer documentation covers vendor oversight. Version pinning covers effective challenge, because a challenger who cannot tell which version they are challenging is not challenging anything.</p><p>The longer form of each control, with the specific failure it addresses, is in <a href="https://www.caventia.com/blog/five-places-sr11-7-breaks-down">the five places SR 11–7 breaks down on AI agents</a>. That post was written five days before the rescission and the mechanics held up.</p><h3>If you are under $30 billion</h3><p>The new guidance says it is “most relevant” to banking organizations above $30 billion, and generally does not apply below that unless the institution has significant model risk exposure.</p><p>That threshold is about this document. It is not about your exposure.</p><p>A $6 billion bank running an agent on consumer lending decisions has the same ECOA obligation as a $600 billion bank. State regulators have their own view. The CFPB has its own view. And a plaintiff’s attorney reading your discovery has never once been impressed by a relevance threshold in supervisory guidance.</p><h3>Where this actually lands</h3><p>For fifteen years the model risk function had an answer to “how much is enough”. SR 11–7 was that answer, and its specificity was the point.</p><p>The agencies just replaced specificity with judgment, then carved out the fastest-moving category of models entirely. The burden of proof moved from the regulator’s document to your evidence trail.</p><p>So the question worth asking this quarter is not whether your AI agents are in scope. It is whether you could reconstruct, today, what one of them decided six months ago and why.</p><p>If the room goes quiet when you ask that, you already know where to start.</p><p>The full framework, with templates and an implementation roadmap, is in the <a href="https://www.caventia.com/sr117">model risk whitepaper</a>. The FDA 510(k) equivalent for clinical AI is in the <a href="https://www.caventia.com/fda">FDA whitepaper</a>. Both are free.</p><p>If your gauntlet looks different in some specific way, that is the kind of conversation the founder takes by email.</p><p><em>Originally published at </em><a href="https://caventia.com/blog/footnote-3-ai-agents-model-risk"><em>https://www.caventia.com</em></a><em> on July 30, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7728a9cfec1c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An Audit Trail Is Not a Screenshot]]></title>
            <link>https://medium.com/@MindbytesAI/an-audit-trail-is-not-a-screenshot-d41b37746837?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/d41b37746837</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[banks]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[finance]]></category>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Fri, 15 May 2026 16:50:56 GMT</pubDate>
            <atom:updated>2026-05-15T16:50:56.499Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>Why the way banks are documenting AI agent decisions today will not survive an examiner walking in the door.</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2JRPqsCmD_f8z0VeqILAdw.jpeg" /></figure><p>Last quarter a Chief Model Risk Officer at a $50 billion regional bank told me a story. Their KYC agent denied a loan application. Three weeks later the Office of the Comptroller of the Currency sent a routine examination request. Show us how this decision was reached.</p><p>The bank’s second line of defense pulled the agent’s chat log. They sent it over.</p><p>Here’s the thing. That wasn’t an audit trail. That was a screenshot.</p><p>A screenshot tells you what happened. An audit trail tells you why it happened, and lets a regulator verify the answer without trusting your story.</p><p>That distinction is the whole game.</p><h3>The screenshot trap</h3><p>You’ve probably seen the workflow. An AI agent processes something important. It returns text.</p><p>Somebody saves the input and the output to a database. When a regulator asks for the record, you produce the input, the output and maybe the date. That’s the record.</p><p>If you’re at a bank running SR 11–7 model validation, that record fails the basic test. The framework asks: how do we know this decision was conceptually sound? How do we know the model behaved the way the validators approved? How do we know nothing was tampered with after the fact?</p><p>A chat log can’t answer any of those questions. It tells you the words went in and the words came out.</p><p>It doesn’t tell you which version of the model was running. It doesn’t tell you which tools the agent had available, what context it loaded, what policy boundary was active, what the override path looked like. None of that survives.</p><p>Think about how a doctor’s chart works. It is not a transcript of what the patient said. It is a structured record of what the doctor heard, what they ordered, what they ruled out and why.</p><p>The transcript would be useless to a malpractice reviewer. The chart is the artifact that survives the audit.</p><p>The chat log is the transcript. The audit trail is the chart.</p><p>A regulator’s first question is almost never “what did the AI say”. It is “could anything have stopped it from saying that, and did anyone notice when it did?”</p><p>You can’t answer the second question with a chat log.</p><h3>What lives inside an actual audit trail</h3><p>If you wrote down everything a model risk officer needs to reconstruct one AI decision, here is the shape of it.</p><p>The input the agent saw. Not just the user’s question. The full message history, the tools available at that moment, the system prompt, the context window, the retrieval results that landed in front of the model.</p><p>The model itself. Provider, exact version, the parameters at call time (temperature, top-p, max tokens) and a hash of the system prompt so you can prove the prompt has not drifted since validation.</p><p>The action surface. What tools could the agent call? Which did it actually call, in what order, with what arguments? What did the tools return? An agent with five tools is a different model than the same agent with three. The action space is part of the artifact.</p><p>The decision boundary. What policy was in force? What rules were active? What would have escalated to a human? Most banks have these written in a Confluence page nobody updates. The audit trail has to capture which version was actually running.</p><p>The output. Not just the final answer. The full content the model produced, including tool calls and reasoning when the provider exposes it. The finish reason. Token counts. Error states.</p><p>And finally, the chain. Each record cryptographically linked to the one before it. If anyone modifies any field of any record, the chain breaks visibly. The break is the evidence.</p><p>That’s the artifact. Anything less is a screenshot.</p><h3>How we got here</h3><p>You might be wondering why banks don’t already have this. They run model risk frameworks. They run audit programs.</p><p>They have hundreds of compliance people. Why is the artifact missing?</p><p>Two answers. First, AI agents are new. The model risk frameworks at most US banks were written for fixed artifacts. A neural net you train, validate, deploy and run for two years.</p><p>The framework assumes the model does not change between validation and inference. Agents change every turn.</p><p>Second, the tooling did not exist. Most of the “AI governance” products you can buy today were built for the old world.</p><p>They watch model outputs. They flag drift. They show accuracy dashboards. They give you observability into what the model is doing right now.</p><p>They do not produce evidence for what the model did six weeks ago, in a form that survives an examiner.</p><p>The gap is not malicious. It is structural. The category nobody built for is compliance evidence, not model observability.</p><h3>The 2026–2027 forcing function</h3><p>Here is what is coming, in case the runway feels longer than it is.</p><p>Every quarter, more AI agents go into production at regulated institutions. Every quarter, examiners get a little more specific about what they want to see.</p><p>The OCC has signaled they will increase scrutiny on agentic AI under existing SR 11–7 authority. The FDA is moving on diagnostic AI under 510(k) and the Clinical Decision Support carve-out. NAIC is moving on insurance. State financial regulators are watching the federal ones.</p><p>None of these regulators are going to wait for a new framework. They are going to apply the existing one to the new technology and see what breaks.</p><p>What breaks is the screenshot.</p><p>Banks that have been deploying AI agents for the past eighteen months are about to face their first serious examination of that technology. They will get asked to reconstruct decisions. The ones who can will look professional and prepared. The ones who cannot will spend the next six quarters on a remediation track nobody enjoys.</p><p>This is not a 2030 problem. It is a 2027 problem at the latest. For some institutions it is a 2026 problem.</p><p>If you are at one of those institutions, you already know which side you are on.</p><h3>The cryptographic angle, in one paragraph</h3><p>You do not have to understand the math to use the system. But the mechanism is simple enough to explain.</p><p>Every record gets a SHA-256 hash of its own content. The next record stores that hash in its previous-hash field. This forms a chain.</p><p>Tamper with any field of any record and the hash does not match the content anymore. The next record’s previous-hash does not match either. The break is detectable in two ways at once.</p><p>Every day, all the day’s records roll up into a Merkle tree. The root is signed and archived to tamper-evident storage. An examiner can take any record, walk the proof up to the root and verify it existed at a specific date without trusting the vendor that produced it.</p><p>That is the whole tamper-evidence layer. It runs in the background. The bank does not have to think about it.</p><p>The point is not the cryptography. The point is that the evidence holds up when scrutinized by someone with a budget for forensic accountants.</p><h3>Why this is a product, not a feature</h3><p>A reasonable engineer at a bank might read this and think: we should just build it.</p><p>You can. It is a non-trivial chunk of work. You need the SDK that wraps every LLM provider. You need the canonical schema that survives a schema review by your second line. You need the storage layer with append-only enforcement at the database role level, not just at the application level.</p><p>You need the hash chain with proper canonicalization. You need the daily Merkle anchor and the key management ceremony for the signing keys. You need PII redaction before transport, with built-in patterns for SSN, MRN, NPI and the long tail of identifiers your specific industry cares about.</p><p>You need the dashboard a model risk officer will actually open. You need the verification tool an examiner can run independently of you.</p><p>Most banks will not build all of that. The ones who try will end up with a slightly different version that does not interoperate with anyone else’s, will not survive a vendor change and will not be examiner-verifiable without trusting the bank itself.</p><p>That is why this is a product. Specifically, it is the product I have spent the last year building.</p><h3>What examiner-ready actually looks like</h3><p>Here’s the test. Imagine the OCC sits down at your dashboard tomorrow. They pick one decision the agent made last Wednesday. They want the input, the tools, the policy boundary, the override path and proof nothing’s been changed since.</p><p>How long does it take you to produce that?</p><p>Under thirty seconds is examiner-ready. Under five minutes is acceptable. Anything longer means the artifact was not designed for the question.</p><p>If it takes you a week, you are producing the artifact for the first time. That is the screenshot problem.</p><p>You captured chat logs. Now you have to reconstruct everything else from memory, Confluence pages, deployment manifests and Slack threads.</p><p>You will get most of it. You will not get all of it. The pieces you cannot reconstruct become the pieces the examiner asks about.</p><p>When you have the audit-trail spine in place, the answer is immediate. You click the decision in the dashboard. You see every component. The examiner clicks the verification button on their side and checks the proof against the Merkle root themselves.</p><p>That is the bar. That is what every regulated AI deployment needs to clear by 2027.</p><h3>Where this goes from here</h3><p>Caventia is the audit-trail spine for regulated AI agents. SR 11–7 for banks, FDA 510(k) for health systems. We capture every prompt, every tool call, every decision in a form that survives an examiner walking in the door.</p><p>The SDK runs in your process. The records ship to an immutable, hash-chained store. The daily Merkle anchor makes the trail third-party verifiable.</p><p>If you are at a regulated institution running AI agents in production today, the SR 11–7 whitepaper at caventia.com/sr117 goes deeper into the five places the framework specifically breaks. Free, no signup. Nine short animations explain the gaps.</p><p>If you are a model risk officer, a Chief AI Officer or anywhere on the second line of defense, and you are reading this because something here landed, I want to hear from you. Ten design-partner slots are open. Reply or find me on LinkedIn.</p><p>The next examiner who walks into your office is going to ask a question your current tooling cannot answer. The right time to fix that is before they ask.</p><p><em>cavēre necesse est.</em> Caution is necessary.</p><p>— -</p><h4>About</h4><p>Ashish K. Saxena is the founder of <a href="https://caventia.com/">Caventia, Inc</a>. Fifteen years deploying AI inside large institutions. Two Amazon-bestselling books on AI ethics. Forty-two peer-reviewed papers as a reviewer. h-index 8 on Google Scholar.</p><p>Caventia is the audit-trail spine for regulated AI agents. SR 11–7 for banks. FDA 510(k) for health systems. Ten design-partner conversations open.</p><h4>Read further</h4><p>- <a href="https://caventia.com/">caventia.com</a> · the platform<br>- <a href="https://caventia.com/sr117">caventia.com/sr117</a> · the SR 11–7 whitepaper, free, no signup</p><p>If this piece reframed how you think about AI audit evidence, the <a href="https://caventia.com/sr117">SR 11–7 whitepaper</a> goes ten times deeper. Nine short animations explain the gaps. Free. No signup.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d41b37746837" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What happens when you compare 3,000-year-old fasting traditions to cutting-edge cellular biology?]]></title>
