Ethics & Governance

Who Owns the Machine's Mind: The Coming War Over AI Intellectual Property

AI outputs cannot be copyrighted in the US, a rule the Supreme Court made final in March 2026. What is still being decided — by settlements worth $1.5bn, by label-by-label licences, and by a jury trial set for April 2027 — is who owes whom

Gabriele Masetti · 2026-09-02

A New Kind of Author

In March 2023, the United States Copyright Office issued a registration for an AI-assisted graphic novel — then, days later, cancelled the copyright protection for the AI-generated images within it, while maintaining protection for the human-authored text. The message was precise: humans can hold copyright; machines cannot; the line between them is legally significant, practically blurry, and increasingly contested.

That was the opening skirmish in a legal war that has since expanded to encompass every major creative industry, the entire pharmaceutical and materials science pipeline, the software development sector, and the fundamental architecture of intellectual property law itself.

The questions at stake are not procedural. They reach to the foundations of why intellectual property rights exist at all — to incentivise human creativity and innovation — and what happens to that justification when the creator is not human, when the "inventor" is an algorithm trained on the accumulated intellectual output of millions of people who were never asked and never compensated, and when creativity and invention operate at a scale and speed that no IP system designed for human creators can plausibly accommodate.

The Copyright Doctrine in Crisis

The cornerstone of US copyright law's position on AI-generated content was established by the DC Circuit Court in its 2025 ruling in Thaler v. Perlmutter: AI outputs cannot be copyrighted. It stopped being one circuit's holding on 2 March 2026, when the Supreme Court denied certiorari, leaving the rule as settled federal law. The court held that copyright protection requires human authorship — an author who exercised sufficient creative control over the expressive choices in the work — and that prompting a generative AI system, however sophisticated the prompting, does not constitute the kind of human creative expression that copyright was designed to protect.

The United States Copyright Office's January 2025 report on AI-generated content reinforced this position while attempting to draw a workable line: copyright protects those elements of AI-assisted work that reflect human creative choices — the specific framing of a prompt, the selection and arrangement of AI outputs, the human-authored portions of a hybrid work — but not the AI-generated elements themselves.

The practical consequences are not yet fully absorbed by the industries that depend on copyright. A novelist who uses Claude or ChatGPT to generate initial drafts, then revises and refines them, occupies an uncertain position. A graphic designer who uses Midjourney or Stable Diffusion to generate illustrations for a commission may be unable to hold copyright in the generated images, potentially undermining their client relationships and commercial model. A software developer whose entire codebase was generated by an AI tool may have no copyright protection in the code itself.

The "human-in-the-loop" framing that companies and practitioners have adopted as a response — asserting that sufficient human guidance and selection turns AI-generated content into copyrightable human authorship — has not been consistently accepted by courts. How much human creative engagement constitutes sufficient authorship is an unresolved question on which different courts in different jurisdictions have reached different conclusions.

The Training Data Reckoning

Running parallel to the output ownership debate is a second legal battleground that may ultimately prove more consequential: the liability of AI developers for training their models on copyrighted content without authorisation.

The major foundation model developers — OpenAI, Anthropic, Google, Meta, Stability AI, and others — trained their systems on internet-scale datasets that included substantial quantities of copyrighted text, images, code, and audio. These datasets were assembled through web scraping and other automated collection methods. The copyright holders whose work was included in these datasets did not consent, were not informed, and were not compensated.

The litigation response has been substantial. The New York Times filed suit against OpenAI and Microsoft, alleging that its journalism was reproduced in training data and that the model can, in certain circumstances, reproduce Times content verbatim. The Authors Guild, representing thousands of professional writers, filed class actions against OpenAI, Google, Meta, Anthropic, and others. Getty Images filed against Stability AI alleging that its image library was used to train Stable Diffusion without licence.

Two of those matters have moved past the filing stage. In the United Kingdom, the High Court rejected Getty's secondary-infringement claim after Getty discontinued its primary copyright claims, leaving it a narrow trade mark win confined to older Stable Diffusion versions. At the consequentials hearing in December 2025, Mrs Justice Joanna Smith granted Getty permission to appeal the secondary-infringement point, calling it a "pure question of law" on which "the minds of reasonable lawyers may differ." In the United States, Andersen v. Stability AI, the visual artists' case over image-model training, is the first US case set to put that question to a jury; the trial was continued from September 2026 to 5 April 2027.

