<?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 Society on AI Governance and Ethics on Medium]]></title>
        <description><![CDATA[Stories by Society on AI Governance and Ethics on Medium]]></description>
        <link>https://medium.com/@saige_81207?source=rss-6f4908bb22ed------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*hQBPK1t_RFSVqYAOAtF1UA.jpeg</url>
            <title>Stories by Society on AI Governance and Ethics on Medium</title>
            <link>https://medium.com/@saige_81207?source=rss-6f4908bb22ed------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 20:22:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@saige_81207/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[Pulped for Progress? Book Destruction in Project Panama]]></title>
            <link>https://medium.com/@saige_81207/title-pulped-for-progress-book-destruction-in-project-panama-287acc4aa5c9?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/287acc4aa5c9</guid>
            <category><![CDATA[case-study]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-and-law]]></category>
            <category><![CDATA[law]]></category>
            <category><![CDATA[infringement]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 10:46:46 GMT</pubDate>
            <atom:updated>2026-10-29T17:33:56.053Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YHj-GYCT0hRAbYuA.jpg" /></figure><p><strong>Author’s name:</strong> Hari Menon</p><p><strong>Introduction:</strong></p><p>In early 2024, Anthropic launched Project Panama, the premise of which was to acquire physical books, cut off their bindings, scan the pages and subsequently recycle the originals. This process has been coined, ‘destructive printing’, its objective being to obtain a large knowledge base from physical books, many of which are old texts no longer in circulation. Internal documents detailing the project had emerged in US copyright litigation, asking the simple question of whether purchasing and destroying a book make its reproduction for AI training lawful?</p><p><strong>The Indian Position:</strong></p><p>From an Indian legal perspective, section 14 of the Copyright Act 1957 states that copyright holders reserve the exclusive right to reproduce their own works. The purchasing of such a book does not bestow the ability to digitize the text in this sense, the process of which counts as reproduction regardless of the destruction at the end.</p><p>However, whether such reproduction constitutes infringement depends partly on the exceptions available under Section 52. This issue has gained particular relevance following ANI Media Pvt Ltd v. OpenAI OpCo LLC, decided by the Delhi High Court on 24 July 2026. The Court, in this case, declined to grant ANI an interim injunction against OpenAI, holding, on a prima facie basis, that storing ANI’s works for AI training fell within section 52(1)(a), which covers fair dealing for private or personal use, including research. The Court’s observations however, remain preliminary, and the underlying litigation continues.</p><p><strong>The US Position:</strong></p><p>This case provides us a framework to examine Project Panama. In The case of Bartz v. Anthropic (June 2025), a US federal court created a distinction. Prior to the establishment of Project Panama, Anthropic relied primarily on piracy to source literary works, scalping works from large providers such as LibGen. Anthropic is said to have committed a fault in such procurement and eventually settled for a sum of 1.5 billion dollars. However, Project Panama itself was held to be fair, considered as transformative use.</p><p>While the observations in these judgements align in recognising that there exist exceptions provided under copyright law throughout the world for fair use and transformative works, the question before us is whether these judgements are fair. For business tactics and secrecy purposes, Project Panama was deliberately kept hidden. The part of destruction of the physical books in the printing process has no mention in Bartz v. Anthropic in relation to the final judgement, yet still remains an vital fact to understand when considering the consequences of such judgement.</p><p><strong>Conclusion:</strong></p><p>Anthropic and other large AI companies have the resources to procure large volumes of works from physical bookstores, and removing old books that are no longer being produced amounts to erasing works, and therefore, knowledge from the world. While destruction as a ground has not been identified earlier in relation to works, the question of whether it ought to be permitted arises when such destructive purpose serves as a means to eliminate competition from using the same works to train their AI models.</p><p>Should destructive methods of training for AI models be cause for amendments to copyright law? This is the question that law ought to answer in the days to come.</p><p><strong>Author’s Bio:</strong></p><p>Hari Menon is a first year student at Bits Law School, Mumbai pursuing his B.B.A LL.B (Hons.) undergraduate degree. He is interested in how older judgements and statutes become applicable when questioned with new developments, particularly with AI.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=287acc4aa5c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Headnotes to Neural Networks: Interpreting Fair Use in the Age of Generative AI]]></title>
            <link>https://medium.com/@saige_81207/from-headnotes-to-neural-networks-interpreting-fair-use-in-the-age-of-generative-ai-811a929a7bda?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/811a929a7bda</guid>
            <category><![CDATA[copyright-and-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[case-analysis]]></category>
            <category><![CDATA[copyright]]></category>
            <category><![CDATA[copyright-infringement]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Wed, 21 Oct 2026 16:44:25 GMT</pubDate>
            <atom:updated>2026-10-21T16:44:25.391Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/930/1*FQyzb9z-XAeypv2ZUg0xNg.jpeg" /></figure><p><strong><em>~By Rakshat Gupta</em></strong></p><p><strong><em>Keywords: Fair Use, AI Training, Transformative Use, Copyright</em></strong></p><p><em>“A block of raw marble, like a judicial opinion, is not copyrightable. Yet a sculptor creates a sculpture by choosing what to cut away and what to leave in place. … A headnote is a short, key point of law chiseled out of a lengthy judicial opinion.”</em></p><p><strong><em>Keywords: Fair Use, AI Training, Transformative Use, Copyright</em></strong></p><p>Thomson Reuters owns Westlaw, whose editorial headnotes summarize points of law in judicial opinions. Ross Intelligence, a competitor, commissioned a third-party to prepare approximately 25,000 training memoranda for non-generative legal search engines, which drew on these Westlaw headnotes, of which 2,243 were held to have been copied by the U.S. District Court for the District of Delaware. The Third Circuit affirmed that the headnotes satisfy the originality threshold and Ross’s usage could not constitute ‘<em>fair use</em>’ under §107 of the Copyright Act, 1976. This case demonstrates how difficult it is to distinguish technological use of copyrighted material from commercial function of the product that is ultimately developed. Although defensible on the facts, the decision exposes an unresolved doctrinal tension between intermediate technological uses and substitutive uses and must be construed as a fact specific application of the fair use doctrine, not as a categorical prohibition on AI training.</p><p>§107 of the Act weighs four factors: purpose and character of the use, nature of the work, amount taken, and its effect on the market. The first and fourth factors carry the analysis. On purpose and character, the court noted that Ross’s use was commercial and then asked whether it was transformative, balancing the degree of difference in purpose against commerciality as prescribed in <em>Warhol v. Goldsmith</em>. Ross contended that the headnotes were merely an intermediate input for a system returning only judicial text. The Court conceded that this step arguably put forth a slight difference but held that Ross trained its system for the benefit of a platform which shared ‘the same ultimate purpose’ as Westlaw, which rendered the use minimally transformative at best.</p><p>Under the fourth factor, the Court identified harm in two markets. Firstly, legal research platforms, wherein Ross competed directly and secondly, an emerging market for licensing headnotes, as training data which Thomson Reuters had not by itself entered but which Ross’s unauthorized copying pre-empted. The second factor favoured fair use slightly. The third factor tied to the absence of justifying transformative purpose, did not.</p><p>Characterising purpose at the level of the platform fixed Ross as a substitute. That framing then rendered the copying unnecessary under the third factor and market injury was particularly self-evident under the fourth. The first factor in effect determined the remainder.</p><p>The Court’s conclusion draws its strongest support from the facts themselves. Ross openly aimed to supplant Westlaw and won some of its customers, a Court deciding on undisputed evidence of such intent was thus entitled to treat commercial rivalry as having a bearing on both purpose and market effect. Under <em>Warhol</em>, shared purpose and commerciality of the use are likely to weigh against the copier unless further justification exists.</p><p>The analytical difficulty lies in the level of abstraction at which purpose is identified. <em>Warhol </em>directs the first factor towards the specific use alleged to be infringing, not the defendant’s enterprise at large. The specific use here was the incorporation of headnote text into memoranda from which a model leaned to correlate queries with opinion passages. By equating that use with the finished product’s function, the Court treated intermediate copy as indistinguishable from the end service. The difference conceded was never quantified. The contrasting approach in <em>Sega Enterprises v. Accolade Inc. </em>is instructive, wherein the Ninth Circuit assessed intermediate copying by reference to its purpose of discovering unprotected functional elements, rather than treating the commercial function of the rival games ultimately marketed as dispositive.</p><p>This decision foregrounds the fundamental question of whether a novel technological practice should be accommodated through the interpretation of existing statutory exceptions or through specific legislative intervention. Courts must apply statutes as enacted and legislative intervention becomes necessary where the proposed permission cannot reasonably be grounded in the existing framework, a distinction the Third Circuit maintained by expressly treating the dispute as “no more than an ordinary copyright case”.