Author – Anurag Kumar
College – Lloyd Law College
ABSTRACT
AI is no longer confined to the lab – it is in the courthouse, on Capitol Hill and in boardrooms at an speed that often outpaces the law attempting to rein it in. Generative AI tools can write, draft, decide and even diagnose and, with that, comes a flood of thorny, unanswered questions: who wrote that work? Who is liable? What’s the privacy issue here?
And who is actually on the hook for that big mistake? To understand that intersection of the law and the machines, we consider AI from the perspective of three things: the global regulatory efforts taking shape (especially the EU AI Act); the copyright and tort law battles already underway; and the cases currently making new law on the fly. S. Copyright Office’s position on copyrighting material generated by AI-we’ll discuss the premise of the article in a piece here in the Washington Monthly and tell you that, it turns out, the law isn’t so much slow-walking AI so much as it’s engaging in the time – honoured common law tradition of analogy to work toward a conclusion it has yet to reach.
Everything’s on the line and the last 18 months in case law is going to shape the legal framework of AI for a generation to come.
To The Point
AI has no legal personhood. Courts and copyright offices all over the globe have stated numerous times that only humans may be “authors” and “inventors” – machines aren’t owners of IP. Training AI on existing works has a particularly cloudy and debatable grey area: some judges and juries have deemed such usage to be fair and transformative; others have ruled that the way data was sourced to train AI(licensed or pirated) dictates the liability even if the AI training itself might not be, depending on jurisdiction. This is forcing legal practitioners to take on AI with some well worn frameworks – like product liability, negligence, and vicarious liability – for want of new legislative intent to address AI in an appropriate way.
The European Union AI Act – though it will serve as the first in the world’s comprehensive, risk tiered legal regulation ofAI–is still gradually rolling out until 2027-2028 and its most stringent rules are being postponed for a second time. 5 billion between the AI company Anthropic and writers (the most reported US copyright recovery in history) is currently serving as a legal blueprint. This litigation is being settled out of court before any courtroom decision is rendered via either settlements or licensing arrangements.
Use of Legal Jargon
Anyone who follows AI litigation reporting is sure to notice a certain language used by people discussing the cases right away. Here’s a quick working glossary to explain the terms:
Fair use U.S. Copyright law that allows limited reuse of copyrighted works without obtaining permission because, the use “transforms” it (is used for a new, different purpose and not a mere substitute). What counts as transformative use when training generative AI on copyrighted text or music is the central question, so far without a definitive judicial answer.
Human authorship requirement The centuries-old standard, reiterated by U.S. Copyright law again and again, that any work eligible for protection by copyright has to have been the work of a human being, leaving purely AI-generated outputs registrable with no human creative involvement.
General-Purpose AI (GPAI) model A regulatory definition, particularly from the E.U. AI Act, for an AI system “that is capable of performing a wide range of general tasks rather than a specialized, purpose-built one.” GPAI providers have specific transparency and documentation requirements that are distinct from those governing the narrowly designed AI.
High-risk AI system Under the E.U. AI Act, it is a high-risk “product” if it is “intended to be used in a context that adversely impacts its health, safety and fundamental rights,” for example, “biometric identification applications used by law enforcement” or in “employment and worker management or self-employment.”These high-risk systems face demanding requirements related to assessment of conformity, provision of technical documentation, and monitoring of human supervision.
Class certification A procedural determination that enables a class action lawsuit to proceed on behalf of a group of similarly situated persons – as in, for example, “authors whose novels have been incorporated into the training data set of” a certain large-language model – instead of requiring every individual author to file her or his own suit.
Discovery The pre-trial procedure through which parties to a lawsuit are required to produce information to each other. In some of the most prominent AI lawsuits, discovery – the process where the legal system compels one side to hand over its information – has taken on critical importance; courts have already ordered AI companies to share tens of millions of user logs, and most legal observers believe that a case will be settled based on evidence found in discovery, rather than from argument in court.