            <link>https://medium.com/@MindbytesAI/your-grandmothers-fasting-ritual-and-the-2016-nobel-prize-are-telling-the-same-story-3c3dfe573417?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c3dfe573417</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[health]]></category>
            <category><![CDATA[intermittent-fasting]]></category>
            <category><![CDATA[aging]]></category>
            <category><![CDATA[biohacking]]></category>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Sun, 12 Apr 2026 04:43:33 GMT</pubDate>
            <atom:updated>2026-04-12T05:07:21.196Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IspPr33nO6ewG88TB9AACA.png" /></figure><h3>What happens when you compare 3,000-year-old fasting traditions to cutting-edge cellular biology? The overlap is uncomfortably specific</h3><p>In 2016, Yoshinori Ohsumi won the Nobel Prize in Medicine for figuring out how cells eat themselves.</p><p>That sounds grim, but it is actually how you stay alive. The process is called <strong>autophagy</strong>, from Greek for “self-eating” and it is the cellular equivalent of taking out the trash. Your cells build tiny double-membrane bubbles that engulf damaged proteins, broken mitochondria, and miscellaneous junk, then deliver the whole package to a lysosome for recycling. When autophagy works well, your cells stay clean. When it slows down, the junk piles up, and you age faster.</p><p>Here is the part that made me sit up straight. The single most reliable way to trigger autophagy in a human being is fasting. Not a pill, not an injection, not a supplement. Just not eating for long enough that your cells get the message that external food is scarce, and switch to eating their own broken parts instead.</p><p>Now here is the part that made me write a 32-page academic paper. My grandmother knew this.</p><p>Not in so many words. She did not talk about mTOR or AMPK or LC3-II ratios. But every two weeks, on the eleventh day of the lunar fortnight, she would fast for 24 to 36 hours. It was called Ekadashi, and it has been part of Hindu devotional practice for somewhere between one and three thousand years, depending on which text you trust. She thought of it as purification. The Nobel committee would call it peak autophagy induction.</p><p>I kept running into coincidences like this, and eventually I could not ignore them.</p><h4>The concept that broke me</h4><p>In Ayurveda, the traditional medicine system of India, there is a concept called <strong>Ama</strong>. It translates roughly as “undigested metabolic waste”. Classical texts from around 200 BCE describe Ama with three specific properties:</p><p>- <strong>Picchila</strong>: sticky, viscous<br>- <strong>Guru</strong>: heavy<br>- <strong>Srotorodha</strong>: channel-obstructing</p><p>Ama, the texts say, accumulates when the body’s metabolic fire (<strong>Agni</strong>) weakens. It accumulates with age. It causes inflammation, weakness, and systemic dysfunction. The prescribed treatment is fasting.</p><p>Now open a modern cell biology textbook and look up proteostasis failure, the age-related collapse of the cell’s protein quality control system. What you find is a description of three main culprits:</p><p>- <strong>Amyloid fibrils</strong>, which aggregate through hydrophobic interactions. Literally sticky.<br>- <strong>Lipofuscin</strong>, an insoluble aggregate of oxidized proteins that accumulates as dead weight inside aged neurons. Literally heavy.<br>- <strong>Advanced glycation end-products</strong>, cross-linked proteins that physically obstruct molecular trafficking through tissues. Literally channel-obstructing.</p><p>These substances accumulate when autophagy weakens. They accumulate with age. They cause inflammation, tissue dysfunction, and systemic decline. The cleanest way to clear them is fasting-induced autophagy.</p><p>Read those two paragraphs again. The correspondence is not vague. It is point-for-point, property-for-property, mechanism-for-mechanism. Someone writing in Sanskrit twenty-two centuries ago gave a phenomenological description that lines up, uncomfortably well, with what we now see through an electron microscope.</p><p>I am not going to tell you that ancient practitioners understood molecular biology. They did not. But I am going to tell you that the pattern they described, the thing they pointed at with the word Ama, appears to be the same thing we now point at with the word proteostasis failure. Different epistemologies, different vocabularies, same underlying phenomenon.</p><h4>The timing problem</h4><p>This alone would be a cute coincidence. What made me actually write the paper was the timing.</p><p>The Ekadashi fast is specifically 24 to 36 hours long. Twice per lunar month. About 24 times per year. For a lifetime.</p><p>That window is not arbitrary from a cellular standpoint. Here is what your body does during a fast:</p><p>- <strong>Hours 0 to 12</strong>: Running on glycogen from your last meal. Business as usual.<br>- <strong>Hours 12 to 16</strong>: Hepatic glycogen depletes. The metabolic switch flips. You begin burning fat and producing ketone bodies.<br>- <strong>Hours 16 to 24</strong>: Beta-hydroxybutyrate rises to 0.5 to 2.0 millimolar. It turns out this ketone is also an inhibitor of class I histone deacetylases, meaning it epigenetically upregulates autophagy genes. Your cells are now actively rebuilding their recycling machinery.<br>- <strong>Hours 24 to 36</strong>: Peak autophagic induction. Experiments in mice show neuronal autophagy kicking into high gear in exactly this window.<br>- <strong>Hours 48+</strong>: Still firing, but now you are in territory where muscle catabolism starts to become a concern without medical supervision.</p><p>The Ekadashi window, 24 to 36 hours, sits squarely inside the peak autophagy zone and stops exactly where extended fasting starts to get medically dicey without supervision. If you were an intermittent fasting researcher today trying to design a safe, repeatable, maximum-benefit protocol for the general population, you would land on something remarkably close to what the Puranas already prescribed.</p><p>Did ancient practitioners figure this out by accident? By trial and error over thousands of years? By calendrical convention that happened to land on a biologically meaningful number? I do not know. What I know is that the timing is not random.</p><h4>Jain monks and the hormesis model</h4><p>The Jains took this further than anyone else in recorded history.</p><p>Jain ascetic practice involves a graduated fasting progression that unfolds over years. You start with Ekasana, one meal per day. You advance to Biyasana, two restricted meals. Then Ayambil, one meal of plain food. Then Upavasa, complete 24-hour fasts. Then Atthai, eight consecutive days water-only. And in advanced cases, Maskhaman, a 30-day water-only fast.</p><p>Each stage is held for weeks or months before advancing. The explicit rationale in the tradition is that the body must be trained, gradually, to tolerate longer fasts. This is not dogma. It is a practical observation that if you go from zero to Maskhaman in a week you will probably die.</p><p>Modern physiology has a word for this: hormesis. It is the principle that repeated, progressively intensified exposure to a stressor builds adaptive capacity. The same principle that makes exercise work for your muscles makes cold exposure work for your cardiovascular system makes sauna sessions work for your heat shock response.</p><p>Does progressive fasting training build progressively more autophagic capacity? Nobody has tested this. A 24-hour fast in a person who has never fasted before and a 24-hour fast in a Jain monk who has been doing this for thirty years are almost certainly not the same experiment at the cellular level. But until someone actually puts both groups in a clinical trial and measures LC3-II trajectories, we are guessing.</p><p>That guess, though, is the entire premise of hormesis. And we know hormesis works for every other cellular stress pathway we have studied.</p><h4>Why I am not telling you to fast</h4><p>I want to be careful here. This is not a “ancient wisdom beats modern science” post. It is not a “try this biohack” post. And it is definitely not medical advice.</p><p>What this is, instead, is an argument that traditional knowledge systems are an underused source of structured, testable hypotheses for longevity research. They are not proof. They are not data. They are leads. They point at specific protocols, specific timings, specific combinations, specific populations you could recruit for a study.</p><p>And the cost of testing them is modest. You could run a cross-sectional study comparing autophagy markers in long-term Ekadashi practitioners against matched non-fasting controls from the same community for somewhere around $100,000 to $150,000. That is couch-cushion money by biomedical research standards. If the practitioners light up with elevated LC3-II ratios and lower p62 levels, you have learned something real. If they do not, you have learned that too.</p><p>In my paper I lay out seven specific testable hypotheses and three experimental designs. One tests whether Ekadashi practitioners have measurably enhanced baseline autophagy. One tests whether experienced Jain fasters show a faster metabolic switch during an acute fast than fasting-naive volunteers. One tests whether traditional Ayurvedic clinical assessment of Ama actually correlates with blood biomarkers of proteostasis failure. All three could run in under a year at an Indian research institution with existing infrastructure.</p><p>Nobody has done them. That is the part that frustrates me most.</p><h4>The bigger picture</h4><p>I have started calling this framework <strong>contemplative geroscience</strong>. It is the systematic study of correspondences between traditional contemplative practices and the molecular biology of aging.</p><p>Fasting and autophagy is one chapter. There are others I think are worth writing. Meditation and chronic low-grade inflammation through vagal and neuroimmune pathways. Pranayama breath practices and autonomic nervous system regulation. Rasayana herbal compounds, several of which (berberine, curcumin, urolithin A from triphala) have already been independently validated as autophagy modulators in modern pharmacology labs that had no idea they were replicating Ayurvedic prescriptions from two millennia ago.</p><p>The pattern in each case is the same. A tradition evolved a practice over centuries. The practice persisted, probably in part because it worked. Modern biology, moving from the other direction, identified a molecular mechanism that the practice appears to engage with specificity. The two frameworks were developed by people who could never have talked to each other and used incompatible vocabularies, yet they keep pointing at the same underlying phenomena.</p><p>The skeptical reading is that I am pattern-matching, that confirmation bias is doing most of the work, and that if you looked at any sufficiently large traditional knowledge system you would find convergences with modern science that mean nothing. Fair. Maybe.</p><p>The other reading is that humans have been careful observers of their own bodies for a very long time, and that the patterns they found, encoded in rituals and conceptual frameworks, sometimes point at something real.</p><p>I do not know which reading is correct. But the cost of finding out, I keep coming back to, is so low, and the potential upside so large, that I cannot understand why we are not testing this yet.</p><p>My grandmother fasted every other week for about forty years. She lived to 91 in excellent cognitive health. Is that because of Ekadashi? I have no idea. It is one data point, and one data point means nothing.</p><p>But it would be interesting to know.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c3dfe573417" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Future of AI in Banking: How AI is Revolutionizing Three Key Areas]]></title>