</p><p>Ross rejects fair use on the facts before it, where a commercial competitor appropriated original editorial work despite the availability of unprotective alternatives and failed to rebut evidence of market harm. Yet its reasoning exposes a deeper doctrinal fault line by measuring technological transformation through the finished product rather than the purpose of copying itself. Its holding must remain confined to the statutory framework applied, lest a fact specific judgement be mistaken for a categorical prohibition on AI training. The question copyright law must now confront is whether protecting original expression can justify granting its owners effective control over the technological uses of that expression or whether such control risks converting copyright from an instrument of creative incentive into a barrier for technological progress.</p><p>Author’s Bio: Rakshat Gupta B.A. LL.B. (Hons.) is a student at BITS Law School, Mumbai, interested in Jurisprudence, Intellectual property law. His work explores the doctrinal and institutional challenges posed by emerging technologies, particularly the intersection of technological development and the limits of existing legal frameworks.</p><p><strong>References</strong></p><p><strong>Cases</strong></p><p><a href="https://www.supremecourt.gov/opinions/22pdf/21-869_87ad.pdf">Andy Warhol Foundation for the Visual Arts Inc v Goldsmith 598 US 508 (2023)</a></p><p><a href="https://law.resource.org/pub/us/case/reporter/F2/977/977.F2d.1510.92-15655.html">Sega Enterprises Ltd v Accolade Inc 977 F 2d 1510 (9th Cir 1992)</a></p><p><a href="https://chatgptiseatingtheworld.com/wp-content/uploads/2026/09/Thomson-Reuters-v.-ROSS-Intelligence-252153p.pdf">Thomson Reuters Enterprise Centre GmbH v Ross Intelligence Inc No 25–2153 (3d Cir, 29 September 2026)</a></p><p><strong>Legislation</strong></p><p><a href="https://www.govinfo.gov/content/pkg/USCODE-2024-title17/html/USCODE-2024-title17-chap1-sec107.htm">Copyright Act 1976, 17 USC § 107 (United States)</a></p><p><strong>Secondary Sources</strong></p><p><a href="https://www.authorsalliance.org/2025/02/13/thomson-reuters-v-ross-the-first-ai-fair-use-ruling-fails-to-persuade/">https://www.authorsalliance.org/2025/02/13/thomson-reuters-v-ross-the-first-ai-fair-use-ruling-fails-to-persuade/</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=811a929a7bda" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How India’s outdated Copyright Laws fail to address the use of publicly available content in AI…]]></title>
            <link>https://medium.com/@saige_81207/how-indias-outdated-copyright-laws-fail-to-address-the-use-of-publicly-available-content-in-ai-db607ed91b11?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/db607ed91b11</guid>
            <category><![CDATA[public-policy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[supreme-court]]></category>
            <category><![CDATA[copyright]]></category>
            <category><![CDATA[ai-regulation]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 14:00:39 GMT</pubDate>
            <atom:updated>2026-10-12T14:00:39.529Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>How India’s outdated Copyright Laws fail to address the use of publicly available content in AI training?</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tmTStayg2jnbE6KYIIMshQ.jpeg" /></figure><p><strong><em>By Lalithaanjali Gudimetla</em></strong></p><p><strong>Keywords: AI training, copyright, data scraping, reproduction, legality</strong></p><p>The onset of the technological revolution through the development of Artificial Intelligence (AI) has converted an abstract futuristic possibility into an everyday reality. The pursuit of highly capable AI models in our everyday lives, striving to increase efficiency, productivity and capacity have created an insatiable demand for data to train these models. However, this growing trend towards data-oriented technologies points to the growing disparity of regulation in India and all around the world. This disparity can specifically be highlighted in India from the recent case of <em>ANI Media Pvt. Ltd. v. OpenAI.</em> The factual circumstances included ANI suing the latter in 2024 for using its published news reports and articles without permission or payment to train its Large Language Models (LLMs) — an advanced AI system trained on vast amounts of text to understand, summarize, translate, and generate human-like language. This dispute has raised significant questions of whether the Indian copyright law of 1957 is equipped to govern the increasingly complex nature of disputes surrounding copyrighted works and AI training.</p><p>Generally, AI training includes teaching an AI model to recognise patterns and information by feeding it massive amounts of data in order for it to generate content and make decisions. This data is often scraped from public sources like the Internet and involves making temporary copies of the data, for the model to fully understand and learn from it. The grey area surrounding such AI training and copyright infringement claims is directly linked with the false assumption that since data is publicly available, it is freely available for reproduction and training purposes.</p><p>This assumption is expressly negated by the Copyright Act of 1957, where Section 14 states that copyright holders possess the exclusive rights to reproduce, store, and communicate their work to the public. This means that a news article or news commentary, which is publicly available, is still copyrighted and its reproduction is regulated by the law. Therefore, the question that remains is whether scraping of copyrighted material from the internet to train AI models constitutes copyright infringement under the Copyright Act of 1957. This is the contestation that arose in the case of ANI v OpenAI wherein it was contested by ANI that OpenAI trained its chatbot, ChatGPT using the former’s copyrighted news content to answer user queries and also alleged that the chatbot generated fake stories and fabricated news while falsely attributing the source to ANI, threatening its credibility and reputation. Yet the scope of Section 14 falls short in answering whether temporary storing and training of models with publicly available data counts as reproduction and copying.</p><p>Given the stated contestation, it must be noted that Section 52 of the Copyright Act permits the use of copyrighted works for purposes such as research, criticism, private study and review. However, the legislation falls short in addressing the legality of using copyrighted material for AI training. The scope of Section 52 has been restrictive in answering the legal ambiguity that arises at the intersection between the law developed even before the advent of prevailing technologies and newer issues that come into light with this development.</p><p>This legal ambiguity, as highlighted above, cuts both ways, affecting the holders of these copyrighted works and the ones who rely on these works to build and advance the efficiency and capacity of their AI models. On one hand, the absence of a decisive regulation on AI companies scraping copyrighted works for training threatens creators’ economic rights and dilutes the value and credibility of their work (Joshi). On the other hand, over-regulation of copyrighted works threatens to curb the development of new technologies and AI systems. In order to resolve this dispute, it is necessary to provide clear, statutory regulations on this matter (Joshi). Additionally, training of AI models by scraping of public data may be enabled, subject to certain conditions that safeguard the rights of copyright holders, allowing them to reserve their rights for their works to be excluded from data scraping (Joshi).</p><p>In conclusion, the fundamental source of conflict within Indian copyright laws must be recognised and sought to be legislated upon in consultation with various stakeholders involving legal and technical professionals from both fields. Although the Indian Copyright Act, 1957, provides an adequate framework for regulation of Copyrights in India, its scope and ambit must be extended to include the rising nuances of conflict surrounding AI and Intellectual Property. India can draw upon legislative developments around the world in order to balance and meet the needs of all those involved, balancing technological advancement with Intellectual Property Rights. For example, India can draw focus to the European Union’s Copyright Directive (2019/790) that permits commercial data scraping from the internet while also permitting an opt-out for the copyright holders to preserve their right to reserve, which plays the balancing act as opposed to the working paper proposed by the Department for Promotion of Industry and Internal Trade which recommends a mandatory licensing model with no option for copyright holders to opt out. Until India clears this ambiguity in the application of the Act of 1957, parties in disputes over such infringements would have to look to the courts to draw clear interpretations guided by the new-found notions of public interest; rendering too wide and unclear to provide real answers (Lath).</p><p>Author bio: Lalithaanjali Gudimetla is a member of the SAIGE writing team at BITS Law School, Mumbai. While pursuing her Bachelors in LLB (Hons.), she strongly believes in the impact of AI in numerous fields while considering the ethical and governance challenges posed by its rapid advancement.</p><p><strong>References used</strong></p><p>1. <em>ANI Media Pvt. Ltd. v. OpenAI OpCo LLC</em>. CS(COMM) 1028/2024</p><p>2. India, Ministry of Commerce and Industry. “DPIIT Publishes First Part of Working Paper on AI–Copyright Interface.” <em>Press Information Bureau</em>, 9 Dec. 2025, <a href="http://www.pib.gov.in/PressReleasePage.aspx?PRID=2200741">www.pib.gov.in/PressReleasePage.aspx?PRID=2200741</a>. Accessed 6 Oct. 2026.</p><p>3. Mella, Victor, and Juan Ramón Robles. “What Do Companies Need to Know about the Data Mining Exception of the Copyright Directive?” <em>Hogan Lovells Cadwalader</em>, 10 Mar. 2021, <a href="http://www.hlc.com/en/publications/what-do-companies-need-to-know-about-the-data-mining-exception-of-the-copyright-directive_1">www.hlc.com/en/publications/what-do-companies-need-to-know-about-the-data-mining-exception-of-the-copyright-directive_1</a>. Accessed 6 Oct. 2026.</p><p>4. Joshi, Moomal. “Web Scraping and AI Training Data: Copyright Challenges in the Age of Generative AI.” <em>Khurana and Khurana</em>, 29 Apr. 2026, <a href="http://www.khuranaandkhurana.com/web-scraping-and-ai-training-data-copyright-challenges-in-the-age-of-generative-ai">www.khuranaandkhurana.com/web-scraping-and-ai-training-data-copyright-challenges-in-the-age-of-generative-ai</a>. Accessed 8 Oct. 2026.</p><p>5. Lath, Aparajita. “Public Interest, Copyright and Fair Dealing: <em>ANI v OpenAI</em>.” <em>Supreme Court Observer</em>, 19 Aug. 2026, <a href="http://www.scobserver.in/journal/public-interest-copyright-and-fair-dealing-ani-v-openai/">www.scobserver.in/journal/public-interest-copyright-and-fair-dealing-ani-v-openai/</a>. Accessed 7 Oct. 2026.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db607ed91b11" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Socioeconomic Impact of AI-Generated Creativity in Copyright Law]]></title>