The central legal question is whether training on copyrighted data constitutes fair use under US law — or equivalent doctrines in other jurisdictions. The analysis turns on four factors: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect on the market for the copyrighted work.

The fourth factor has proved to be the most contested and the most consequential. US District Judge Vince Chhabria, in the class action brought by book authors against Meta, characterised the central question as whether AI training on copyrighted books "obliterates the market" for those books by producing AI systems that can substitute for the original works. If AI-generated text can substitute for human-authored books in commercial markets, the training use cannot be fair use regardless of its technological novelty.

The Anthropic settlement is no longer a proposal. Judge Araceli Martínez-Olguín granted final approval in Bartz v. Anthropic on 20 July 2026: $1.5 billion, roughly $3,000 per work, four times the $750 statutory minimum for ordinary infringement. Attorneys' fees were cut to about $101.56 million, 6.8% of the fund, from the 20% requested. The release was drafted narrowly enough to preserve future claims, so the largest number yet attached to AI training settles a price without settling the doctrine.

The Patent Law Parallel

While copyright battles have attracted the most public attention, an equally fundamental disruption is occurring in patent law — with potentially greater consequences for industrial innovation.

Patent systems were designed around a central concept: invention by human beings. The doctrines of inventorship, disclosure, and non-obviousness all presuppose a human inventor who conceived of the claimed invention, understood its principles, and could explain it to others skilled in the relevant field.

The challenge posed by AI-assisted invention is not new but has intensified. The case of DABUS — an AI system developed by Stephen Thaler (the same figure involved in the copyright litigation) that autonomously identified novel inventions — has been litigated in multiple jurisdictions with divergent results. The European Patent Office, UK Intellectual Property Office, and US Patent and Trademark Office have all held that inventors must be human. The Australian Federal Court briefly recognised AI inventorship before being reversed on appeal. South Africa and some jurisdictions have taken more accommodating positions.

Drug discovery is perhaps the most consequential domain. AI systems — notably DeepMind's AlphaFold, Insilico Medicine's generative chemistry platform, and Recursion Pharmaceuticals' phenomics approach — are identifying novel molecular candidates for pharmaceutical development at rates that human researchers could not match. When those candidates proceed to patent applications, the question of inventorship arises directly. It is no longer a hypothetical pipeline: Insilico's TNIK inhibitor rentosertib, which the company describes as the first end-to-end generative-AI-assisted drug, has reported positive Phase IIa results in idiopathic pulmonary fibrosis. A molecule proposed by an AI against a target identified by an AI is now a clinical asset with a patent estate attached to it.

If AI systems cannot be inventors, and if the AI-generated candidate is substantially the product of the AI's autonomous reasoning rather than human conception, then either: (a) the human researchers who directed the AI claim inventorship on questionable grounds; (b) the candidate is not patentable at all, removing the commercial incentive for development; or (c) patent doctrine evolves to accommodate a new category of AI-assisted invention.

Option (c) is the most practically necessary but requires legislative action in most jurisdictions — and the legislative timelines for IP reform are measured in years to decades, while AI capability in drug discovery and materials science is advancing on timelines measured in months.

International Divergence: A Fractured Landscape

If the US position on AI and IP is contested, the international landscape is even more fragmented — and the fragmentation itself creates significant problems for globally operating AI companies and creators.

The United Kingdom's approach has been shaped by a distinctive provision of its copyright law: Section 9(3) of the Copyright, Designs and Patents Act 1988, which provides copyright protection for computer-generated works with no human author, attributing authorship to "the person by whom the arrangements necessary for the creation of the work are undertaken." This provision — written in 1988 to address primitive computer generation of text and music — has been interpreted by some UK courts and scholars to provide a route to copyright protection for AI-generated content, albeit through an indirect attribution mechanism.

The UK Intellectual Property Office has been engaged in extensive consultation on whether this position should be maintained, modified, or abandoned. The current trajectory appears to lean toward maintaining some form of protection for computer-generated works while clarifying its scope.

The European Union has taken the most restrictive approach. The European Parliament's 2025 resolution on generative AI explicitly excluded AI-generated works from copyright protection and emphasised the importance of protecting human creativity from unfair competition by AI-generated substitutes. The EU AI Act's transparency requirements — mandating disclosure when content is AI-generated — are no longer prospective. The Article 50 duties applied from 2 August 2026, and systems already on the market were given until 2 December 2026 to meet the Article 50(2) marking obligation. The rest of the Act went the other way: the Digital Omnibus agreement postponed the high-risk obligations, with stand-alone Annex III systems now due to comply by 2 December 2027 and AI embedded in regulated products under Annex I by 2 August 2028. Europe kept the labelling on schedule and deferred most of the enforcement architecture around it.