            <link>https://medium.com/@MindbytesAI/the-future-of-ai-in-banking-how-ai-is-revolutionizing-three-key-areas-3b0d4bde5366?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b0d4bde5366</guid>
            <category><![CDATA[futureoffinancialmodeling]]></category>
            <category><![CDATA[fraud-detection]]></category>
            <category><![CDATA[ai-in-banking]]></category>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Tue, 24 Jun 2025 00:13:25 GMT</pubDate>
            <atom:updated>2025-06-24T00:13:25.789Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial intelligence (AI) is actively improving the banking and financial services offered to customers. The progression of AI technologies in the banking industry benefits financial institutions and their customers.</p><p>The overall goal of AI in banking is to reduce fraud, increase security, improve operational efficiency, and allow customers access to innovative tools to build wealth and personalize their banking experience.</p><p>This article will explore how AI is revolutionizing key areas in the banking industry and how it will benefit financial institutions and customers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JX659WFwFVezTCHd68EE-A.png" /></figure><p><strong>Customer Service &amp; Personalization</strong></p><p>Bank customers never like waiting long for support for their questions, transactions, and inquiries. Banks have struggled with this for years because of understaffing or poorly trained employees.</p><p>AI technology enables banks to streamline the customer service experience by allowing AI-powered chatbots and voice assistants to help customers more quickly and efficiently. It reduces customer service expenses for banks while eliminating long waiting and response times for customers.</p><p><strong>AI Chatbots</strong></p><p>AI chatbots are AI-driven automated customer support agents that can communicate with customers via live chat. They are programmed to answer common customer questions and personal queries associated with their financial accounts.</p><p>AI chatbots eliminate the need for customers to wait for human customer support agents because the chatbots provide immediate assistance with accurate and detailed information to help the customers.</p><p>For example, Bank of America has an AI-driven chatbot named Erica that gives customers financial advice. HSBC has an AI-driven virtual assistant named Amy that handles customer inquiries.</p><p><strong>Voice Assistants</strong></p><p>Voice assistants are virtual assistants with integrated AI voice recognition technology that allows customers to use voice commands to conduct online banking and manage their accounts. If you have ever interacted with Google Assistant, Amazon Alexa, or Apple’s Siri, you already understand how voice assistants operate.</p><p>Your voice assistant will learn to recognize your voice and enable you to perform typical bank management actions, such as checking your bank account balances, transferring funds, and setting up bill payments.</p><p><strong>Hyper-Personalized Banking</strong></p><p>AI can now help banks offer highly personalized banking products and services to customers by analyzing enormous amounts of data related to each customer and their history with the bank.</p><p>The latest AI machine learning algorithms are programmed to review each customer’s transaction history, spending patterns, financial preferences, and behavior to make exclusive offers they might like.</p><p>Some of these personalized recommendations and offerings include financial products, personalized investment strategies, budgeting advice, spending insights, and dynamic interest rates.</p><p><strong>Risk Assessment &amp; Fraud Detection</strong></p><p>Banks have relied on traditional credit scoring models for many decades. The problem with these old-school models is that they only look at a limited amount of your financial information to determine your creditworthiness, such as your credit history and income. This information alone can no longer help banks assess risks and prevent fraud.</p><p>AI-powered credit scoring and risk assessment technology is changing everything for banks. Many financial institutions now use AI-driven tools to conduct comprehensive risk assessments when considering loan applicants and preventing fraudulent banking transactions.</p><p>The AI tools are much more accurate and thorough in conducting risk assessments and detecting fraud than human bankers.</p><p><strong>AI-Driven Loan Approvals</strong></p><p>AI-powered risk assessment models review more than your credit score and income level when considering you for a loan. They also analyze your spending patterns, employment history, cash flow, expenses, online purchasing behaviors, and other specific details about your finances.</p><p>This in-depth assessment of your current and historical finances will help banks make more informed decisions about whether to approve your loan application. Since the decision-making is AI-driven, you will know the approval results much faster than waiting for a human to assess everything manually.</p><p><strong>Real-Time Fraud Prevention</strong></p><p>More banks are using AI-powered fraud prevention systems to track all your banking transactions in real-time to look for unusual spending patterns or other suspicious activities associated with your account. If potential fraud is detected, the system will alert you via text message or email so that you can confirm or deny the legitimacy of the suspicious activity.</p><p>AI-powered fraud prevention systems also include biometric security technology to prevent unauthorized access to your bank accounts. You can train the AI to recognize your identifiable physical features, such as your fingerprints, voice patterns, or face, to verify that you are the account’s rightful owner.</p><p><strong>Conclusion</strong></p><p>The future of AI in banking will bring more efficiency and financial opportunities to banks and their customers. AI will enable banks to make better loan approval decisions and reduce the risk of fraud while giving customers better banking tools for faster, safer, and more personalized customer support, financial advice, banking experiences, and wealth management opportunities.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b0d4bde5366" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Ethical Implications of AI Regulation]]></title>
            <link>https://medium.com/@MindbytesAI/the-ethical-implications-of-ai-regulation-3d1a649f3bdc?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d1a649f3bdc</guid>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[technology-and-society]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[policy-and-regulation]]></category>
            <category><![CDATA[global-perspectives]]></category>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Wed, 10 Jul 2024 01:29:33 GMT</pubDate>
            <atom:updated>2024-07-10T01:29:33.645Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-xh3_mFw4HaoiyG1NpBIHQ.jpeg" /></figure><p>The ethical implications of AI regulation are vast, encompassing issues of transparency, accountability, bias, and societal impact. The rapid advancement of artificial intelligence (AI) technology, and its potentially dangerous effects, necessitate robust regulatory frameworks to ensure ethical practices.</p><p>This article provides a comprehensive overview of AI’s ethical implications, supported by statistical data and credible sources, and the ways in which they are and may continue to be mitigated.</p><h3><strong>Transparency and Accountability</strong></h3><p>Transparency in AI development refers to developers’ openness surrounding how their AI was coded, and the data upon which they trained it. By making this information available to the public, it’s possible to:</p><p>1. Preempt the AI’s potential biases, which may already exist in the training data.</p><p>2. Assign accountability when an AI model makes an ethical infraction or is used irresponsibly.</p><p>These reasons underpin much of the global regulation about AI, which often mandates that companies disclose the necessary information to explain their AI’s decision-making processes. For example:</p><h3><strong>The European Union AI Act</strong></h3><p>The EU is one of the first governing bodies to pass sweeping regulations concerning AI. According to Dragos Tuderache, the AI Act’s co-rapporteur, this will ensure AI models “do not present systemic risks to the Union,” instead offering “strong safeguards for our citizens and our democracies against any abuses of technology by public authorities.” (<a href="https://www.europarl.europa.eu/news/en/press-room/20231206IPR15699/artificial-intelligence-act-deal-on-comprehensive-rules-for-trustworthy-ai">Europarl</a>).</p><p>In particular, the EU AI Act, passed on March 13th, 2024, categorizes AI models into three separate risk categories to determine their ethical implications and potential harm: unacceptable risk, high risk, and other AI models. Models of unacceptable risk are automatically banned from release.</p><p>High-risk models could be anything from AI that determines a child’s access to education based on their test scores without human verification, the failure of which may result in unjustified barring and an infringement on the human right to education, to AI that operates road traffic signal systems, the failure of which could result in serious physical harm.</p><p>Companies developing AI models in the high-risk category are subject to stringent compliance protocols, requiring them to document their processes meticulously. This includes transparency about the data sets used, the algorithms’ decision-making processes, and regular audits. The EU has also developed an online AI compliance checker, making it far easier for AI development companies to ensure their offerings have ethical implications in mind from conception to release. (<a href="https://artificialintelligenceact.eu/assessment/eu-ai-act-compliance-checker/">EU AI Act</a>).</p><h3><strong>EU General Data Protection Regulation (GDPR)</strong></h3><p>In addition to passing an act dedicated to categorizing AI in terms of potential ethical harm, the EU has made it clear that its existing regulation should be applied to AI. Companies developing AI models must apply the tenets of GDPR to their offerings and ensure their AI product abides by this regulation.</p><p>For example, the GDPR tenet of purpose limitation, which mandates the collection of an individual’s data only when it is completely necessary, limited to the purposes of its processing, and agreed upon by that individual, can also be applied to AI.</p><p>According to a study by the EU Panel for the Future of Science and Technology, “The information made available to data subjects should enable them to understand the purpose of each AI-based processing and its limits, even without going into unnecessary technical details.” Essentially, it must always be clear to those using AI why their data is being collected so they can come to their own conclusions about the potential ethical implications. (<a href="https://www.europarl.europa.eu/RegData/etudes/STUD/2020/641530/EPRS_STU(2020)641530_EN.pdf">Europarl</a>).</p><p>_______________________________</p><p>The above examples show that governing bodies are taking steps to mitigate AI’s potentially ethically harmful effects, ensure its fairness and accountability, and prevent potential misuse.</p><p>Though the EU AI Act is the first large-scale regulation pertaining to the fair use of AI, many countries have begun to pass AI-centric laws to protect their citizens. The UK has put forward an AI Regulation Bill, while the US passed an executive order on the ‘Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.’ (<a href="https://lordslibrary.parliament.uk/research-briefings/lln-2024-0016/#:~:text=In%20addition,%20the%20bill%20would,AI%20at%20some%20future%20point.">Lords’ Library</a>) (<a href="https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/">White House</a>).</p><p>Studies show that countries with strict AI regulations report fewer instances of misuse, proving that requiring transparency and accountability on a legal basis can mitigate the potential ethical damage of AI models. These laws also set an important precedent and can be used as benchmarks by countries wishing to pass their own AI regulations. (<a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>) (<a href="https://www.unesco.org/en/artificial-intelligence/recommendation-ethics">UNESCO</a>).</p><h3><strong>Bias Mitigation</strong></h3><p>If not properly regulated, AI systems can perpetuate and even exacerbate existing biases. This is because AI is ‘trained’ on existing data sets compiled and managed by human beings, reflecting inherent biases present in our societies. Ensuring that AI is developed and deployed fairly requires stringent measures to identify and mitigate these biases.</p><h3><strong>Examples of Bias in AI Systems</strong></h3><p>Despite AI being a relatively new advancement, many instances of bias have already occurred, usually unfairly affecting minorities that societal institutions already target.</p><p>In 2018, Amazon did away with an AI-powered hiring system that showed preference for CVs written by men over women due to the presence of words such as ‘captured’ and ‘executed’ that were more commonly found in men’s applications. (<a href="https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G/">Reuters</a>).