Vicarious and contributory liability The legal concept under which one party can be held liable for the wrongdoing of another party – as when a website or platform facilitates, or profits from, copyright infringement by users, or for copyright infringement accomplished through its product or services, without having itself directly infringed.
Personhood / legal capacity Whether AI systems could potentially someday become subjects of rights and responsibilities themselves rather than solely objects of human or corporate agency, which currently no country recognizes.
The Proof
Evidence that the laws surrounding AI, rather than some kind of fixed rulebook for AI law, are being forged in real-time and on the spot is apparent when we watch two activities unfolding at the same time – the lawmakers producing regulations that govern a technological landscape constantly being redefined and the courts struggling to slot that technology into old frameworks, many created several decades ago.
The EU’s AI Act came into force in August 2024 and the world’s first legislation on general and risk-based AI regulates how risk is associated with any AI system – rather than how it’s made. Prohibited AI uses and duties of knowledge regarding AI have been enforced since February 2025, and since August 2025 general purpose AI models have come under rules and regulations regarding the duty, and by August 2026 many other regulations have yet to be implemented. This flagship policy hasn’t gone unnoticed, and it needs updates. A recent “Digital Omnibus” law approved in May 2026 pushes back the stringent high-risk AI obligations until after 2027.
It indicates that neither standards, conformity, nor authorities can meet the demands and schedule. The new law also introduces new illegal acts that would ban the creation of or manipulation of Non-Consensual Intimate Imagery (NCII) and Child sexual abuse material and was effective December 2026. The United States, in contrast, is without a similar Federal law and is dependent on litigation and enforcement actions by agencies such as the Federal Trade Commission as well as court decisions for an operating AI legal system, making the legal framework scattered but perhaps quick because the court is able to adjudicate immediate conflicts without having to wait for potentially never enacted legislation.
Doctrinal Course of Action. The dilemma at the heart of almost every AI lawsuit – that the use of copyrighted material in training AI systems is an infringement or transformative fair use – might seem to be the most difficult question; however, the most straightforward answer is the most accurate for this time, as for the mid-2026 landscape, it all depends on the circumstances of the suit and how the information was obtained. Generally, when the courts have struggled with two commonly mixed concepts – whether obtaining training data illegally constitutes a violation regardless of how it is used and if training is transformative fair use – a divide between the two has started to materialize. Every subsequent case is guided by this fundamental concept, and it highlights why some AI firms are subject to charges because of the methods of data collection used, even in instances where they received more sympathetic rulings regarding AI training overall.
Economic Course of Action. The number of cases resolved prior to trial continues to grow, and settlement figures are now serving as reference prices for entire industries. The highest known AI copyright settlement to date involved Anthropic and covered approximately 482,000Works at an average price of few thousand dollars per Work, receiving interim court approval in September 2025.Industry attorneys have viewed this deal as the standard for the cost of AI training by unauthorized parties using books. Music Labels likewise went similar paths-in November 2025, Warner Music settled with AI music company Suno and licensed a deal, whereas Universal Music Group-settled with Ud io in October 2025 and announced a collaboration to create a licensed AI music platform . Sony Music, however, has settled neither case and instead seeks to have either one decided, anticipating a favourable fair-use decision might be worth more.
Case Law
jA handful of cases have become the boilerplate that all AI lawyers know by heart. And none of them are fully resolved-a key fact in itself. S. Copyright for a work that he said was created solely by his AI, naming only the AI as the author.
The Copyright Office rejected his application, a decision upheld by courts that insist that works be the creation of humans to be copyrightable. Thaler asked the Supreme Court to take up the case-claiming that the decisions had stifled human innovation with AI. The court declined in March 2026, reaffirming the idea that works have to be authored by human beings. Legal experts have noted that this particular ruling might be narrowly applicable because Thaler didn’t claim any human input; we don’t know how much AI assist is too much.
Thomson Reuters v. ” The court ruled that the headnotes were copyrighted works of authorship, and the use was not fair use. This ruling is being appealed in the Third Circuit, and it’s one of only a few cases that have ruled against fair use for AI training out of hand. Bartz v.