            <link>https://medium.com/@saige_81207/the-socioeconomic-impact-of-ai-generated-creativity-in-copyright-law-99daba4b5455?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/99daba4b5455</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[socioeconomic-disparity]]></category>
            <category><![CDATA[copyright-law]]></category>
            <category><![CDATA[ai-policy-and-regulation]]></category>
            <category><![CDATA[ai-ethics]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 17:00:21 GMT</pubDate>
            <atom:updated>2026-10-05T17:00:21.234Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O4m4fda6K1q-izn94Oe4cQ.jpeg" /></figure><p>Keywords: AI, creativity, copyright, socioeconomic, legal</p><p>Creativity is widely recognized as one of the hallmark aspects of the workings of the human mind. Our ability to bring imaginative ideas to life in the form of creative pursuits such as art and literature, remains one of the most important distinguishing features of human beings. The question then arises, what happens when a machine exercises creativity derived from human thinking? This article argues that AI generated creativity is not truly accessible to all but is ridden with social disparities.</p><p>It is well known within the sphere of intellectual property rights that “human creativity is the sine qua non at the core of copyrightability” (<em>Thaler</em>, 687 F. Supp. 3d at 146). Copyright is essentially a form of legal protection offered to human creations. Not only does it provide the original creator with the right to make profit from their own works; it awards due credit to the rightful owner for their creation. However, the issue of creativity and copyright is confounded by the advent of artificial intelligence. This machine, which derives its innovative abilities from human inputs, brings up unique complications. On one hand, all creative work is technically derived from existing ideas, so it can be said that AI possesses its own creativity much like humans. On the other hand, AI’s creativity stems from reaching a specific goal, unlike human creativity which flows from original experiences, emotions and relationships with other beings. Does this emotional quality not contribute significantly to the originality of the creator’s works? These questions can only be answered by further investigation into the differences between AI’s thinking process and that of humans. Considering the fact that AI is controlled by large corporations, concerned with profit-generation, conferring the same copyright laws on AI risks alienating small artists who cannot afford equal access to the legal system.</p><p>While it is necessary to consider the legal definition of creativity, the socioeconomic aspect of copyright laws is of paramount importance to the project of AI governance. There already exists an enormous disparity in access to copyright protection. It is often the case that owners and creators are not one and the same. Large corporations, with the ability to afford the best of legal representation, have the ability to exercise their intellectual property rights far more effectively than individual self-employed artists and creators. Conferring copyright protection on AI works without appropriate regulations can further compound this disparity. A fear of replacement of human labour by AI is already pervasive among the creative community. AI poses a major threat to job security, particularly in the creative field. If tech giants gain ownership rights over AI creations that feed off existing human works, is it ethical to afford the same protections to AI works as is granted to those made by humans? The risk of compromising a level playing field for the sake of profit-generation and efficiency is one that does not outweigh the benefits of generative AI.</p><p>In the grand scheme of AI development, it is important to recognize the ethical implications of this new relationship between humans and machines. One of the novel challenges that AI has brought with it is the socioeconomic imbalance compounded by unequal access to legal protection. Overall, there must be a good balance created between human interests and the rapidly advancing technology in this new age of Artificial Intelligence.</p><p><strong>Author’s Bio:</strong></p><p>Lasya Guha is a member of the writing team for SAIGE in Bits Law School, with an understanding of the importance of Artificial Intelligence and its impact on modern society. She understands that it comes with its disadvantages but believes that with proper regulation it can be used to help society.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99daba4b5455" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who Is Behind the Wheel?]]></title>
            <link>https://medium.com/@saige_81207/who-is-behind-the-wheel-488cb660d92e?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/488cb660d92e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Thu, 13 Aug 2026 17:05:18 GMT</pubDate>
            <atom:updated>2026-08-13T17:05:18.686Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Autonomous Vehicles, Legal Vacuums, and the Unfinished Architecture of Accountability</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/976/0*B_oeLzXoHELJ6mzq.jpg" /></figure><p>In March 2018, Elaine Herzberg became the first known pedestrian fatality involved in a self-driving vehicle accident. An Uber — operated Volvo SUV struck and killed her while she was wheeling her bicycle across the street. Investigators said that car’s safety driver, Rafael Vasquez, had been streaming an episode of the Voice at the time. According to reports, the vehicle sensors detected her presence 6 seconds before the impact but misjudged it and took no corrective measures. Mr. Vasquez pleaded not guilty and was released to await the trial. The Arizonian Court ultimately charged the driver with negligent homicide while Uber, the principal architect, was acquitted “as there was no basis for the criminal charges being pressed upon the corporation.” The precedent it has set is as unsettling as the tragedy itself. This single case hints at a far larger crises: The law is playing catch-up in a race it never signed up for.</p><p><strong>The Illusion of Safety through Automation</strong></p><p>Self-driving cars are one of the most consequential experiments of our era but remains also a legally uncharted arena. Its promise is straightforward: replace the human drivers with sensors, data, and algorithm and the fatalities will go down too. According to the National Highway Traffic Safety Administration (NHTSA), 94% of road fatalities in United States are attributable to the fallible human driver error. Remove the driver and the risk automatically gets uprooted. Yet the reality that has emerged from the lawsuits, product liability, and engineering audits is far from the utopian promises, one in which the automated cars inherit the drivers position without inheriting legal liabilities surrounding them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/964/1*zfpmOrXs9gltVwKereCHWw.png" /></figure><p>To understand how the legalities work, one needs to understand the how the self-driving car is architecturally composed. Harvard’s analysis suggests that the car is composed of three interlocking systems. Firstly, the mechatronic system that encompasses the physical hardware like LiDAR sensors, cameras, radar arrays, and actuators. Secondly, the autonomous mechatronic system which is the software brain comprising of the perception algorithms, convolutional neural networks (CNNs), object classifiers, Markov Decision Processes, and reinforcement learning modules that translate sensor data into driving decisions. And thirdly, the cross-linked mechatronic system connects the vehicle to the cloud infrastructure, other vehicles, and a centralized HD mapping system. Each layer introduces its own set of legal liabilities.</p><p>Consider the sensor stack alone. GPS (Global Positioning System) alone operates at the positional error margin of one meter and updates at only 10 times per second (10 Hz. It sounds fast until you are driving a car at 100Km/Hr. (27.7 meters per second) and with the 10Hz frequency you will receive position updates at 0.1 seconds meaning that the car travels at almost 3 meters between updates. This is fine for navigation but for collision avoidance which requires precision of every centimeter and milli-second reaction time, it’s weak. So autonomous vehicles need other sensors apart from the GPS alone.</p><p>Now comes LiDAR (Light detection and ranging). They work like a spinning laser radar firing laser pulses and calculating time they take to bounce back. From that timing, it constructs a 3-D map of the surroundings High-end systems like the Leica RTC360 are around $100,000. This means that the sensors alone cost around average of two cars.</p><p>Now the third component, Cameras. Human eyes can handle enormous light differences given the presence of specialized photoreceptor cells (rods and cones). Machines struggle with this. Dynamic range (It determines whether the cameras can simultaneously see the road in shadow and bright light without going blind or losing detail) is measured in Decibels, and to operate safely on roads, cameras need around 130 Db of Dynamic range to handle situations like bright sunlight, dark shadows and the light being reflected back from the metal surfaces or glass. High End automotive system capable of these costs around 16000 USD for a systematic 8-camera array.</p><p>The 2016 Tesla Model -3 fatality in Florida is a definitive case study here. The vehicle’s autopilot system failed to detect the white trailer against the bright sky, and the car drove beneath it at the full speed. This can’t be just dismissed as a sensor failure, but it is a failure of epistemic design. No manufacturer fully discloses the boundary conditions of its system as to when the system is bound to fail, the situations in which the algorithm will hand over the control to the driver. This information gap is not incidental, but it is a liability shield.</p><p><strong>The Code</strong></p><p>Every autonomous vehicle carries a set of pre-coded instructions. When the collision is unavoidable who does the car prioritize, the passenger, the motor vehicle, or the pedestrian? These are not some trolley problems thought experiments, but engineering decisions buried in the proprietary codes. The legal system is built for the human driver error and is completely unprepared to adjudicate on the instances when the decisions were premeditated.