</p><p>Researchers at Carnegie Mellon University found that Google ads powered by AI were more likely to display high-paying job roles to male users, unintentionally exacerbating gender-biased corporate structures. (<a href="https://www.washingtonpost.com/news/the-intersect/wp/2015/07/06/googles-algorithm-shows-prestigious-job-ads-to-men-but-not-to-women-heres-why-that-should-worry-you/">The Washington Post</a>).</p><p>Predictive policing is one of AI’s most promising yet concerning applications. These systems use historical arrest data to predict where and when a crime might happen. A partnership between researchers at Carnegie Mellon University and the University of Texas found that predictive policing algorithms over-predicted the number of crimes that were going to take place, and commonly attributed them to racial minorities that police forces have unfairly targeted in the past. (<a href="https://www.technologyreview.com/2021/02/05/1017560/predictive-policing-racist-algorithmic-bias-data-crime-predpol/#:~:text=It&#39;s%20no%20secret%20that%20predictive,lessen%20bias%20has%20little%20effect.">MIT Technology Review</a>).</p><p>In 2020, a study by the National Institute of Standards and Technology found that many facial recognition technologies have higher error rates for people of color than white individuals. (<a href="https://news.yale.edu/2024/02/21/uncovered-unknown-and-uncertain-guiding-ethics-age-ai">YaleNews</a>) (<a href="https://hyscaler.com/insights/critical-ai-ethics-insights-2024-ai-expo/">HyScaler</a>).</p><p>These are just a few of the most high-profile instances of AI systems reflecting and displaying bias, but there are many more examples. With AI’s vast applications, biases could appear in everything from education to healthcare.</p><h3><strong>AI Bias Mitigation Strategies</strong></h3><p>Michael Chui, James Manyika, and Mehdi Miremad, authors of ‘What AI can and can’t do (yet) for your business,’ stated that “recognizing [biases in AI models], and taking steps to address them, requires a deep mastery of data-science techniques, as well as a more meta-understanding of existing social forces, including data collection.” This can make it incredibly difficult to manage bias, but it also reinforces the necessity of such individuals to stop these biases from becoming further embedded. (<a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/what-ai-can-and-cant-do-yet-for-your-business">McKinsey</a>).</p><p>AI bias regulation must be enforced at the highest level to lessen the potential impact on groups that have been victims of human bias in the past. The aforementioned EU AI Act has gone some way to do this, mandating that high-risk AI systems must be trained on diverse data that reflects the unique needs and behaviors of different demographics. (<a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>).</p><p>Additionally, non-governmental groups have exercised their influence to ensure the mitigation of bias in AI. In particular, the Organisation for Economic Co-operation and Development’s AI Policy Observatory has compiled a list of tools to help developers create ethically sound AI systems. One such tool is the ‘Unsupervised bias detection tool,’ which uses a binary algorithmic classifier to detect bias, allowing developers to address this issue before the AI’s release. (<a href="https://oecd.ai/en/catalogue/tools/unsupervised-bias-detection-tool">OECD.AI</a>).</p><p>Though bias is potentially one of AI’s most damaging ethical implications due to its ability to create feedback loops between AI and society, concerted efforts are already being made to mitigate it. Going forward, it’s vital that these efforts keep up with AI’s rapid advancement.</p><h3><strong>Societal Impact</strong></h3><p>Since AI has begun to touch almost every area of society, AI regulation must also address its broader societal impacts, including personal and data privacy, security, and economic effects.</p><h3><strong>Privacy and Security</strong></h3><p>The protection of user data is a primary concern in AI ethics. Without proper safeguards for user data, AI systems may be subject to breaches, resulting in the loss of personal and financial information. In fact, according to a report by Hidden Layer, 77% of businesses reported breaches to their AI systems in 2024. (<a href="https://hiddenlayer.com/threatreport2024/">Hidden Layer</a>).</p><p>With cybersecurity lacking when it comes to AI, it’s up to regulatory frameworks to pick up the slack. As mentioned, existing regulations like GDPR have been amended to ensure their application to AI. This law requires companies to obtain explicit user consent before collecting their data, providing transparency about data usage. Non-compliance can lead to significant fines, which acts as a deterrent against unethical data practices by AI developers. (<a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>) (<a href="https://www.unesco.org/en/artificial-intelligence/recommendation-ethics">UNESCO</a>).</p><h3><strong>Economic Impacts</strong></h3><p>Interestingly, the economy is one area where strict regulations may hamper the benefits of widespread AI use in different industries. AI regulations can influence economic activities by setting standards that affect innovation and competition.</p><p>With 60% of jobs in advanced economies expected to feel the positive effects of AI, such as automation and enhanced productivity, regulations that prevent the release of these systems due to bias and data protection laws create a tricky balancing act between potential ethical damage and economic advantages. (<a href="https://www.imf.org/en/Blogs/Articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity#:~:text=Roughly%20half%20the%20exposed%20jobs,of%20these%20jobs%20may%20disappear.">International Monetary Fund</a>).</p><p>But even if the beneficial effects of AI systems aren’t felt immediately due to increased care during the development stage and a longer wait for rollouts, this could increase trust and stability in the AI market, ultimately leading to long-term economic well-being. In fact, according to a study by McKinsey, regulated AI markets tend to have more sustainable growth as they balance innovation with ethical considerations. (<a href="https://news.yale.edu/2024/02/21/uncovered-unknown-and-uncertain-guiding-ethics-age-ai">YaleNews</a>) (<a href="https://hyscaler.com/insights/critical-ai-ethics-insights-2024-ai-expo/">HyScaler</a>).</p><h3><strong>Global Perspectives</strong></h3><p>While far-reaching regulations like the EU AI Act are incredibly important, they need to consider the nuances that affect specific countries and even locales within those countries. This is why different regions must take different approaches to AI regulation, tailored to the unique ethical implications of AI use in that area.</p><p><strong>United Kingdom: </strong>The UK is one country aiming to balance innovation with regulation, as highlighted above. Their ‘pro-innovation approach to AI regulation’ would see regulating authorities handling governance and regulation of specific AI systems. At the same time, the UK parliament would merely set the standards for how this governance should be carried out. This will ensure the UK can continue to invest in technological advancement while ensuring safety in every industry where AI is used. (<a href="https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper#executive-summary">GOV UK</a>).</p><p><strong>United States</strong>: The US adopts a more hands-off approach, with fewer federal regulations but significant input from state-level laws and industry standards. California, Florida, and New York are at the forefront of this approach, each passing AI regulations that focus on transparency and accountability. However, aside from the executive order previously mentioned, no federal regulation has been passed yet. This approach can lead to innovation and state-centric AI regulation but may result in inconsistent ethical governance. (<a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>) (<a href="https://www.unesco.org/en/artificial-intelligence/recommendation-ethics">UNESCO</a>).</p><p><strong>China</strong>: China originally published a New Generation Artificial Intelligence Development Plan in 2017 that set out goals for AI creation and release in the country. However, there has yet to be a sweeping AI regulation in the country, with laws pertaining to specific elements of AI, like generative AI, instead. This allows for quick regulatory responses and a comprehensive governmental understanding of the particular applications of AI and its effect on the population but may lack a cohesive ethical strategy. (<a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>).</p><p><strong>Japan</strong>: Like many other countries, Japan intends to allow AI development to flourish due to its prospective benefits and not to overestimate its potential risks. This has led to a policy of agile governance, with the government publishing the Social Principles of Human-Centric AI that outline how and why AI should be released and used in the country. However, though they guide developers, the Japanese government relies on smaller regulatory bodies in specific industries to police potentially harmful AI systems. (<a href="https://www.csis.org/analysis/japans-approach-ai-regulation-and-its-impact-2023-g7-presidency">Center for Strategic and International Studies</a>).</p><p><strong>India</strong>: The Ministry for Commerce and Industry has established a task force that can balance India’s desire to become a global leader in AI development and use it with its potential ethical effects. Though there are no laws that govern AI in India yet, this task force intends to make recommendations to the government on how AI can be regulated at both a federal and industry level. (<a href="https://www.aitf.org.in/">AITF.org</a>).</p><p><strong>Australia</strong>: Similar to how GDPR must be applied to AI in the EU, Australia has recommended that its existing regulations governing technology, ethical considerations, data privacy, and security be applied to the development of AI. While this means AI will be brought under the same far-reaching ethical laws that affect other advancements in the country, these laws won’t be tailored to the specific ways in which AI can infringe on the rights of Australian citizens, and may result in a lack of AI innovation altogether. (<a href="https://www.mindfoundry.ai/blog/ai-regulations-around-the-world">Mind Foundry</a>).</p><p>Other countries such as Brazil and Canada have also begun to take steps to regulating AI, and it’s likely that the next few years will see more robust regulation passed. This will ensure the unique ethical implications of AI in different geographical, social, and economic contexts can be mitigated by the power of the law, creating a safer and fairer technological landscape.</p><h3><strong>Conclusion</strong></h3><p>The ethical implications of AI regulation are multifaceted, impacting transparency, accountability, bias mitigation, and societal effects. Robust regulatory frameworks are essential for ensuring that AI technologies are developed and used in ways that are fair, safe, and beneficial to society. As AI evolves, ongoing dialogue and research will be critical in refining these regulations to address emerging ethical challenges.</p><h3><strong>References</strong></h3><p>EU AI Act Press Conference: <a href="https://www.europarl.europa.eu/news/en/press-room/20231206IPR15699/artificial-intelligence-act-deal-on-comprehensive-rules-for-trustworthy-ai">Europarl</a></p><p>EU AI Compliance Checker: <a href="https://artificialintelligenceact.eu/assessment/eu-ai-act-compliance-checker/">EU AI Act</a></p><p>GDPR and AI: <a href="https://www.europarl.europa.eu/RegData/etudes/STUD/2020/641530/EPRS_STU(2020)641530_EN.pdf">Europarl</a></p><p>UK AI Reform Bill: <a href="https://lordslibrary.parliament.uk/research-briefings/lln-2024-0016/#:~:text=In%20addition,%20the%20bill%20would,AI%20at%20some%20future%20point.">Lords’ Library</a></p><p>US AI Executive Order: <a href="https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/">White House</a></p><p>AI Misuse Studies: <a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>, <a href="https://www.unesco.org/en/artificial-intelligence/recommendation-ethics">UNESCO</a></p><p>Amazon AI Bias: <a href="https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G/">Reuters</a></p><p>Google Ads Bias: <a href="https://www.washingtonpost.com/news/the-intersect/wp/2015/07/06/googles-algorithm-shows-prestigious-job-ads-to-men-but-not-to-women-heres-why-that-should-worry-you/">The Washington Post</a></p><p>Predictive Policing Bias: <a href="https://www.technologyreview.com/2021/02/05/1017560/predictive-policing-racist-algorithmic-bias-data-crime-predpol/#:~:text=It&#39;s%20no%20secret%20that%20predictive,lessen%20bias%20has%20little%20effect.">MIT Technology Review</a></p><p>Facial Recognition Bias: <a href="https://news.yale.edu/2024/02/21/uncovered-unknown-and-uncertain-guiding-ethics-age-ai">YaleNews</a>, <a href="https://hyscaler.com/insights/critical-ai-ethics-insights-2024-ai-expo/">HyScaler</a></p><p>AI Bias Mitigation: <a href="https://www.mckinsey.com/capabilities/quantumblack/our-insights/what-ai-can-and-cant-do-yet-for-your-business">McKinsey</a></p><p>Bias Detection Tool: <a href="https://oecd.ai/en/catalogue/tools/unsupervised-bias-detection-tool">OECD</a></p><p>AI Data Breaches: <a href="https://hiddenlayer.com/threatreport2024/">Hidden Layer</a></p><p>AI and the Economy: <a href="https://www.imf.org/en/Blogs/Articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity#:~:text=Roughly%20half%20the%20exposed%20jobs,of%20these%20jobs%20may%20disappear.">International Monetary Fund</a></p><p>UK AI Legislation: <a href="https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach/white-paper#executive-summary">GOV UK</a></p><p>US AI Legislation: <a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>, <a href="https://www.unesco.org/en/artificial-intelligence/recommendation-ethics">UNESCO</a></p><p>Chinese AI Legislation: <a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a></p><p>Japanese AI Legislation: <a href="https://www.csis.org/analysis/japans-approach-ai-regulation-and-its-impact-2023-g7-presidency">Center for Strategic and International Studies</a></p><p>Indian AI Legislation: <a href="https://www.aitf.org.in/">AITF.org</a></p><p>Australian AI Legislation: <a href="https://www.mindfoundry.ai/blog/ai-regulations-around-the-world">Mind Foundry</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d1a649f3bdc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ethical AI For Children]]></title>