China's approach has been more pragmatic and, in some respects, more permissive. The Beijing Internet Court's 2023 ruling in Li Yunkai v. Liu Yuanchun held that AI-generated images could be protected by copyright when a sufficient degree of intellectual input was demonstrated by the human user. This approach — focusing on the effort and creativity of the prompter rather than applying a categorical rule about AI authorship — is closer to the "human creative contribution" standard some US practitioners advocate and US courts have been reluctant to adopt.

The practical consequence of this international divergence is a complex legal geography in which the same AI-generated work may be copyrightable in one jurisdiction, unprotectable in another, and protected under different rules in a third. For global content creators, this creates compliance complexity and strategic uncertainty that has already begun to affect investment decisions about where to locate AI development activities.

The Ownership Architecture of AI Output

Beneath the specific doctrinal debates, a more fundamental question is emerging about the basic ownership architecture of AI-generated intellectual output.

Current intellectual property frameworks distribute rights through a chain: the creator of a work holds copyright; the inventor of a technology holds patent rights; employers may hold these rights through work-for-hire arrangements. The AI disruption breaks this chain in multiple places simultaneously.

Consider the layers of potential claimants to an AI-generated creative work:

The developers of the foundation model that produced the output have trained a system that is demonstrably central to the creation of the work. They invested billions in developing this capability. Without their model, the work would not exist.

The operators — companies that have fine-tuned the model on specific datasets, built applications around it, or integrated it into products — have made additional investments and creative choices that shape what the model produces in their context.

The users who provided the prompts and selected from the outputs have made the most proximate creative choices, however constrained those choices are by the capabilities of the models and applications they are using.

And the human creators whose work was included in training data — without their consent or compensation — provided the raw creative material from which the model's capabilities were synthesised.

No existing IP framework provides a coherent mechanism for distributing rights or revenues among all of these stakeholders. The current de facto arrangement — in which terms of service grant users some rights while reserving others for the platform, and training data contributors receive nothing — is a commercial expedient rather than a principled legal framework.

The Second-Order Economic Effects

The IP uncertainty created by the current legal landscape is not merely a problem for lawyers and legal scholars. It is producing measurable second-order effects on investment and innovation that deserve attention.

The pharmaceutical industry's interest in AI-assisted drug discovery is substantial — the potential to accelerate development timelines from a decade to a few years has obvious commercial and public health significance. But the patent question creates a significant chilling effect on investment in AI-generated drug candidates: if the IP protection for a compound discovered by an AI system is uncertain, the commercial case for investing in the development of that compound (at costs of hundreds of millions to billions of dollars) is correspondingly weakened.

The creative industries are facing a different but structurally similar problem. The inability to establish clear copyright protection for AI-assisted work reduces the commercial value of that work and creates uncertainty about licensing, enforcement, and infringement claims. Some publishers, music labels, and creative agencies have responded by adopting internal policies that prohibit or significantly restrict the use of AI tools in content creation — not because they object to AI in principle, but because the IP uncertainty makes AI-assisted work legally and commercially risky.

The irony is that IP uncertainty may simultaneously incentivise secrecy — retaining AI capabilities as trade secrets rather than filing patents that require disclosure — and discourage the investment in AI development that would produce the most socially beneficial innovation. A system designed to incentivise disclosure and reward innovation is producing, in the AI context, incentives for opacity and defensive IP positioning.

Policy Options on the Table

The policy response to the AI-IP crisis is constrained by the speed of change, the international character of the problem, and the deeply established institutional architecture of IP law — which in many countries requires legislative action to change in fundamental ways.

Several directions of reform are under active discussion:

A new "AI-generated works" category — creating a sui generis form of protection for AI-generated content that is weaker and shorter in duration than human-authored copyright, in acknowledgement that the social justification for copyright protection (incentivising human creative effort) applies differently. This approach is advocated by some technology companies and academics but faces resistance from human creators who fear it would reduce the competitive disadvantage of AI-generated content.