</p><p>Outside observers face even more ambiguity. Watching a vehicle in the traffic, one cannot tell whether the person sitting behind the wheel is in control or merely sitting. The social cues we rely on — eye contact at the crossroads or yielding a wave become meaningless if an algorithm is the driver. The law however still assigns the liability on the driver even its role has been reduced.</p><p><strong>The Semiconductors shortage, Infrastructure and the connectivity traps</strong></p><p>The viability of the autonomous vehicle rests on the infrastructure that is itself fragile. The systems in place depend heavily on the real time services provided by the Apple Maps and Google Maps, with no credible third alternative providing services with such comparable precision. Inertial measurement Units (IMUs), which provide the centimeter level accuracy employed in aerospace applications which is best suited for our cause, but the exorbitant cost and miniaturization still pose as a challenge.</p><p>Adding on to this, we have the global semiconductor shortage. It has already demonstrated that a disruption in the supply chain can halt the automative production line overnight. A fully networked fleet of automative cars connected to the central grid can turn out to be a catastrophe from a cybersecurity perspective if a bad actor penetrates it whether a state level adversary or a ransomware attack, the consequences won’t be restricted to just data breaches.</p><p>Starlink might be the answer to all our connectivity issues. However, concentrating the navigational sovereignty in the hands of a private actor whose regulatory authority remains at best contested won’t be a wise decision.</p><p><strong>Conclusion</strong></p><p>The Herzberg case ended with a human conviction question as to who bears the responsibility if a pre coded machine decides to kill people. The legal system came up with the answer that was administratively convenient until the legislators agree on the transparency owed to the public about the system boundaries, ethical encodings, and the data collection practices and cyber security systems. Until then, the autonomous vehicle will just be a remarkable technological phenomenon driving through a legal vacuum, at speed.</p><p><em>Written by Rakshat Gupta.</em></p><p><a href="http://www.linkedin.com/in/rakshat">www.linkedin.com/in/rakshat</a></p><p>Works Cited</p><p>1.) Chung, Joanna. <em>Mind, Machine, and Society: Legal and Ethical Implications of Self-Driving Cars.</em> 2018. Harvard University, DASH (Digital Access to Scholarship at Harvard), <a href="http://nrs.harvard.edu/urn-3:HUL.InstRepos:38811556.">http://nrs.harvard.edu/urn-3:HUL.InstRepos:38811556.</a></p><p>2.) PA Tech. <em>“Self-driving Cars Coming to a Slow Lane Near You, Government Announces.”</em> BBC Science Focus, 29 Apr. 2021, <a href="https://www.sciencefocus.com/news/self-driving-cars-coming-to-a-slow-lane-near-you-government-announces">https://www.sciencefocus.com/news/self-driving-cars-coming-to-a-slow-lane-near-you-government-announces</a>.</p><p>3.) <em>“Uber’s Self-Driving Operator Charged over Fatal Crash.”</em> BBC News, 16 Sept. 2020, <a href="https://www.bbc.com/news/technology-54175359">https://www.bbc.com/news/technology-54175359</a>.</p><p>4.) Stamp, Helen. “The Reckless Tolerance of Unsafe Autonomous Vehicle Testing: Uber’s Culpability for the Criminal Offense of Negligent Homicide.” Case Western Reserve Journal of Law, Technology &amp; the Internet, vol. 15, no. 1, 2024, pp. 37–, <a href="https://scholarlycommons.law.case.edu/jolti/vol15/iss1/2">https://scholarlycommons.law.case.edu/jolti/vol15/iss1/2</a>.</p><p><em>5.) </em>Schweber, Bill. <em>“The Autonomous Car: A Diverse Array of Sensors Drives Navigation, Driving, and Performance.”</em>Mouser Electronics, 2018, <a href="https://www.mouser.com/applications/autonomous-car-sensors-drive-performance/">https://www.mouser.com/applications/autonomous-car-sensors-drive-performance/</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=488cb660d92e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The legal implications of GEMA v OpenAI for training of Generative AI]]></title>
            <link>https://medium.com/@saige_81207/the-legal-implications-of-gema-v-openai-for-training-of-generative-ai-75c40317c53f?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/75c40317c53f</guid>
            <category><![CDATA[legal]]></category>
            <category><![CDATA[law]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Mon, 09 Mar 2026 07:11:00 GMT</pubDate>
            <atom:updated>2026-03-09T07:11:00.798Z</atom:updated>
            <content:encoded><![CDATA[<p>Written by Maansi Singhal ‘28</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*x5mXbIUnYVleEjNA" /><figcaption>Image Source: TechnoLlama</figcaption></figure><p><strong>Introduction</strong></p><p>Recently, the 42nd Civil Chamber of the Landgericht München I in the case of <em>GEMA v OpenAI </em>tried to answer the question of whether use of copyrightable material for AI training infringes on the rights of the copyright holder.<a href="#_ftn1">[1]</a> Particularly, in this case, GEMA, a collecting society representing the lyricists and composers, brought claims against two companies of the OpenAI group. It was alleged that the OpenAI’s language models (including GPT-4 and GPT-4o) had memorised the lyrics of various songs, for which GEMA holds the exclusive licensing rights. The Munich Regional Court gave the judgement by stating that OpenAI is infringing the copyright by training and operating ChatGPT and the Text-and-Data Mining (TDM) exception does not apply.<strong> </strong>This piece aims to examine the legal reasoning behind the judgment, as it advocates for creator rights in the AI and copyright debate and assesses its implications for generative AI companies, particularly regarding training data practices.<a href="#_ftn2">[2]</a></p><p><strong>OpenAI’s defence</strong></p><p>OpenAI did not contest the use of the copyrighted works while training its language model. Instead, its defences included the argument that: (1) language models do not “store” specific works but encode the work into statistical correlations. The work is stored as “technical copies” in the memory of the model and the saved model cannot be “opened” to read a song as it’s a compressed statistical representation, not in a human-readable format<a href="#_ftn3">[3]</a>; (2) the outputs are generated by the prompts of users, and not OpenAI; and (3) any reproduction was covered by the EU text and data mining (TDM) exception under the copyright law. The Munich I court rejected these defences and upheld GEMA’s claims for injunctive relief, information, and damages.<a href="#_ftn4">[4]</a></p><p><strong>Copyright Reproduction and Memorisation</strong></p><p>Under Article 2 of Directive 2001/29/EC (the InfoSoc Directive)<a href="#_ftn5">[5]</a> and Section 16 of the German Copyright Act (Urheberrechtsgesetz, UrhG)<a href="#_ftn6">[6]</a>, authors enjoy the exclusive right to authorise or prohibit the reproduction of their works. The Munich I Regional Court delved onto whether the incorporation of song lyrics into a large language model’s parameters could amount to “reproduction” within the meaning of EU and German copyright law. The Chamber rejected the argument that because large language models rely on statistical representation of patterns rather than storing files in a conventional sense. Instead, it focused on functionality and effect of such storage and stated that if protected work can be generated in recognisable form through interaction with the language model, then it must contain an embodiment of that work sufficient to meet the legal threshold of reproduction.<a href="#_ftn7">[7]</a> This interpretation expands the right of reproduction to include generative AI models within the traditional framework of copyright protection, rather than treating them as analytically distinct.<a href="#_ftn8">[8]</a></p><p><strong>Text and Data Mining Exception</strong></p><p>Section 44(b) of UrhG provides for an exception wherein reproductions made “in the course of text and data mining” for certain qualifying purposes, primarily research and analysis, are allowed. However, OpenAI’s reliance on the EU TDM exception was rejected by the court. It was stated that where a model’s parameters embody a copyrighted work in an extractable form, the utilitarian premise of the TDM exception does not apply as the copyrighted work’s exploitation rights are materially affected by such use.<a href="#_ftn9">[9]</a></p><p>The narrow interpretation of the TDM exception has significant implications on generative AI companies, including OpenAI as it now suggests that the mere presence of copyrighted material within training data cannot be automatically excused as legitimate data mining. Models that yield verbatim or near-verbatim outputs based on copyrighted training data may now fall outside the protective ambit of the TDM exception.</p><p><strong>Operator Responsibility and Output Liability</strong></p><p>OpenAI’s argument that users are responsible for any infringing outputs as the outputs are generated as a response to user prompts was also dismissed. The court emphasised that the OpenAI curates the training data, designs model architectures, and deploys models in configurations that shape the output, irrespective of the user prompt. The fact that users trigger, through various prompts, the infringing output, does not absolve the responsibility of the operator, especially where the operator’s decisions to use certain data sets to train the model materially contribute to the model’s capacity to reproduce copyrighted material.<a href="#_ftn10">[10]</a></p><p>Language model operators exercise control over the language model’s training date set, algorithmic design, and deployment parameters, thereby positioning them as actors who facilitate the reproduction of copyrightable material rather than acting as mere neutral intermediaries.<a href="#_ftn11">[11]</a> The ruling thus imposes direct accountability on AI developers for the outputs of their systems when those outputs embody infringing reproductions.</p><p><strong>Governance Implications for Generative AI companies</strong></p><p>The decision highlights a basic conflict in AI governance: established legal protections for artistic expression may clash with training data practices that prioritize performance enhancements through large and diverse datasets. The implications for generative AI companies, including OpenAI are numerous.