            <link>https://medium.com/@MindbytesAI/ethical-ai-for-children-716f6ff317d6?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/716f6ff317d6</guid>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Tue, 25 Jun 2024 05:40:21 GMT</pubDate>
            <atom:updated>2024-06-25T05:40:21.680Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NziWC_C_ZV6ejcLbbujbuQ.jpeg" /></figure><p>The rapid integration of AI technologies into various aspects of life, including education, entertainment, and social interaction, raises significant ethical considerations, particularly concerning children. While this demographic stands to benefit from growing up alongside AI, due to their developmental stage, they’re also uniquely vulnerable to the technology’s potentially adverse impact.</p><p>Research in this field aims to ensure that AI applications are designed and deployed in ways that are safe, beneficial, and supportive of children’s growth and well-being. This article summarizes the ethical implications of AI’s use by and for children, and how regulations and governance can ensure children’s data privacy, safety, and well-being.</p><h3><strong>Privacy and Data Protection</strong></h3><p>In the digital age, data privacy is a significant concern. This is especially true when it comes to AI, which collects vast amounts of data and uses it to ensure its algorithms and output improve over time. Children’s data is susceptible, so ensuring robust data protection measures is crucial, like:</p><p>1. <strong>Sensitive Data Handling</strong>: AI systems must comply with stringent privacy laws such as COPPA (Children’s Online Privacy Protection Act) in the US, which mandates explicit consent from parents or guardians before collecting the data of children under the age of thirteen.</p><p>2. <strong>Data Minimization</strong>: A long-held tenet of data privacy, data minimization is the practice of collecting and using only the data necessary for a technology’s basic functioning. In the case of AI and children’s data, minimization reduces the risk of misuse and exposure of sensitive information. (<a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>) (<a href="https://www.sciencedaily.com/releases/2024/03/240320122433.htm">ScienceDaily</a>).</p><p>The risks associated with the privacy and data protection of minors require preventative measures built into AI during the developmental stage and continually updated after systems’ releases. With a concerted effort toward information security, AI models can remain safe for children to use throughout their lives.</p><p>Source: Federal Trade Commission, “Children’s Online Privacy Protection Rule (‘COPPA’),” 2021.</p><h3><strong>Bias and Fairness</strong></h3><p>Biased AI was a concern even when it was a theoretical concept, and since its widespread rollout and integration, instances of AI systems showing bias are not uncommon. Biased AI output usually occurs due to inherent ‘human’ biases in the datasets these systems are trained on. This could be particularly harmful to children who may internalize the bias shown by AI, creating a feedback loop that unfairly targets marginalized groups.</p><p>For example, according to research by Stanford University, children whose first language is not English are more likely to have their work incorrectly flagged as AI-generated by AI-powered detection systems. To avoid reinforcing biases such as these, developers must train their AI systems on diverse datasets that represent all children, regardless of their background. This helps in creating fair and unbiased AI interactions, reinforcing equitable learning. (<a href="https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7?_returnURL=https://linkinghub.elsevier.com/retrieve/pii/S2666389923001307?showall=true">Cell</a>) (<a href="https://news.yale.edu/2024/02/21/uncovered-unknown-and-uncertain-guiding-ethics-age-ai">YaleNews</a>) (<a href="https://www.mckinsey.com/featured-insights/in-the-balance/from-principles-to-practice-putting-ai-ethics-into-action">McKinsey &amp; Company</a>).</p><p>AI tools should also be designed to cater to children with varying abilities and socio-economic backgrounds, ensuring equal access and benefit from AI technologies. This means incorporating features such as text-to-speech, speech-to-text, and visual aids, as well as adaptive learning technology that adjusts to the pace and level of each child. One such example is Adobe Captivate, which also assists with distance learning and e-learning, providing supplemental education to children who, for financial or social reasons, may not have access to it. (<a href="https://belitsoft.com/custom-elearning-development/ai-in-education/adaptive-learning-ai">Belitsoft</a>).</p><p>Source: Gebru, T., et al. (2018). “Datasheets for Datasets,” arXiv preprint.</p><p>With AI use in daily life increasing and showing no signs of slowing down, it’s no stretch to say that ensuring unbiased and fair systems for children’s use could create a more equitable world in the future.</p><h3><strong>Transparency and Explainability</strong></h3><p>The way AI works can be difficult to explain even to adults, so teaching children about the technology requires simple, age-appropriate examples that make complex concepts accessible without understating potential risks and proper methods of use. In particular, this requires:</p><p><strong>Clear Communication</strong>: AI systems interacting with children should be transparent about how and why they use AI. Children should understand when they are engaging with an AI and what data it collects, even if they may need their parents to explain further. (<a href="https://www.mckinsey.com/featured-insights/in-the-balance/from-principles-to-practice-putting-ai-ethics-into-action">McKinsey &amp; Company</a>)..</p><p><strong>Explainable AI</strong>: The decision-making processes of AI systems should be explainable in simple terms, allowing parents and educators to understand and trust the AI’s actions and recommendations. This is particularly important for enabling children to think critically about how AI may ‘choose’ to assist them.</p><p>Source: European Commission, “Ethics Guidelines for Trustworthy AI,” 2019.</p><p>Psychological and Social Impact</p><p><strong>Screen Time Management</strong>: Many parents and guardians already worry about managing children’s screen time on devices. AI applications should help manage screen time effectively to prevent overuse and its negative effects on children’s physical and mental health.</p><p><strong>Positive Reinforcement</strong>: AI should be used to reinforce positive behavior and learning, providing personalized feedback and encouragement tailored to each child’s needs and progress. This approach means AI can support individual development, and create a positive and motivating learning environment for all children.</p><p>Source: American Academy of Pediatrics, “Media and Young Minds,” 2016.</p><h4>Safety and Security</h4><p><strong>Robust Security Measures</strong>: AI systems are subject to frequent hacking attempts and leaks, which often result in data misuse, necessitating strong security protocols. While considerations regarding bias and data privacy must be made in the developmental stage, such attention must not cease after AI systems have been put into use by and for children. Generally, this includes regular audits and updates to address any vulnerabilities.</p><p><strong>Parental Controls</strong>: AI applications for children should offer comprehensive parental controls, allowing parents to monitor and guide their children’s use of AI technologies. AI assistants tailored to children, like Angel AI, already provide such functionality, even alerting parents when their child asks questions that may require further clarification from an adult. (<a href="https://newatlas.com/children/angel-ai-parental-controls/#:~:text=AI%20offers%20a%20chance%20to,inappropriate%20content%20or%20privacy%20risks.">New Atlas</a>) (<a href="https://www.sciencedaily.com/releases/2024/03/240320122433.htm">ScienceDaily</a>) (<a href="https://www.apa.org/monitor/2024/01/trends-generative-ai-evolution-psychology#:~:text=URL%3A%20https%3A%2F%2Fwww.apa.org%2Fmonitor%2F2024%2F01%2Ftrends,100">APA Monitor</a>)..</p><p>Source: National Institute of Standards and Technology (NIST), “Security and Privacy Controls for Information Systems and Organizations,” 2020.</p><p>Building responsible and effective AI tools is incredibly important if they are going to be used by children. However, should the AI fall short, transparency in the way it works and explainability that enables children to understand these processes mean they can critically assess potential limitations.</p><h3><strong>Case Studies and Examples</strong></h3><p>Below are some examples of AI tools that have been developed for use by children, the ethical considerations their developers made, and how those considerations affected their usability and success.</p><p><strong>AI in Education</strong></p><p>As mentioned, children who speak English as a second language can sometimes struggle to use specific AI systems or be unfairly targeted by their existing biases. In the case of language learning, AI-powered apps may not recognize the spoken practice as accurate, unfairly penalizing these children and hindering their learning process.</p><p>Azure AI sought to address this issue with improved pronunciation assessment capabilities. Language-learning system Berlitz then used this technology to improve user experience. They estimate that this focus on addressing the shortcomings of educational AI has enabled them to help 500,000 individuals learn English. (<a href="https://customers.microsoft.com/en-us/story/1650519804730300378-berlitz-language-azure-ai-speech-usa">Microsoft</a>).</p><p>However, researchers must ensure these systems do not widen the educational gap between children from different backgrounds. In the case of Berlitz, access to their AI-powered pronunciation practice technology requires a place on one of their courses, starting at $330.</p><p>With the cost of AI development and technological improvements such as those in this case study increasing, children from less economically developed areas will not be able to access this tool. Accessibility and inclusivity will always be paramount to AI’s ethical deployment, necessitating more widely available tools in the future. (<a href="https://www.technologyreview.com/2024/01/05/1086203/whats-next-ai-regulation-2024/">MIT Technology Review</a>).</p><p>Source: OECD, “AI in Education: Challenges and Opportunities for Sustainable Development,” 2019.</p><p><strong>AI in Entertainment</strong></p><p>Online games such as Roblox are hugely popular with children, with over 60% of its users under the age of twenty-five. In addition to providing a place for children to communicate with friends, Roblox has a built-in mini-game builder that enables children to code their own games and share them with the world. (<a href="https://www.statista.com/statistics/1190869/roblox-games-users-global-distribution-age/">Statista</a>).</p><p>In 2023, Roblox’s Chief Technology Officer, Daniel Sturman, announced the platform’s plans to incorporate generative AI into this mini-game builder. By making it easier for children with limited coding knowledge and those who cannot yet design detailed models, he hopes that this will “accelerate creator productivity” while “significantly dropping the technical skill needed to bring ideas to life.”</p><p>Should generative AI be added to Roblox, this will help children realize their creative ideas, providing them a foundational knowledge of game development on which to build through future education. However, while the tools will likely be available to all players, addressing accessibility concerns, adding generative AI to Roblox’s game-builder raises other ethical issues. For example, Roblox takes a 30% commission on every purchase of mini-games made by its users, making claims of increased productivity more sinister. (<a href="https://blog.roblox.com/2023/02/generative-ai-roblox-vision-future-creation/">Roblox Blog</a>) (<a href="https://www.youtube.com/watch?v=_gXlauRB1EQ">People Make Games</a>).