Anthropic A class action was filed by a number of authors accusing the company Anthropic of using pirated copies of their books to train its models. A federal judge ruled in a 250-page opinion that AI could use the material to learn but that acquiring books through piracy to do so was separate infringement. 5 billion settlement, covering some 482,000 works; final approval was made in May 2026. S.
The OpenAI/New York Times MDL Sixteen copyright cases that were brought by the New York Times and many of its peers as well as famous writers against OpenAI have been merged before a single federal judge in the Southern District of New York. The trial became as notable for its discovery proceedings as for the core legal arguments. In January 2026, the judge ordered OpenAI to produce 20 million anomyized conversations from ChatGPT; in March, it was more than ten million more conversations. It’s expected that what the content of these conversations reveals could determine the terms of settlement.
Disney, Universal, and Warner v. Midjourney Major film studios have taken image-generator Midjourney to court, claiming it has built its training dataset on their characters and is continuing to produce work using those figures. The case is currently in process, and its significance lies in its expansion of the AI-copyright battle into the realms of high-value visual content that’s arguably more challenging for AI companies to defend than purely textual, stylistic claims. The Music Industry Split The recording industry has come to different decisions regarding AI: Warner Music and Universal have both reached licensing deals with AI Music and other AI companies, while Sony Music is actively litigating against other AI music startups with the expectation of a fair use ruling next summer.
1 billion in the case, much like the earlier one against the company. Taken together, these rulings haven’t created a unified approach to AI and intellectual property law, but they establish a clear trend: courts are increasingly inclined to find AI training methods transformative, but not at the expense of how the data itself was acquired, and they are collectively united in denying AI full ownership rights.
Conclusion
AI hasn’t really broken the law as much as it’s revealed just how much of it was tailored to a world of discrete, accountable human actions-the singular authorship of a book, the inventive act of an individual applying for a patent, the decision by a single company to copy a competitor’s product.AIblurs that accountability through the flow of data pipelines, training runs, and probabilistic results, and now the law is slowly stitching it back together-case by case and sentence by sentence-determining where responsibility should fall when one single, discernible human action did not lead directly to the damage.
For now, a couple of things are probably safe:
No AI will soon get copyright, patent inventorship or the full dignity of legal personhood. The questions of how AI was trained-not just what it does with training data-matter immensely. The law isn’t going to be set by one landmark case; it is being assembled out of discovery requests, license agreements, settlements, and small appeals filtering through three different continents. For those companies building and deploying AI, as well as artists and writers who create its training material, and the lawyers mediating for them all, AI law in 2026 is not a defined set of guidelines.
It’s a continuing negotiation, and with the expected rulings later this year, the needle will move again.
Frequently Asked Questions:
1. Can a computer program, created by itself, be eligible for a copyright of what it produces?
Nope, the courts have repeatedly affirmed a computer system can’t be the “author”. The Thaler Case – when a Patent and Copyright Office denied patent protections for an AI-created artwork – this was reaffirmed by the Supreme Court in 2022 when they refused to override that prior judgment in that case – that a work is not eligible for copyright protection when a human creator refuses credit.
A work produced with AI assist can still obtain protection, only the portion created by the human can get the copyright protections.
2. Are AI companies breaking the law by using copyrighted material for their model training data?
It’s still murky, complex, and facts-specific.
There has recently emerged a growing distinction between training the AI (is training “transformative” – a potential part of “fair use” protections? and how that training data was legally acquired in the first place. 5B settlement over claims of using pirated books in their model-building. So the company is being accused not necessarily of the act of building models (vs whether it’s fair use to do so), but rather for how they sourced their raw material.
3. If a company’s AI is classified under the EU AI Act’s high-risk provisions, what does that mean?
Actually, it means nothing yet (well – most of it, anyway). The highest and strictest requirements regarding biometric identification systems, hiring processes, credit scoring or the police tools haven’t taken effect until 2027 or 2028 if the AI is integrated into existing products such as medical equipment.
However, companies already have to designate such systems now, because they have to deal with required documentation, risk management and an oversight body and preparation work.