</p><p>First, granular rights assessments must be supported by data governance frameworks. Aggregating or anonymizing training data without taking extractability into consideration is no longer adequate. It might be necessary for model developers to use technical strategies like filtering, post-training editing, or differential privacy techniques to reduce the amount of copyrighted content that is memorized.</p><p>Second, licensing solutions may be used by AI companies. Therefore, there is a possibility that structured markets for rights clearances will become essential to legitimate training programs of the language models. Collecting societies and authors may leverage to ensure remuneration for the use of protected works.</p><p>Third, transparency obligations, whether statutory or contractual, will be critical. While existing EU proposals such as the Digital Services Act and EU AI Act do not directly address copyright, emerging regulatory discussions emphasise documentation of training data sources, risk assessments, and governance safeguards. Compliance regimes that integrate legal risk assessments with technical documentation will be essential.</p><p>Finally, the decision invites reflection on global harmonisation challenges as divergent approaches to copyright limitations, fair use, and data mining exceptions complicate an already complex problem with companies like OpenAI operating across multiple jurisdictions. Generative AI companies may need to adopt jurisdiction-specific safeguards that account for varying legal thresholds of infringement and permissible use.</p><p><strong>Conclusion</strong></p><p>The decision in <em>GEMA v OpenAI</em> can reasonably be read as a deliberate effort by the Munich I Regional Court to answer at least some questions in the AI and copyright debate, including the non-applicability of the TDM exception and assigning operational liability to model operators for infringing output, irrespective of the user prompt. the Munich I court has articulated a framework that foregrounds rights holder protection without excusing technological abstraction.<a href="#_ftn12">[12]</a></p><p>For generative AI companies, the ruling highlights the importance of aligning their training data practices with the licensing requirements. The ruling also signals that European courts and potentially the Court of Justice of the European Union are willing to apply existing principles, without waiting for changes at the legislative level. At the same time, the ruling raises difficult questions about innovation. A strict application of reproduction rights as interpreted in this case may increase transaction costs and legal uncertainty, particularly for start-ups and research-driven entities.<a href="#_ftn13">[13]</a></p><p>The concern is especially acute for generative AI systems designed to create music, where the language model is primarily trained using copyrightable material like music and melodies. If courts interpret memorisation and similarity expansively, developers of music-generating models may face increased legal uncertainty, even in situation when outputs are not direct copies but derivative in aesthetic character. The long-term effect of this judgement could be a slower pace of innovation in creative AI sectors, unless regulatory or licensing frameworks balance such protection with a licensing mechanism.</p><p><a href="#_ftnref1">[1]</a> <em>GEMA v OpenAI</em> (Ref 42 O 14139/24) Landgericht München I (14 November 2024).</p><p><a href="#_ftnref2">[2]</a> Ibid.</p><p><a href="#_ftnref3">[3]</a> Giancarlo Frosio, ‘Copyright in Formaldehyde: How GEMA v OpenAI Freezes Doctrine and Chills AI — Part 1’ <em>Kluwer Copyright Blog</em> (10 December 2025) <a href="https://legalblogs.wolterskluwer.com/copyright-blog/copyright-in-formaldehyde-how-gema-v-openai-freezes-doctrine-and-chills-ai-part-1/?utm_source=chatgpt.com">https://legalblogs.wolterskluwer.com/copyright-blog/copyright-in-formaldehyde-how-gema-v-openai-freezes-doctrine-and-chills-ai-part-1/</a> accessed 27 February 2026.</p><p><a href="#_ftnref4">[4]</a> <em>GEMA v OpenAI</em> (Ref 42 O 14139/24) Landgericht München I (14 November 2024).</p><p><a href="#_ftnref5">[5]</a> Article 2, Directive 2001/29/EC of the European Parliament and of the Council of 22 May 2001 on the harmonisation of certain aspects of copyright and related rights in the information society (InfoSoc Directive).</p><p><a href="#_ftnref6">[6]</a> s 16, German Copyright Act (<em>Urheberrechtsgesetz</em>, UrhG).</p><p><a href="#_ftnref7">[7]</a> <em>GEMA v OpenAI</em> (Ref 42 O 14139/24) Landgericht München I (14 November 2024).</p><p><a href="#_ftnref8">[8]</a> Florian Mueller, ‘GEMA v OpenAI: Munich Copyright Trial Signals Liabilities for AI Training and Outputs’ (IPFray, 14 November 2024) <a href="https://ipfray.com/gema-openai%20accessed%2027%20February%202026">https://ipfray.com/gema-openai accessed 27 February 2026</a>.</p><p><a href="#_ftnref9">[9]</a> <em>GEMA v OpenAI</em> (Ref 42 O 14139/24) Landgericht München I (14 November 2024).</p><p><a href="#_ftnref10">[10]</a> <em>GEMA v OpenAI</em> (Ref 42 O 14139/24) Landgericht München I (14 November 2024).</p><p><a href="#_ftnref11">[11]</a> Ibid.</p><p><a href="#_ftnref12">[12]</a> Ibid.</p><p><a href="#_ftnref13">[13]</a> Ibid.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=75c40317c53f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Hallucination and its Myriad of Problems: A Review of KMG Wires Pvt Ltd v NFAC & Ors.]]></title>
            <link>https://medium.com/@saige_81207/ai-hallucination-and-its-myriad-of-problems-a-review-of-kmg-wires-pvt-ltd-v-nfac-ors-28a9f58ab975?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/28a9f58ab975</guid>
            <category><![CDATA[india]]></category>
            <category><![CDATA[supreme-court]]></category>
            <category><![CDATA[regulation]]></category>
            <category><![CDATA[law]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Sun, 22 Feb 2026 06:36:41 GMT</pubDate>
            <atom:updated>2026-02-22T06:36:41.732Z</atom:updated>
            <content:encoded><![CDATA[<p>By Rushika Shah, B.A LL.B ‘28</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_RBaXecxHzKZ9MNA" /></figure><p><em>AI or “the science of making machines do things that would require intelligence if done by men”</em><a href="#_ftn1"><em>[1]</em></a><em> has a problem, and the problem is that it hallucinates.</em></p><p>Artificial Intelligence (AI), particularly generative AI systems such as Large Language Models (LLMs), has increasingly been adopted for research, drafting, and analytical tasks across multiple professional domains, including law.<a href="#_ftn2">[2]</a> These systems are capable of producing fluent and persuasive text that often resembles expert human output. However, within these capabilities lies a serious and persistent problem: AI hallucination, a phenomenon where AI systems generate information that is factually incorrect, misleading, or entirely fabricated, but presenting such information with apparent confidence and as fact.</p><p>There was a case wherein a lawyer used generative AI for legal research and ended up submitting claims supported by imaginary judgments.<a href="#_ftn3">[3]</a> And another, where fabricated authorities were relied upon in formal pleadings, leading Courts to issue warnings and impose costs.<a href="#_ftn4">[4]</a> And another, and another — for these incidents are but two of the dozens of reported instances,<a href="#_ftn5">[5]</a> from which a clear picture emerges: AI, at its current stage of development, <strong>cannot</strong> be trusted with real facts and verified information without human oversight. Notably, and avoiding any sort of liability, Generative AI models such as ChatGPT and Gemini themselves warn users at the outset that they may make mistakes and should not be relied upon as authoritative sources.</p><p>The principal contributors to such hallucinations are the training data which LLMs utilize to gain their general capabilities and factual knowledge as well as alignment data, which teach LLMs to follow user instructions and align with human preferences.<a href="#_ftn6">[6]</a> In addition, their design focuses on generation of plausible content and not accuracy, resulting in the technology’s inability to discern what is truth.<a href="#_ftn7">[7]</a> This primarily manifests in the presence of misinformation and biases in the flawed pre-training data sources, the knowledge boundary inherently bounded by the scope of the pre-training data, and the hallucinations induced by inferior alignment data.<a href="#_ftn8">[8]</a></p><p>This phenomenon of AI hallucination is compounded by the “black box” problem, for LLMs are often black box AI. The black box problem is that neither users nor developers are fully aware or can explain the internal reasoning processes through which AI systems arrive at particular outputs.<a href="#_ftn9">[9]</a> Because of this opacity, AI hallucination presently remains both prevalent and unsolvable under existing technological paradigms. While researchers have proposed various mitigation strategies — such as fact checking methodologies such as external retrieval and internal checking,<a href="#_ftn10">[10]</a> no method has succeeded in eliminating hallucinations entirely. Knowing all of this, it would seem entirely pragmatic to avoid using AI-generated content in legal briefs, claims, or quasi-judicial decision-making, where factual accuracy and verified authority are foundational requirements. Wouldn’t it?</p><p>Yet, AI has already found its way into precisely such contexts.</p><p><em>KMG Wires Private Limited vs. The National Faceless Assessment Centre, Delhi &amp; Others</em><a href="#_ftn11"><strong><em>[11]</em></strong></a> is a recent landmark judgement in India decided on 6th October, 2025, and more importantly, a cautionary tale. Herein, an NFAC Assessing Officer issued a tax assessment order that substantially increased the petitioner’s declared income, and also included refusal of certain substantial purchases on the ground that the supplier failed to respond to a notice under Section 133(6) of the IT Act which was factually untrue, as the response was on record. However, concerningly and relevant for this brief, the officer also involved an alleged peak balance of Rs. 22.66 crore as part of unsecured loans from directors, including the opening balances in this computation under Section 68 of the Act, while citing three <strong>non-existent </strong>judicial precedents. These judgements — having been generated using AI — were not traceable in any recognised legal database.</p><p>A Division Bench in the Bombay High Court quashed this assessment order, along with the issuance of a consequential demand notice as well as penalty proceedings. The Court held that reliance on fabricated, AI-generated case law amounted to a serious breach of the principles of natural justice and reflected a complete non-application of mind by the assessing officer as well as breach of procedure.