</p><p><strong>AI in Social Media</strong></p><p>In the modern world, children grow up with the internet as a powerful tool to help them learn and make and keep in contact with friends across the world. Undoubtedly, social media is the best place to do this. But social media is no longer simply a tool for communication. It’s almost seen as a requirement to have a presence on these sites, despite the myriad dangers that may come with this.</p><p>Unmoderated social media feeds may mean children see content that is disturbing, unkind, or that influences them in negative ways. AI-powered content moderation provides significant protections in this regard, with sites like Facebook and Instagram entrusting up to 98% of their moderation decisions to image recognition and natural language processing tools. (<a href="https://www.lemonde.fr/en/pixels/article/2023/11/14/content-moderation-key-facts-to-learn-from-facebook-instagram-x-and-tiktok-transparency-reports_6252988_13.html#:~:text=Meta%20also%20highlights%20the%20efficiency,the%2045%25%20reported%20by%20TikTok.">Le Monde</a>).</p><p>Algorithms designed for this purpose must be fine-tuned to effectively filter inappropriate content without infringing on free expression. To this end, Facebook employs approximately 30,000 human content moderators to ensure the AI removes content accurately. This results in a safe social media site that protects children from harmful content while allowing them to enjoy the benefits of web-based connection. (<a href="https://www.newamerica.org/oti/reports/everything-moderation-analysis-how-internet-platforms-are-using-artificial-intelligence-moderate-user-generated-content/case-study-facebook/">New America</a>).</p><p>Source: International Journal of Child-Computer Interaction, “The Use of AI in Moderating Online Content for Children,” 2020.</p><h3><strong>Future Directions</strong></h3><p>Children born today will likely grow up alongside AI, and improvements to the technology are likely to continue throughout their lives. A concerted effort now means that future generations can benefit from AI systems that are effective yet ethical, and work to increase accessibility, enhance their learning, and improve their daily lives…</p><p><strong>Interdisciplinary Research and Collaboration</strong></p><p>Developing ethical AI for children requires collaboration between developers, psychologists, educators, and policymakers. This interdisciplinary approach ensures that AI tools will be beneficial, safe, and aligned with children’s developmental needs, no matter what area of life they intend to improve or assist with. (<a href="https://www.pymnts.com/artificial-intelligence-2/2024/this-years-model-openai-salesforce-and-others-boost-efforts-for-ethical-ai/">PYMNTS.com</a>).</p><p>Source: Nature Machine Intelligence, “The Role of Interdisciplinary Research in AI Ethics,” 2021.</p><p><strong>Regulatory Frameworks and Legislation</strong></p><p>The ongoing development of regulatory frameworks is crucial to addressing emerging ethical issues in AI. Governments and international bodies must work together to establish comprehensive guidelines that protect children globally.</p><p>Though regulations explicitly concerning children’s use of AI are scant, provisions regarding children’s data and online safety are standard worldwide. An example is GDPR (General Data Privacy Regulation) in the EU, which gives children the same rights to their data as adults and mandates that the collection and processing of this data require parental consent if the child is under the age of sixteen.</p><p>The EU has stated that tenets outlined in GDPR must be applied to AI tools, setting a clear precedent that countries around the world can follow.</p><p>Source: World Economic Forum, “AI Governance: A Holistic Approach,” 2020.</p><p><strong>Ethical AI Frameworks, Guidelines and Standards</strong></p><p>Establishing clear guidelines and standards for ethical AI development, tailored specifically for children’s interactions, can help developers and companies create responsible AI technologies. (<a href="https://www.mckinsey.com/featured-insights/in-the-balance/from-principles-to-practice-putting-ai-ethics-into-action">McKinsey &amp; Company</a>) (<a href="https://www.technologyreview.com/2024/01/04/1086046/whats-next-for-ai-in-2024/">MIT Technology Review</a>).</p><p>A leader in this area is Imisi 3D, a Nigerian start-up that created an AI-powered VR game for non-autistic children to practice interacting with individuals with ASD. As well as the overall aim to benefit children, Imisi 3D operates by three ethical requirements outlined by UNICEF; “supporting children’s development and well-being, ensuring inclusion, and prioritizing fairness and nondiscrimination.” (<a href="https://www.unicef.org/innocenti/media/1726/file/UNICEF-Global-Insight-AI-case-study-Imisi-3D-2021.pdf">UNICEF</a>).</p><p>With ethical considerations built into developers’ policy, the end product will enhance children’s learning and lives and ensure fair and accessible operation for as long as possible.</p><p>Source: IEEE, “Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems,” 2019.</p><h3><strong>Conclusion</strong></h3><p>AI is becoming a part of daily life. With many entertaining and educational applications for the technology, children are coming into contact with it now more than ever. Ensuring ethical AI for children, therefore, should be a primary concern for developers. This involves addressing privacy issues and bias, ensuring transparency, and considering psychological impacts through robust design and regulation.</p><p>However, while children and parents alike must be able to trust the AI that they use, it’s also important to teach them how to assess these tools critically. To this end, explainability should also be a key factor in the release of AI for children, perhaps even more so than for standard AI tools.</p><p>By focusing on key areas such as those listed in this article, we can develop AI technologies that support children’s growth and well-being, providing them with safe, ethically sound, and enriching experiences in their digital interactions.</p><p>For more detailed information and latest updates, consult sources such as <a href="https://www.technologyreview.com/">MIT Technology Review</a>, <a href="https://www.sciencedaily.com/">ScienceDaily</a>, and <a href="https://news.yale.edu/">YaleNews</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=716f6ff317d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Code: How TRiSM Redefines AI’s Promise for Society]]></title>
            <link>https://medium.com/@MindbytesAI/beyond-code-how-trism-redefines-ais-promise-for-society-2616b1f37ed6?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/2616b1f37ed6</guid>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Fri, 23 Feb 2024 22:57:26 GMT</pubDate>
            <atom:updated>2024-02-23T22:57:26.253Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VgNiWtCLnYPwgD8-5BtU4g.png" /></figure><p>AI might feel like it has come from nowhere, but the biggest developments in the last two years have been years in the making. As AI becomes more commonplace in day-to-day life, even in professions where it can feel superfluous, attention has been drawn to its positives and negatives. For all of the benefits AI can provide, such as idea generation, automation, and efficiency, risks exist around trust and security.</p><p><strong>As such, AI Trust, Risk, And Security Management (TRiSM) strategies have been developed.</strong></p><p>How can TRiSM strategies be implemented to ensure AI systems are designed with these policies?</p><p>How can we ensure that AI is secure and reliable but also capable of meeting human-set ethical standards?</p><p>For any business intending to implement AI into their business, ensuring any AI being used follows agreed TRiSM strategies is vital. When implemented into AI development from the start, TRiSM allows for AI that is effective but also safe to use.</p><h3>TRiSM: At A Glance</h3><p>In the development of any AI solutions, it is vital that TRiSM is applied vigorously. These practices ensure that an AI system is designed in a way that makes it trustworthy and free from the risks often associated with using AI. Not only can this ensure that AI is used ethically, but also that the risk and potential for security vulnerabilities are reduced significantly. Today, the core foundations of TRiSM include:</p><ul><li><strong>Trustworthiness</strong>. Any AI systems developed have to be trustworthy. This means that their usage and output must be transparent and that the results produced are consistently reliable.</li><li><strong>Risk Management</strong>. Another factor of TRiSM is ensuring that all AI projects are thoroughly assessed to identify potential risks arising from its usage.</li><li><strong>Security Measures</strong>. AI systems must be designed from the ground with a strict focus on protecting the system from potential security threats, vulnerabilities, and backdoor entrances.</li></ul><p>These are considered the ‘pillars’ of TRiSM, and ensuring all AI is developed to the above standards should become a priority.</p><h3>Why Do We Need TRiSM?</h3><p>Put simply, people are not yet fully ready to trust AI. While much of this is a build-up of assumptions based on works of fiction, there are justifiable reasons to be concerned about AI. Given that over one-third of AI systems show at least some form of bias that impacts their decision-making, public concerns around the ethical strength of AI further highlight the importance of making TRiSM a foundation of development.</p><p>Indeed, a 2022 report from Deloitte found that 3 out of 4 executives in major global markets intend to gain the trust of their consumers by creating trustworthy AI tools. However, with around 58% of adults in the United States concerned about AI and robotics lacking human ethics, there is work to be done to convince that same consumer public.</p><p>Given that studies by the MIT Sloan Management Review found that 1 in 4 AI projects fall apart due to a failure to implement risk management frameworks early on, these public fears can feel justified. Indeed, the Bank for International Settlements found that some $217m had been lost in 2020 thanks to AI and machine-learning risks and mistakes.</p><p>Also, individuals do not need to look too far to find negative stories about AI being used negatively. Indeed, in 2023, there has been a 50% increase in AI-powered cyberattacks. By 2025, it is expected that around $10 billion will be spent on AI security systems, jumping from $4 billion in 2021. This shows a natural concern within the industry that AI is not as safe as it should be.</p><h3>How Can TRiSM Be Implemented In AI?</h3><p>The challenge in ensuring TRiSM is implemented properly is that it is included from the start. This means incorporating TRiSM into already-built AI systems can be challenging. When every stage of the conceptualization and development of the AI platform takes TRiSM into account, though, it is much easier to use effectively.</p><h3>Implementing Trustworthiness</h3><p>The first step to implementing TRiSM is ensuring that AI developments are designed with ethics in mind. This means that the developers must first understand the importance of ethics within the context of AI. When ethics are a primary consideration during the design and development phase, ensuring that an AI tool can be used for the right purposes is much easier.</p><p>This also helps to guarantee that the AI provides unbiased outcomes, reducing the risk of inherent bias and unethical thinking being ‘baked in’ to the AI system’s output. It is also vital that any AI system developed can be transparent in how it comes to its outcomes and conclusions. Any AI should be able to show it working and let the user know the steps taken to reach the conclusion that is arrived at. This helps the user to understand that the AI has followed a strict, ethical process instead of providing pre-determined answers to fit a specific agenda.</p><h3>Implementing Risk Management</h3><p>The same goes for risk management within TRiSM. A thorough risk assessment should be conducted to determine what challenges might exist for an AI project. What challenges might make the AI unsafe or potentially compromise its fairness?</p><p>Once the risks are identified, developers must implement strategies that would remove or mitigate the risk. This would allow for the AI system to be used within an acceptable threshold of what would be determined as a risk. While all risks are unlikely to be removed, designing AI with risk management in mind will improve and reduce the level of risk in usage.</p><p>With some 60% of AI risks being focused on data privacy, it is natural that users and implementors are concerned. TRiSM helps overcome these risks from the first design and development phase.</p><h3>Implementing Security Measures</h3><p>Security measures should be designed into the AI project from the offset. AI systems that are found to have vulnerabilities and holes in their security can be hard to correct if developed with the issues present. By designing the AI project from the start with an intention to avoid this, the risk of such security breaches being found later down the line is decreased exponentially.