Compulsory licensing for training data — requiring AI developers to obtain licences for training data from a central licensing body (analogous to music performance rights organisations like ASCAP or PRS), with proceeds distributed to rights holders based on usage. Several European collecting societies have been developing frameworks along these lines. The market did not wait for them. Universal settled with Udio in October 2025, with Udio committing to a platform trained only on authorised and licensed music; Warner settled with Suno in November 2025, its artists and songwriters able to opt in and be paid when their names, voices and compositions are used in AI-generated music; BMG signed a global licensing deal with Suno in August 2026 that also settled its past use. Universal and Sony remain plaintiffs against Suno. What has emerged is not a compulsory licence but a label-by-label private one, negotiated under litigation pressure, routing money to the catalogue owners best placed to sue and nothing to anyone else.

Opt-out registries with legal standing — creating official registries in which rights holders can record their objection to use of their work in AI training, with legal consequences for developers who train on registered works without licence. This approach has been adopted in limited form by some AI companies voluntarily but has not been established with legal backing in any major jurisdiction.

Mandatory disclosure and provenance — requiring AI-generated content to be labelled, with metadata establishing the AI system, operator, and generation parameters. This does not resolve the ownership question but creates the infrastructure for downstream legal analysis and market differentiation.

None of these options is sufficient alone, and the international coordination required to make any of them effective at scale faces the same obstacles as other AI governance coordination efforts — the difficulty of aligning the interests of jurisdictions with very different legal traditions, economic interests, and political systems.

The Philosophical Stakes

Behind the legal and commercial questions lies a philosophical one that deserves to be stated directly: does intellectual property law, as a system designed to incentivise and reward human creativity, have a coherent future once industrial-scale creative and inventive capability no longer requires human creators or inventors?

The utilitarian justification for IP rights — that they provide incentives for creative and inventive activity by ensuring that creators can capture some of the economic value of their work — applies with different force to AI-generated content. AI systems do not respond to incentives in the way that human creators do. They do not need the prospect of royalties to motivate them to write novels or compose music. The social benefit of IP protection in this context is unclear.

The Lockean justification — that creators have a natural right to the fruits of their labour because they have mixed their labour with the raw material of the work — applies even more awkwardly. What does it mean to "mix your labour" with a generative AI prompt? Who has "laboured" in the production of an AI-generated pharmaceutical compound — the developers who built the model over years, the scientists who directed the search over months, or the AI system that identified the candidate over hours?

These philosophical questions are not merely academic exercises. They determine what IP systems are for, who they should benefit, and how they should be designed. A legal system that fails to engage with them will continue to produce ad hoc, contradictory, and commercially disruptive answers to questions that will not go away.

The Coming Resolution

The current period of IP law uncertainty around AI is unsustainable. The accumulation of contradictory rulings, the expanding volume of litigation, the chilling effects on investment, and the growing international divergence all create pressure for resolution.

Resolution will come through one of two mechanisms, or both: judicial development of doctrine through case law across jurisdictions, or legislative reform establishing clear rules for AI-generated content and AI-assisted invention.

Judicial development is already underway, and the outlines of a possible doctrinal equilibrium are becoming visible: AI-generated content without significant human creative input will not receive copyright protection; AI-assisted content with substantial human creative engagement will receive protection commensurate with the human contribution; training on copyrighted data will require licence in some circumstances and constitute fair use in others, with the boundary determined by the degree to which AI output substitutes for the original; AI inventors will not be recognised as such, but human directors of AI research will receive patents on AI-identified inventions in recognition of the creative and intellectual contribution of the research direction.

Parts of that equilibrium have since become observable rather than forecast. The human-authorship rule was made final by the March 2026 denial of certiorari. The training-data question is being answered twice at once — by cheques already written ($1.5 billion from Anthropic, label-by-label licences from Suno and Udio) and, from April 2027, by a jury in San Francisco. The price mechanism is arriving ahead of the doctrine, which is not how IP law usually settles anything.

This doctrinal equilibrium is imperfect and will not satisfy all stakeholders. The training data compensation question, in particular, remains structurally unresolved by judicial development alone — the scale of the problem exceeds what case-by-case adjudication can address.

Legislative solutions will take longer, will vary by jurisdiction, and will themselves require ongoing revision as AI capabilities continue to develop. The IP law that emerges from the current period of disruption will likely look different from what exists today in every major jurisdiction — and the process of reaching it will be contested, expensive, and consequential for the entire global creative and innovation economy.

The war over who owns the machine's mind is not ending. It is just beginning.