<a href="#_ftn12">[12]</a> Significantly, the Bombay High Court went beyond the facts of the individual case and issued a broader caution regarding the use of AI in administrative and quasi-judicial functions.<a href="#_ftn13">[13]</a> While acknowledging that AI tools may assist in research, the Court categorically warned against blind reliance on AI-generated outputs, emphasising that any such material must be independently verified by human decision-makers before being relied upon.<a href="#_ftn14">[14]</a> This judgment stands as the first explicit judicial acknowledgement in India of the dangers posed by AI hallucinations in governance and legal decision-making.</p><p>We can come to the conclusion that AI hallucinations are not simply isolated anomalies or minor errors but a structural limitation of contemporary generative AI. The persistence of hallucinations, despite extensive research and mitigation efforts, demonstrates how AI cannot presently function as an autonomous source of verified knowledge. The ruling in <em>KMG Wires Pvt Ltd v NFAC &amp; Ors</em> reinforces this reality, making it clear that while AI can be useful in legal research, it cannot replace judicial prudence and application of mind. It underscored the need for strict human verification and restraint in the use of AI in legal and administrative contexts. All this to say: please verify any and all information derived from generative AI and even more importantly, stay away from the em dash — lest you be mistaken for AI. Really bad humour aside, we hope you enjoyed your update on AI hallucination!</p><p><strong>Other sources:</strong></p><p>1. Matt O’Brien, ‘Chatbots Sometimes Make Things Up. Is AI’s Hallucination Problem Fixable?’ <em>Associated Press News</em> (1 August 2023) &lt;<a href="https://apnews.com/article/artificial-intelligence-hallucination-chatbots-chatgpt-falsehoods-ac4672c5b06e6f91050aa46ee731bcf4">https://apnews.com/article/artificial-intelligence-hallucination-chatbots-chatgpt-falsehoods-ac4672c5b06e6f91050aa46ee731bcf4</a>&gt; accessed 11 January 2026.</p><p>2. Leonardo Nicoletti and Dina Bass, ‘Humans Are Biased. Generative AI Is Even Worse: Stable Diffusion’s text-to-image model amplifies stereotypes about race and gender — here’s why that matters’<em> Bloomberg </em>(2023) &lt;<a href="https://www.bloomberg.com/graphics/2023-generative-ai-bias/">https://www.bloomberg.com/graphics/2023-generative-ai-bias/</a>&gt; accessed 31 January 2026.</p><p>3. Jiarui Li, Ye Yuan and Zehua Zhang, ‘Enhancing LLM Factual Accuracy with RAG to Counter Hallucinations: A Case Study on Domain-Specific Queries in Private Knowledge-Bases’ (2024) <em>arXiv preprint</em> arXiv:2403.10446 &lt;<a href="https://arxiv.org/abs/2403.10446">https://arxiv.org/abs/2403.10446</a>&gt; accessed 21 February 2026.</p><p><a href="#_ftnref1">[1]</a> Liran Antebi, ‘What Is Artificial Intelligence?’ in <em>Artificial Intelligence and National Security in Israel</em> (Institute for National Security Studies 2021) 31–40 <a href="http://www.jstor.org/stable/resrep30590.7">http://www.jstor.org/stable/resrep30590.7</a> accessed 31 January 2026.</p><p><a href="#_ftnref2">[2]</a> ‘Attorneys Sanctioned for AI Hallucinations in Court Filings’ <em>Washington Post</em> (3 June 2025) &lt;<a href="https://www.washingtonpost.com/nation/2025/06/03/attorneys-court-ai-hallucinations-judges/">https://www.washingtonpost.com/nation/2025/06/03/attorneys-court-ai-hallucinations-judges/</a>&gt; accessed 31 January 2026.</p><p><a href="#_ftnref3">[3]</a> <em>Mata v Avianca Inc</em> 2023 WL 4114965 (SDNY).</p><p><a href="#_ftnref4">[4]</a> ‘Abu Dhabi Court Slaps ₹63 Lakh Costs on Law Firm for Citing Non-Existent AI-Generated Authorities’ <em>Bar &amp; Bench</em> &lt;<a href="https://www-barandbench-com.ap2.proxy.openathens.net/news/litigation/abu-dhabi-court-slaps-63-lakh-costs-on-law-firm-for-citing-non-existent-ai-generated-authorities">https://www-barandbench-com.ap2.proxy.openathens.net/news/litigation/abu-dhabi-court-slaps-63-lakh-costs-on-law-firm-for-citing-non-existent-ai-generated-authorities</a>&gt; accessed 31 January 2026.</p><p><a href="#_ftnref5">[5]</a> ‘Attorneys Sanctioned for AI Hallucinations in Court Filings’ <em>Washington Post</em> (3 June 2025) &lt;<a href="https://www.washingtonpost.com/nation/2025/06/03/attorneys-court-ai-hallucinations-judges/">https://www.washingtonpost.com/nation/2025/06/03/attorneys-court-ai-hallucinations-judges/</a>&gt; accessed 31 January 2026.</p><p><a href="#_ftnref6">[6]</a> L Huang and others, ‘A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions’ (2025) 43(2) <em>ACM Transactions on Information Systems</em> Article 42, 1–55 &lt;<a href="https://doi.org/10.1145/3703155">https://doi.org/10.1145/3703155</a>&gt; accessed 11 February 2026.</p><p><a href="#_ftnref7">[7]</a> MIT Sloan Teaching &amp; Learning Technologies, ‘Addressing AI Hallucinations and Bias’ &lt;<a href="https://mitsloanedtech.mit.edu/ai/basics/addressing-ai-hallucinations-and-bias/">https://mitsloanedtech.mit.edu/ai/basics/addressing-ai-hallucinations-and-bias/</a>&gt; accessed 31 January 2026.</p><p><a href="#_ftnref8">[8]</a> Ibid.</p><p><a href="#_ftnref9">[9]</a> Rahul Awati, ‘What Are AI Hallucinations and Why Are They a Problem?’ <em>WhatIs.com (TechTarget)</em>, 2024 &lt;<a href="https://www.techtarget.com/whatis/definition/AI-hallucination">https://www.techtarget.com/whatis/definition/AI-hallucination</a>&gt; accessed 31 January 2026.</p><p><a href="#_ftnref10">[10]</a> L Huang and others, ‘A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions’ (2025) 43(2) <em>ACM Transactions on Information Systems</em> Article 42, 1–55 &lt;<a href="https://doi.org/10.1145/3703155">https://doi.org/10.1145/3703155</a>&gt; accessed 11 February 2026.</p><p><a href="#_ftnref11">[11]</a> <em>KMG Wires Pvt Ltd v National Faceless Assessment Centre &amp; Ors</em>, Writ Petition (L) No 24366 of 2025 (Bom HC).</p><p><a href="#_ftnref12">[12]</a> Ibid.</p><p><a href="#_ftnref13">[13]</a> Ibid.</p><p><a href="#_ftnref14">[14]</a> Ibid.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=28a9f58ab975" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[FROM BINARY TO BIAS: IS AI REINFORCING LINGUISTIC HIERARCHY?]]></title>
            <link>https://medium.com/@saige_81207/from-binary-to-bias-is-ai-reinforcing-linguistic-hierarchy-f0665ae739fd?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0665ae739fd</guid>
            <category><![CDATA[law]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Sun, 08 Feb 2026 05:40:38 GMT</pubDate>
            <atom:updated>2026-02-08T05:43:09.125Z</atom:updated>
            <content:encoded><![CDATA[<p>By Drishti Sankhyan, B.A LLB ‘2028</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HvW0-Su_AVyNATBYPsvoMQ.jpeg" /></figure><p>Only 15% of ChatGpt users are from United States yet English (SAE) primarily dominates the training dataset for ChatGpt models. As a result, speakers of other languages, including non-standard variants of English, frequently experience prejudice due to poor response quality and feel misrepresented.</p><p>Language technologies, such as natural learning processing, are typically offered as neutral tools, implying inclusion and accessibility. The notion that certain language varieties (typically associated with power) are superior and more accurate is known as standard language ideology. However, it is contended that all language variants are equally expressive and accurate.</p><p>In today’s world, English, especially Standard American English (SAE), holds a privileged position due to historical events like colonialism and globalization. Accordingly, large language models inherit and reinforce the discrimination and beliefs set by humans based in race, caste and nationality. They are trained on large datasets extracted from the internet, where English makes up about 60% of the content, despite being spoken by only 17% of the global population.<a href="#_ftn1">[1]</a> As a result, the large language models (LLMs) understand SAE as the default language standard. Consequently, non-standard varieties are often knowingly excluded or underrepresented, creating a hierarchy where deviating from the standard leads to poorer performance in quality of results from AI.</p><p>In 2024, the University of California at Berkeley issued a report that used statistics to show that AI Models demonstrate a standardised bias against non-standard varieties of English.<a href="#_ftn2">[2]</a> The researchers studied how ChatGPT responded to dialects, like Indian English, African-American English and others. Approximately 25% of the content generated by ChatGPT included negative remarks toward the users of non-standard dialects.<a href="#_ftn3">[3]</a></p><p>Further, ChatGPT failed to understand many phrases used in these dialects. Several people have come forward to address the same issues. For example, the Derby City Council used an AI tool to help their employees, but when the employees tried to communicate with the tool using the Derbyshire dialect, the AI Assistant could not understand what they were trying to say because it did not recognize the difference between their local words.<a href="#_ftn4">[4]</a> In yet another instance, one of a job applicant in India when he submitted his job application through ChatGPT, automatically replaced the surname of the applicant with an another surname commonly associated with a higher caste.<a href="#_ftn5">[5]</a></p><p>This demonstrates that there can be linguistic biases in AI results from more than just poor performance and can go on to affect every aspect of people’s social and economic lives. Within the education sector, the use of AI tools that function reliably only in the English language has resulted in extra layers of inequality in many areas as number of studies have shown that worldwide, students are increasingly switching to English when interacting with AI tools, regardless of whether they have completed their assignments in their mother tongue. Teenagers in Seoul and Berlin, for example, are reporting better outcomes when prompting AI in English which then reinforces the implicit connection between being smart and speaking fluent English.<a href="#_ftn6">[6]</a></p><p>However, if nothing changes, there is a high likelihood that the next generation will internalise the linguistic hierarchy associated with English through their everyday interactions with technology despite government policies. The situation in Iceland serves as a cautionary tale. While the government provides extensive support to preserve Icelandic as a living language, young Icelanders are increasingly reported to have been utilising English when performing more complex tasks, due to the shortcomings of AI.<a href="#_ftn7">[7]</a> What started out as a good deed of cultural globalisation has now become a technological requirement to continue to access and thrive in the digital world.