</p><p>However, regular monitoring should be employed during every development and deployment phase. This helps to spot potential security flaws that might have slipped the net, finding them before they can be exploited. One reason security measures are needed is due to the prevalence of data poisoning, the act of contaminating data to make an AI system produce unsatisfactory or biased results.</p><p>Around 1 in 5 AI security problems come from data poisoning. Ensuring that data poisoning can be mitigated and overcome is a vital part of TRiSM, reducing the risk of an AI being trained on incorrect data and thus producing pre-determined or damaging results.</p><h3>Where Is TRiSM Most Effective?</h3><p>Any industry that can utilize AI will benefit from TRiSM becoming commonplace. However, two sectors where we can see the greatest impact are:</p><h3>Healthcare</h3><p>The healthcare sector needs AI tools that can provide accurate diagnoses without compromising the individual’s or their data’s health. TRiSM in healthcare AI improves diagnostic assistance, provides transparency in how that diagnosis was reached, and gives patients extra peace of mind, knowing that their data is kept secure, thus improving treatment and diagnosis without breaching ethical standards.</p><h3>Finance</h3><p>The finance sector will also benefit from TRiSM, as AI is already a hugely useful tool within the finance industry. This helps to reduce things like counting errors and improves accuracy in fraud detection. On top of that, TRiSM will ensure that finance-focused AI tools are transparent in their decision-making while ensuring that all suggestions and actions meet industry regulations.</p><h3>Implementing TRiSM: Is It Possible?</h3><p>While the idea of TRiSM is fine, its implementation is easier said than done. It provides us with an easy way to cap AI’s dangers whilst benefiting from efficiency, accuracy, and performance improvements.</p><p>However, implementing TRiSM into AI is difficult for both developers and organizations. For one, there is the age-old resistance to change. Things were fine the way they were, so why change?</p><p>Companies often just see the financial benefits of implementing AI and want it in place, regardless of potential consequences. Therefore, stakeholders must be a key part of the TRiSM implementation process. By highlighting the benefits of reducing AI risks, stakeholders can come around to the benefits of implementation without taking so many risks in the future.</p><p>This can be done through continuous negotiation and explanation and comprehensive training. The more key stakeholders understand the benefits of TRiSM, the less likely they are to see it as intrusive or limiting. Thus, its implementation and uptake are more likely to become an industry standard.</p><p>However, as AI continues to develop and improve, TRiSM has to stay ahead of the curve. New complications that could impact TRiSM standards emerge with every new AI development. As such, there has to be a company-wide embrace of continuous learning and improvement. This helps ensure that AI can continue to be beneficial whilst staying ahead of potential security threats and emerging exploits.</p><h3>Using AI To Implement TRiSM</h3><p>Coincidentally, one of the most effective ways to implement TRiSM into AI will be by using AI. AI can provide the most comprehensive data crunching for risk assessments and security analysis. Therefore, using an AI trained on TRiSM for risk assessment can make it easier to assess future AI systems for their suitability.</p><p>TRiSM is going to be a vital part of all future AI developments. As AI continues to evolve into the mainstream, global standards will be developed to mitigate risk and ensure that trust, risk, and security management become industry-wide.</p><p>In the future, we expect AI regulations to be adopted globally, ensuring that the industry follows the high standards set in other global industries.</p><h3>TRiSM: Unlocking The Full Potential Of AI</h3><p>Many of the biggest worries about AI come from the perception of risk. We have all seen the various science fiction tropes of where AI can go if left unchecked. However, by making TRiSM a standard for all AI development, developers and companies using AI can reap the rewards of AI without having so many concerns about security risks and vulnerabilities.</p><p>As each day passes, AI becomes more important in our personal and professional lives. To ensure that AI is accepted within wider society, TRiSM is vital for creating a future where AI is not feared. By ensuring that the risks are understood from the start, instead of simply looking at the potential benefits, we can make AI as safety-conscious as other industries that have created the world better.</p><p>AI comes with many risks and concerns — arguably as many problems as it provides beneficial opportunities. By ensuring that TRiSM becomes a standard across the industry, we can create a future where AI is utilized without fear of breaking ethical boundaries or producing opaque results that cannot be understood. This, in turn, makes AI a much more useful tool for everyone.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2616b1f37ed6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The “Tailored” Truth: AI News Feeds and Your Brain]]></title>
            <link>https://medium.com/@MindbytesAI/the-tailored-truth-ai-news-feeds-and-your-brain-883d1355d665?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/883d1355d665</guid>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Fri, 05 Jan 2024 06:32:25 GMT</pubDate>
            <atom:updated>2024-01-05T06:32:25.460Z</atom:updated>
            <content:encoded><![CDATA[<p>Have you ever wondered why your news feed seems to read your mind? As you mindlessly scroll through your phone to unwind after a long day, you find stories that align so perfectly with your interests that it feels like your news app is eavesdropping on your thoughts. That’s AI-curated news feeds for you — a blend of convenience and clairvoyance, but at what cost?</p><p>Today, we’ll take a look at the intricate world of AI-curated news feeds and what this development means for us all. How do algorithms manage to provide you with news that keep your eyes glued to the screen? And how is this AI-driven personalization shaping our views of the world?</p><h3>The Rise of AI in News Media</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/1*hs8igz7PbtX8_rbBUzZg2A.png" /></figure><h3>What are AI-curated news feeds?</h3><p>What are AI-curated news feeds, you ask? Think of a chef who knows exactly what you like. He serves dishes that feel exclusively made for you. That’s exactly what AI does when it comes to the news you read or view.</p><p>AI in news media uses algorithms to sift through your digital footprints — it analyzes your past interactions, preferences, and even the duration you spend on various news topics. This process then tailors your news feed to mirror your unique interests.</p><p>Why do they do this? If the algorithms can pique your interest, then you will spend more time-consuming news on their platform. A win for them, but is it a win for you too?</p><h3>How is news consumption changing?</h3><p>Remember the days of browsing through a newspaper or waiting for the evening news on TV? That was the era, where news was a universal experience. However, this experience has shifted dramatically. Now, AI-driven news feeds have transformed news consumption into a very personalized experience.</p><h3>User Engagement with AI-curated News Feeds</h3><p>Have you noticed a change in how we digest news? Well, it’s not just your imagination. Recent findings from the Pew Research Center paint a clear picture: there’s been a notable 20–25% uptick in daily engagement with news since AI stepped into the picture. This isn’t merely about reading more. We interact on a deeper level with the content we consume.</p><h3>The shift in source variety</h3><p>But there’s another side to this story. While AI-curated feeds keep us more plugged in, they also guide us toward a limited selection of news sources. Data from the Nieman Journalism Lab indicates a 10–15% reduction in the variety of news sources people explore post AI-adoption. This shift toward a more uniform set of news sources is a subtle, yet impactful, consequence of how AI reshapes our access to information.</p><h3>The Psychological Twists</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/402/1*-InMeuuB8rPMbIS7tbG2YQ.png" /></figure><p>If you thought that this increase in getting our news from a limited number of sources had no effect on our psychology and beliefs, I have a bad surprise for you. There are two main offenders when it comes to the psychological effects of news personalization that are on the rise:</p><h3>The digital echo chamber phenomenon</h3><p>Have you ever felt like your news feed knows you a little too well? This is no wonder since AI-curated news feeds subtly create digital echo chambers. These chambers reflect back our own beliefs, feeding us news that aligns perfectly with our existing views. It’s as if we’re in a room where the walls whisper back our own thoughts. While this might feel reassuring, it often prevents us from encountering diverse viewpoints and leads to a more limited perspective of the world.</p><h3>The reinforcement of confirmation bias</h3><p>Closely linked to these echo chambers in media is the phenomenon of confirmation bias. This is our mind’s natural inclination to believe and seek out information that reinforces our pre-existing beliefs.</p><p>A notable study by the American Psychological Association sheds light on intriguing media consumption patterns: users of AI-curated news feeds have shown a 30% increase in the tendency to gravitate towards news that echoes their ideological leanings. This isn’t just about staying in familiar territory. It’s more like wearing blinders that narrow our view of the world and make us more prone to one-sided stories.</p><h3>Behavioral Shifts and the Rise of Political Polarization</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/392/1*ydWAk0yjJQt7ovEcsO7ghg.png" /></figure><h3>Changing Patterns in News Consumption</h3><p>Have you noticed a shift in what kind of news catches your eye? AI is subtly but surely altering our news consumption habits. A striking figure from the Journal of Communication indicates a 40% surge in the consumption of news that aligns with individuals’ political beliefs after they start using AI-curated feeds. This uniformity in the type of news can narrow our view of the world and reinforce existing divisions.</p><h3>The Influence on Political Perspectives</h3><p>This change in consumption has a significant impact: an increase in political polarization. Research from Stanford University brings a key insight to light. It shows a 25% rise in political polarization that can be traced back to the news feed algorithms. Our online reading habits are not just shaping our news preferences; they’re also coloring our political views and contribute to a more fragmented and divided society.</p><h3>Echo Chambers: Just Like Our Favorite Movie Trailers?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/406/1*mocViygdpRgpw24pwsU46g.png" /></figure><p>You might be wondering what it looks like when someone is stuck in this echo chamber. Well, at our latest family BBQ, my Uncle Joe was passionately arguing about a headline he read. He was claiming that a small town had discovered a “fountain of youth” water source.</p><p>Trying to convince him to consider other news sources, or even think about how this would be possible on a physical and chemical level, was like persuading a cat to enjoy a bubble bath — highly unlikely. Sounds like a humorous story, but it paints a clear picture of how our digital echo chambers can cement our beliefs — no matter how outlandish they may seem.</p><p>Think of your AI-curated news feed as your personal movie trailer editor. It’s great at giving you the action-packed headlines you love, but what about the slow-burn, thought-provoking stories? Just like sticking to one movie genre, a one-track news feed can keep us from the full story spectrum. It’s a bit like only watching thrillers and never knowing the charm of a classic romance.</p><h3>The Weight of Misinformation</h3><p>Nowadays, AI-curated news are everywhere. We often find ourselves bombarded with misleading or sensational news. This steady stream of skewed stories can stir up significant anxiety and mistrust among us.</p><p>Think about it: when you’re bombarded with exaggerated or false news, doesn’t it sometimes twist your view of reality and leave you feeling unsettled or even doubtful? The impact on mental health is substantial. It doesn’t just touch on individuals but even on the very fabric of our communities.</p><h3>Consider our mental well-being</h3><p>But it’s not just about stress and anxiety. The psychological influence of these AI-generated news loops goes deeper. The echo chambers in media and the confirmation biases they create can intensify negative feelings and beliefs. This situation often leads to a society that’s more divided and less understanding of each other. It’s important, now more than ever, to be mindful of these hidden impacts on our mental health as we sift through our digital news feeds.</p><h3>Can We Outsmart AI?</h3><p>One way to combat the effects of AI-curated news is to consciously diversify our news sources. Stepping out of our comfort zones and exploring a variety of news outlets can provide a more balanced view of the world. This practice not only broadens our perspectives but also helps in developing critical thinking skills. These are essential in discerning truth from misinformation.</p><h3>Try this experiment:</h3><p>Why not turn this into an experiment? For a week, try reading news from sources outside your usual preferences. This venture into “enemy territory” can be enlightening and expose you to different viewpoints. It might even challenge your preconceived notions and help you understand other peoples’ perspectives on the matter you’re reading about. It’s like a palate cleanser for your news diet. You get a fresh perspective on the world.</p><h3>AI Transparency and Accountability — Tech Ethics in News Curation</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/374/1*Q-J2f4aWf5Su5CTQejdJyw.png" /></figure><h3>What will futuristic journalism look like?</h3><p>Have you ever wondered what’s behind the AI that curates your news? It’s vital to understand these algorithms. Knowing how they work and their potential biases empowers us to navigate the news with more insight and caution. This understanding is a step towards consuming news more intelligently and responsibly.</p><h3>How do we uphold journalism’s core values?</h3><p>With AI becoming a staple in newsrooms, the spotlight is on ethical journalism. This new era calls for journalists and media houses to stick firmly to accuracy, fairness, and impartiality. After all, the trustworthiness of our news hinges on these principles. Futuristic journalism should not be left in the (unsupervised) hand of AI.</p><h3>Let’s spark a conversation!</h3><p>This blog post isn’t the end but the beginning of an important conversation. As readers and consumers of news, your participation in discussions about AI’s influence on news curation and its wider societal implications is vital.</p><p>So, over to you, dear readers. How have AI-curated news feeds shaped your view of the world? Ever ventured out of your news comfort zone to broaden your perspective? What is your personal opinion on the impact of AI on journalism?</p><p>In contemplating the “tailored” reality AI news feeds offer, let’s think about this: Are we getting a complete picture of the world, or just slices that mirror our own biases?</p><p><em>Originally published at </em><a href="https://mindbytesai.com/the-tailored-truth-ai-news-feeds-and-your-brain-9/"><em>https://mindbytesai.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=883d1355d665" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 Surprising Ways AI Algorithms Control What You See on Social Media]]></title>
            <link>https://medium.com/@MindbytesAI/5-surprising-ways-ai-algorithms-control-what-you-see-on-social-media-c5ec463315b3?source=rss-e5a9853bb3d4------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5ec463315b3</guid>
            <dc:creator><![CDATA[Ashish K Saxena]]></dc:creator>
            <pubDate>Mon, 11 Dec 2023 02:54:40 GMT</pubDate>
            <atom:updated>2023-12-11T02:54:40.449Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="A social media feed influenced by AI personalization" src="https://cdn-images-1.medium.com/max/895/1*pQ5oy204VrRYRVfL6fg2mw.jpeg" /></figure><p>Have you ever wondered why your social media feeds seem to know you better than you know yourself? In a world where every click, like, and share is monitored, AI algorithms are constantly shaping our online experiences. Artificial intelligence knows how to learn what we like and don’t like, ultimately allowing it to reinforce our preferences online. These systems make the behind-the-scenes decisions that determine every post you see in your feed, which comes with surprisingly serious implications.</p><p>Let’s explore how AI algorithms work to understand us and then use that information to influence what we see online.</p><h3>How AI Algorithms Detect User Preferences</h3><p>These detection methods help AI identify and understand user preferences.</p><h3>User Interactions</h3><p>A <a href="https://blog.hootsuite.com/social-media-interaction/">user interaction</a> occurs every time a user takes an action on social media in direct response to content. AI learning models use information from these interactions to better understand what attracts engagement.</p><p>Examples of Interactions Include:</p><ul><li>Likes</li><li>Comments</li><li>Shares</li><li>View Time</li></ul><figure><img alt="Infographic showing how AI algorithms impact user engagement on social media platforms" src="https://cdn-images-1.medium.com/max/896/1*SZtvzWRQFIVPo7RhN4_EEw.jpeg" /></figure><h3>Direct User Input</h3><p>When social media users set up their profiles and engage with the platform, they often share specific inputs. Some of this information will come from the questions that arise when setting up a profile. Others come from search queries and other user actions. Each piece of information helps the system learn more.</p><p>Information like your age or the latest trend you looked up will all factor into what you see in your feed.</p><h3>Behavioral Data</h3><p>For the average social media user, behaviors happen naturally. More often than not, users will not even realize they have specific behaviors when using social media. How much thought do you put in when scrolling, liking, and sharing?</p><p>Artificial intelligence observes how we behave in real-time. In fact, <a href="https://snowplow.io/what-is-behavioral-data/">behavioral data</a> helps AI systems to better understand engagement and preferences on an individual basis and across communities.</p><p>Examples of Behaviors Include:</p><ul><li>Scrolling Patterns</li><li>Click-Through Rates</li><li>View Time on Specific Pieces of Content</li></ul><h3>Common Machine Learning Models for Social Media</h3><p>These top machine-learning models influence digital spaces daily. Are they influencing you?</p><figure><img alt="Machine learning processes used in social media algorithms" src="https://cdn-images-1.medium.com/max/896/1*wfud2SpcR_JphbS-qCx9fA.jpeg" /></figure><h3>Predictive Analytics</h3><p>When it comes to looking to the future, <a href="https://online.hbs.edu/blog/post/predictive-analytics">predictive analytics</a> play a leading role. These data-backed developments rely on monitoring past user behavior. Based on this information, the machine learning models begin to make informed guesses regarding what the user will do or what kind of content they will enjoy.</p><p>If you’re always liking videos of adorable animals, AI knows to give you more!</p><h3>Natural Language Processing (NLP)</h3><p>Language defines cultures, and it defines subcultures as well. For this reason, platforms use artificial intelligence to better understand how users interact with textural content. Using <a href="https://www.ibm.com/topics/natural-language-processing">NLP</a>, AI systems explore the types of language users interact with and find other content that aligns with these interests.</p><p>Have you ever noticed that watching videos of text posts often leads the algorithm to share more content that touches on those same topics?</p><h3>Image and Video Analysis</h3><p>Social media is a multimedia game, and images and videos bring it to life. While these pieces of content attract new users, they also provide a perfect reference point for community analysis. Using <a href="https://www.techtarget.com/whatis/definition/object-recognition#:~:text=Object%20recognition%20is%20the%20area,video%20and%20still%20camera%20images.">object recognition</a>, AI learns more than you might think about different types of content.</p><p>When <a href="https://premioinc.com/blogs/blog/what-is-ai-based-video-analytics#:~:text=AI%2DBased%20Video%20Analytics%20uses,how%20computers%20can%20%E2%80%9Csee%E2%80%9D.">analyzing images and videos</a> on a platform, artificial intelligence identifies clear themes and cross-references that information with user habits. This creates a comprehensive understanding of the patterns in content pieces that have high engagement.</p><h3>Collaborative Filtering and Similarity Indexing</h3><p>Every user is unique, but user preferences often come in trends. Using AI algorithms that focus on user reviews and information and then applying that to other similar user profiles helps provide tailored content for a better user experience. By aligning users with similar preferences, this approach adds more significance to top trends and helps people with similar profiles and behaviors find content they love.</p><h3>How AI Reinforces User Preferences</h3><p>Tailored to you or limiting what you see? AI reinforces user preferences and limits exposure to certain viewpoints.</p><h3>Content Personalization</h3><p>The big difference between social media platforms when they first started and the social media platforms of today is the introduction of <a href="https://www.semrush.com/blog/personalized-content/">content personalization</a>. While users used to have to deal with seeing whatever content was recent or trending, today’s user experience focuses on personalization.</p><p>With personalization, users see content that they want to see more often than not. Through the acts of identifying preferences and lifting up information from relevant communities, feeds give users what they want to see, even as their moods change.</p><p>If you have ever felt like purchase recommendations can read your mind, you know this goes well beyond the kind of content you see. Targeted advertising lets brands lock onto your preferences to market products to you directly.</p><h3>Feedback Loops</h3><p>Modern social media platforms create <a href="https://theconversation.com/feedback-loops-and-echo-chambers-how-algorithms-amplify-viewpoints-107935">feedback loops</a> using engagement metrics and reinforcement learning. In feeds, content that receives more engagement also receives more views, making algorithms favor certain content pieces over others.</p><p>As the algorithm observes the preferences and reactions of users, it will adapt to maximize engagement based on the information it collects.</p><h3>Echo Chambers and Filter Bubbles</h3><p>Given the tendency for artificial intelligence to highlight certain pieces of content for specific user groups, two clear risks arise — the creation of <a href="https://edu.gcfglobal.org/en/digital-media-literacy/what-is-an-echo-chamber/1/">echo chambers</a> and <a href="https://nordvpn.com/blog/what-is-filter-bubble/">filter bubbles</a>.</p><p>These problems happen when users engage with limited content that reinforces their beliefs and preferences. While this is not always bad, it may limit exposure to diverse perspectives and perpetuate misinformation sharing.</p><p>At the end of the day, no one should only see views they personally agree with. Society needs larger discussions with multiple viewpoints.</p><h3>Implications and Challenges Faced by Platforms and Users</h3><p>Creating more ethical online spaces requires platforms and users to work together to improve information sharing and access.</p><figure><img alt="Diverse group of people engaging with social media platforms influenced by AI technology" src="https://cdn-images-1.medium.com/max/896/1*4vgx4lGdS7fuXeBs6lgVCg.jpeg" /></figure><h3>Ethical Considerations</h3><p>Social media impacts individuals, groups, and belief systems. As these realizations come to light, ethical considerations become more important than ever before. Both businesses and users must understand the potential for <a href="https://www.allerin.com/blog/ai-can-reinforce-existing-human-biases">reinforcing bias</a>, as well as the potential for creating polarized online communities.</p><p>Underneath these issues, another concern lurks — user privacy. As social media platforms use AI to create feeds based on user profiles, users give up personal insights into themselves and their preferences. For this reason, we must think critically about data collection and user profiling to create more ethical AI. Already, privacy concerns in social media lead to important discussions every day.</p><p>In some ways, AI may know more about your tendencies than you do–and this can be very dangerous.</p><h3>Transparency and Accountability</h3><p>Social media platforms and users alike must acknowledge the risks that algorithms pose for the general public. There needs to be a clear understanding of how these algorithms work, as well as their impact on user behaviors online and in the real world.</p><p>To create truly ethical AI, we must push back against unintentional discrimination and <a href="https://insights.som.yale.edu/insights/how-social-media-rewards-misinformation">misinformation sharing</a>.</p><h3>User Awareness and Control</h3><p>Empowering users to manage their data and understand the implications of these data profiles provides an excellent approach to creating more ethical AI. Since AI algorithms play a pivotal role in shaping the user experience, we all must think more critically about what we reinforce through our user preferences.</p><p>Educated users are users who can help to improve these algorithms and create stronger and more diverse online communities.</p><p><strong>What does the future of AI and social media mean to you? Share your thoughts on how we can create better digital spaces below!</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5ec463315b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>