</p><p>The effects of these trends culminate in a form of digital linguistic imperialism. There are over 7000 languages in the world; however there are only approximately 100 substantial AI resources.<a href="#_ftn8">[8]</a> If AI systems are becoming critical components of Governance, Education and other sectors and citizens do not have access to appropriate language support, this can translate into more trouble as well as risk for a citizen’s civic safety.</p><p>Moreover, similar patterns are seen in the case of Sign Languages. A report by European Union of the Deaf has noted how AI systems misinterpret grammar and cultural nuances in Sign Languages, because sign language communities have not been much involved in developing the tools used to create AI tool for Sign Language.<a href="#_ftn9">[9]</a> This reinforces the belief that access to Sign Language is an afterthought rather than a necessity.</p><p>Many developers are advocating for multilingual models and use of dialect translations. However, while these options may be beneficial, it is seen that translation often repeats errors and removes cultural context. In addition, larger models do not automatically solve the problem of data scarcity when developing machine learning systems in a different language. An alternative approach to developing models that have multilingual capacity could be through a community led data collection initiative, such as the Maori Language initiatives being used in New Zealand.<a href="#_ftn10">[10]</a> Community led data collection allows for community members to preserve their linguistic sovereignty and to ensure that the collected data is accurate.</p><p>It is interesting when we realise that the corporations who decide what constitutes as an appropriate way for a model to behave are usually small in number, and many times as depicted above are completely disconnected from the communities who speak the language in question. Therefore, AI systems will continue to replicate the existing linguistic hierarchy if not for any type of participation from the public to reduce the data scarcity and thus, it is time to begin with the first step and help in changing the way AI sees and learns.</p><p><strong>Citations</strong></p><p><a href="#_ftnref1">[1]</a> Genevieve Smith and others, ‘Standard Language Ideology in AI-Generated Language’ (2018) 37(4) ACM/IMS J. Data Sci. 111 &lt; <a href="https://arxiv.org/abs/2406.08726">https://arxiv.org/abs/2406.08726</a> &gt; accessed 15 January 2026.</p><p><a href="#_ftnref2">[2]</a> Fintan Burke, ‘AI chatbots are ‘alarmingly’ biased against dialect speakers’ (<em>Deutsche Welle</em>, 29 December 2025) &lt; <a href="https://www.dw.com/en/ai-chatbots-are-alarmingly-biased-against-dialect-speakers/a-75247017">https://www.dw.com/en/ai-chatbots-are-alarmingly-biased-against-dialect-speakers/a-75247017</a> &gt; accessed 14 January 2026.</p><p><a href="#_ftnref3">[3]</a> Ibid.</p><p><a href="#_ftnref4">[4]</a> Burke (n 2).</p><p><a href="#_ftnref5">[5]</a> Nilesh Christopher, ‘OpenAI is huge in India. Its models are steeped in caste bias” (<em>MIT Technology Review</em>, 1 October 2025) &lt;<a href="https://www.technologyreview.com/2025/10/01/1124621/openai-india-caste-bias/">https://www.technologyreview.com/2025/10/01/1124621/openai-india-caste-bias/</a> &gt; accessed 14 January 2026.</p><p><a href="#_ftnref6">[6]</a> Tom Cranstoun, ‘The Digital Language Caste System — How AI Creates Linguistic Apartheid’ (<em>All About Network</em>) &lt; <a href="https://allabout.network/blogs/ddt/ai/the-digital-language-caste-system">https://allabout.network/blogs/ddt/ai/the-digital-language-caste-system</a> &gt; accessed 13 January 2026.</p><p><a href="#_ftnref7">[7]</a><strong> </strong>Miranda Bryant, ‘Icelandic is in danger of dying out because of AI and English-language media, says former PM’ (<em>The Guardian</em>, 15 November 2025) &lt;<a href="https://www.theguardian.com/world/2025/nov/15/icelandic-is-in-danger-of-dying-out-because-of-ai-and-english-language-media-says-former-pm">https://www.theguardian.com/world/2025/nov/15/icelandic-is-in-danger-of-dying-out-because-of-ai-and-english-language-media-says-former-pm</a>&gt; accessed 15 January 2026.</p><p><a href="#_ftnref8">[8]</a> Cranstoun (n 6).</p><p><a href="#_ftnref9">[9]</a> Sofia Olofsson, ‘When AI Can’t Understand Your Language, Democracy Breaks Down’ (Tech Policy, 11 December2025) &lt;<a href="https://www.techpolicy.press/when-ai-cant-understand-your-language-democracy-breaks-down-/">https://www.techpolicy.press/when-ai-cant-understand-your-language-democracy-breaks-down-/</a> &gt; accessed 14 January 2026.</p><p><a href="#_ftnref10">[10]</a> Smith ( n 1)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0665ae739fd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI as a Defence: The Military Application]]></title>
            <link>https://medium.com/@saige_81207/ai-as-a-defence-the-military-application-5617a4edd818?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/5617a4edd818</guid>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Tue, 18 Nov 2025 14:06:49 GMT</pubDate>
            <atom:updated>2025-11-18T14:06:51.381Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/747/1*9JVWYLu9YjrNgeSFpXELCw.jpeg" /></figure><p>When you think of AI, you may conjure up an image of the ChatGPT logo. If asked to think more, you may as well visualise the Perplexity image floating beside it. But, in normal usage, would you ever think of tanks and ballistic missiles? Up until now, Artificial Intelligence has been associated with some of its digital applications and implications. What about the on-site effects of these algorithms, though?</p><p>Palantir could be one of the many possible answers.</p><p>Founded in 2003, Palantir has had a novel approach to the application of AI in warcraft. From ground stations to assess the location of assets, to aerial layers scoping the radar for better targeting(as proven in Project Maven), the organisation encompasses an arsenal of varied algorithms, improving the army’s tactical response to immediate and planned dangers. Moreover, in order to maintain a balanced duty to individuals and the government, it aims at preserving civil liberties while upholding planning missions to stop terrorist attacks and neutralise hundreds of adversaries. And this is just scratching the surface of this domain’s genius.</p><p>Using AI in better mapping of the battlefield, does sound like a feat. But what if we take it up a notch? What if we could implement this intelligence in vehicles? And ultimately, even replace boot soldiers? DARPA, DRDO, ISRO, IDF and the defence forces of Russia and China are few of the defence organisations throughout the globe, progressing towards this objective.</p><p>The Joint Artificial Intelligence Centre (JAIC), now the Chief Digital and Artificial Intelligence Office (CDAO), is a United States Agency responsible for the integration of Artificial Intelligence into the battlefield. It entered into an $800 million contract in 2020 to expand its operations to warcraft with Booz Allen Hamilton, a reputed AI company. Similarly, Thales Group and Raytheon Technologies are other players indulging in this sphere, in cohorts with the defence forces.</p><p>These cases of major defence groups from nuclear countries, entering into the modern era’s approach to warfare with corporate players, raise several vital, yet difficult questions. This modern integration is said to serve as a deterrence, but at what cost? More importantly, at whose cost?</p><p>Similarly, in case of a defect in the service, how does the law determine who is to be held liable? Is it the code generator, the company, the organisation applying the services or those deploying the assets? So then, would the possibility of a private military company dabbling in this area be potential sign of concern? Concerns surrounding blurred accountability, false distinction between a civilian and a combatant, and privacy infringement, steer us towards the legal implications of such questions.</p><p>There are two applications that are kept in mind, when the ICRC (International Committee of the Red Cross) looks at AI: Use in the conduct of warfare and use in humanitarian action. When discussing the aspects of armed conflict, autonomous weapon systems pose a risk of unpredictable attacks based on fixed content fed into the system — which may change quite erratically — leading to harms caused to citizens or even relations between states. Now, even if this harm is to be addressed, to whom shall it be? The question of accountability clashes with the crucial condition of <em>jus in bello</em>, wherein a person has to be held liable. Over here, we could call the law to be rigid or the technology, too complex.</p><p>This issue is so complex that it was raised in the form of an open letter in 2015, by the ‘future of life Institute’, calling for a ban on autonomous weapons if it operated beyond human control. It was signed by an influential list of tech moguls like Elon Musk, Steve Wozniak and scientists like Stephen Hawking and Noam Chomsky, as well as other AI and robotics researchers.</p><p>Another initiative was a report presented to the UN Human Rights Council, in 2013 (A/HRC/23/47). It recommended member states to declare and implement moratoria on the “testing, production, transfer, and deployment” of lethal autonomous robotics (LARs) until an internationally agreed upon framework for LARs has been established.</p><p>Taking a broader perspective on the involvement of corporate bodies in military affairs, it has been observed that while private intervention will provide expertise and rapid innovation, it will also lead to an arms race, dismantling the core of international relations. It is important to extend the regulations of the IHL treaties and related conventions to these private players to improve compliance and accountability.</p><p>The black box nature of AI raises significant questions on accountability too. Why it works the way it works, is almost impossible to trace. Algorithms have no legal personality, no basis for intent and its connection to causation. To tackle this, experts suggest that AI systems should be put up for a periodic review with the algorithms displaying a more transparent process towards action. This, coupled with rigorous testing, could act as safeguards to shape the doctrines and culture influencing its usage.</p><p>Thus, we can grasp from the given contentions and possible solutions that artificial intelligence, while bearing inherent issues that may affect citizens and societies on a large scale, has the capabilities in bringing about a transformation to the world order, and the potential to limit casualties using latest technologies. At the end of the day it is important to remember that a weapon should not be held to attack, but to protect!</p><p>Written by Aishwarya Venkat</p><p>&lt;<a href="https://www.linkedin.com/in/aishwarya-venkatasubramanian/">https://www.linkedin.com/in/aishwarya-venkatasubramanian/</a>&gt;</p><p><strong>Works Cited</strong></p><p>1. International Committee of the Red Cross (ICRC), <em>Artificial intelligence and machine learning in armed conflict: A human-centred approach</em> (2021) 102 Int Rev Red Cross 463 <a href="https://international-review.icrc.org/sites/default/files/reviews-pdf/2021-03/ai-and-machine-learning-in-armed-conflict-a-human-centred-approach-913.pdf?utm_source=chatgpt.com">https://international-review.icrc.org/sites/default/files/reviews-pdf/2021-03/ai-and-machine-learning-in-armed-conflict-a-human-centred-approach-913.pdf</a></p><p>2. International Committee of the Red Cross, <em>Artificial intelligence and machine learning in armed conflict: A human-centred approach</em> (2021) 102 Int Rev Red Cross 463 <a href="https://international-review.icrc.org/sites/default/files/reviews-pdf/2021-03/ai-and-machine-learning-in-armed-conflict-a-human-centred-approach-913.pdf?utm_source=chatgpt.com">https://international-review.icrc.org/sites/default/files/reviews-pdf/2021-03/ai-and-machine-learning-in-armed-conflict-a-human-centred-approach-913.pdf</a></p><p>3. Etzioni A and Etzioni O, ‘Pros and Cons of Autonomous Weapons Systems’ (2017) 97 <em>Military Review</em> (May-June) 72–80 <a href="https://www.armyupress.army.mil/Journals/Military-Review/English-Edition-Archives/May-June-2017/Pros-and-Cons-of-Autonomous-Weapons-Systems/?utm_source=chatgpt.com">https://www.armyupress.army.mil/Journals/Military-Review/English-Edition-Archives/May-June-2017/Pros-and-Cons-of-Autonomous-Weapons-Systems/</a></p><p>4. Davit Khachatryan, ‘Military AI Challenges Human Accountability’ (Center for International Policy, 21 July 2025) <a href="https://internationalpolicy.org/publications/military-ai-challenges-human-accountability/?utm_source=chatgpt.com">https://internationalpolicy.org/publications/military-ai-challenges-human-accountability/</a></p><p>5. Purvi Pokhariyal, Amit K. Kashyap &amp; Arun B. Prasad (eds), <em>Artificial Intelligence: Law and Policy Implications</em> (Institute of Law, Nirma University 2020).</p><p>6. Ryder RD and Naren N, <em>Artificial Intelligence and Law: Challenges Demystified</em> (Law &amp; Justice Publishing Co 2022) ISBN 978–93–93850–09–6.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5617a4edd818" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unseen Algorithms: How AI Quietly Shapes Our Daily Lives]]></title>
            <link>https://medium.com/@saige_81207/unseen-algorithms-how-ai-quietly-shapes-our-daily-lives-abe7605e63a9?source=rss-6f4908bb22ed------2</link>
            <guid isPermaLink="false">https://medium.com/p/abe7605e63a9</guid>
            <category><![CDATA[law]]></category>
            <category><![CDATA[regulation]]></category>
            <category><![CDATA[legal]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Society on AI Governance and Ethics]]></dc:creator>
            <pubDate>Mon, 03 Nov 2025 16:21:27 GMT</pubDate>
            <atom:updated>2025-11-03T16:36:17.720Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EjiHEwGBK_14yxZzliqH0A.jpeg" /></figure><p>Presently, when we hear the phrase <em>Artificial Intelligence</em>, we are prompted to think in terms of self-driving cars, cooking robots, or simply, increasingly ubiquitous autonomous machines. But in reality, AI may not be so overt, rather subliminal — quietly forming our opinions, influencing our choices, behaviour, and actions. This article explores how AI influences our everyday lives, the mechanism behind such phenomena, and the ethical and social implications that follow.</p><p>Subliminal influences are those that occur below the threshold of the conscious mind, where our moods, decisions, habits, and opinions are actively shaped by their effects. In an AI-driven world, algorithms work quietly in the background, shaping what we see, hear, like, or dislike, often without our awareness. A recent review highlighted that only a small fraction of our mental processing is conscious, while the vast majority operates subconsciously. AI now has access to this domain through the creation of <em>decision architectures — </em>systems designed to subtly guide us toward certain choices without our awareness.¹ Furthermore, the proposed EU AI Act recognizes <em>subliminal techniques</em> that function beyond or beneath the level of conscious perception and have the capacity to influence a person’s habits and behaviour in ways the individual remains unaware of.²</p><p>In the case of content recommendations and social media feeds, users on Instagram, YouTube, or any other app, are inclined to believe that the videos and ads they come across are often under their control. What we fail to notice is that every microsecond we spend on these platforms is being tracked to understand our preferences. The moments we pause, or skip are recorded by AI algorithms, which then use this data to suggest similar content. In doing so, these algorithms aim to keep us engaged by continuously offering the kind of material we are most likely to interact with. Over time, this not only shapes our opinions, moods, and decisions but can also subtly influence our sense of identity and belonging.</p><p>Further, Online shopping has become one of the most effective platforms for personalized suggestions — not only in recommending products but also in tailoring discounts, pricing, and limited-time offers according to an individual’s financial conditions. Reports have shown that even ride-hailing apps like Uber and Ola display price variations based on subtle factors such as the user’s battery level or the type of phone they use. For instance, cab fares for identical routes in Chennai, checked simultaneously on iPhone and Android devices, revealed consistently higher prices for iOS users!³</p><p>Research has consistently shown that humans continuously generate data through their online activities, through browsing, purchasing, and clicking on ads. This data is then used to train AI systems, which in turn produce suggestions that influence and shape human behavior. Over time, this creates a feedback loop where human actions train AI, and AI simultaneously guides human decisions, leading to a continuous process of human–AI co-evolution.⁴</p><p>These loops show that AI’s influence is not external but deeply embedded in our daily lives and experiences, making it appear normal and less noticeable. We may not feel it directly, yet our moods, decisions, habits, and opinions are subtly shaped by its effects. This subliminal effect can be argued from both perspectives: benefits and risks. On one hand, it can be said that AI-driven personalization makes our daily routines smoother and more efficient by suggesting what we might like, whether it be music, videos, or other content.</p><p>However, the more we rely on these algorithmic suggestions, the more we risk surrendering our autonomy, at times feeling a reduced sense of agency in our decision-making. One of the most concerning risks is the potential manipulation and addiction of vulnerable groups such as children and emotionally sensitive individuals. The fear of missing out and the constant pursuit of social belonging have become increasingly common consequences of this dependency.</p><p>For example, a study found that users on social media platforms often failed to realise how layered AI nudges had gradually altered their patterns of content consumption, with emotional triggers such as FOMO being deliberately exploited.⁵ During the recommendation process, AI systems are also influenced by historical biases related to gender, race, and class, often amplifying these biases rather than eliminating them. This occurs because the data used to train AI originates from a society that is itself inherently biased.</p><p>Suggested solutions in tackling issues as such, include policies that should be formulated by the state to prohibit the use of subliminal techniques or subject them to strict audits. Algorithmic systems must be transparent and should not be permitted to exploit vulnerable individuals, or just users in general. AI systems should provide an option to opt out of recommendation mechanisms, along with interface features that allow users to reset or customise their profiles and browsing histories. Developers should be required to mitigate historical biases and ensure that data collected through algorithmic processes is not used unethically.</p><p>The era we live in today is undeniably the age of AI, and its influence is not always apparent. It quietly shapes what we see, what we like, how we behave, and even what we believe in. While this influence may appear convenient and helpful, it raises profound ethical concerns surrounding autonomy, fairness, and the broader ideological shifts occurring within society. The challenge, therefore, is not only technological but also deeply human. We must reclaim control over our choices by making conscious, cautious decisions while engaging with technology without overlooking the responsibility of technologists and developers to act ethically in designing these systems.</p><p>Written by Ankit Yadav.</p><p><strong>Work Cited</strong></p><p>¹ “Can AI push the boundaries of privacy and reach the subconscious mind?” (Universitat Oberta de Catalunya, 24 October 2023) &lt;<a href="https://www.uoc.edu/en/news/2023/285-AI-and-the-subconscious-mind">https://www.uoc.edu/en/news/2023/285-AI-and-the-subconscious-mind</a>&gt;</p><p>² European Commission, ‘Commission Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act), C(2025) 884 final, 4 February 2025 (Annex)’ &lt;<a href="https://www.legalianet.com/wp-content/uploads/2025/02/prohibited-artificial-intelligence-practices.pdf">https://www.legalianet.com/wp-content/uploads/2025/02/prohibited-artificial-intelligence-practices.pdf</a>&gt;</p><p>³ Taboola Taboola, ‘Ola, Uber fares different on Apple and Android phones? The conspiracy of pricing war’ The Economic Times (7 April 2023) &lt;<a href="https://economictimes.indiatimes.com/news/company/corporate-trends/ola-uber-fares-different-on-apple-and-android-phones-the-conspiracy-of-pricing-war/articleshow/116671509.cms?from=mdr">https://economictimes.indiatimes.com/news/company/corporate-trends/ola-uber-fares-different-on-apple-and-android-phones-the-conspiracy-of-pricing-war/articleshow/116671509.cms?from=mdr</a>&gt;</p><p>⁴ Dino Pedreschi and others, ‘Human-AI Coevolution’ (arXiv, 3 May 2024) arXiv:2306.13723 &lt;<a href="https://arxiv.org/abs/2306.13723">https://arxiv.org/abs/2306.13723</a>&gt;</p><p>⁵ A Qualitative Study of How AI Influences Subconscious Decision-Making in Digital Nudges Across Social Media Platforms’ (2025) International Journal of Research and Innovation in Social Science</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=abe7605e63a9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>