AI Intellectual Property and Copyright

Modern AI models are trained on enormous datasets scraped from the public internet, much of it copyrighted — books, articles, code, images, music. Whether that training constitutes copyright infringement is one of the most consequential unresolved legal questions in AI, currently being litigated across dozens of active lawsuits (The New York Times v. OpenAI and Microsoft, Getty Images v. Stability AI, Authors Guild v. OpenAI, and others), with no single settled precedent as of today.

The central legal battleground is fair use (in the US) — a defense that permits otherwise-infringing use of copyrighted material, evaluated on a four-factor test: the purpose and character of the use (is it transformative, and is it commercial), the nature of the copyrighted work (factual works get less protection than creative ones), the amount and substantiality of the portion used, and the effect on the market for the original work. AI companies generally argue training is transformative — the model learns statistical patterns, not reproducing specific works — analogous to how a human author "trains" by reading widely. Plaintiffs argue the opposite: that training on their work to build a commercial product which can reproduce substantially similar output, and which competes with and can substitute for the original market, fails the fair use test, particularly on the fourth factor.

This is genuinely unsettled law, not a solved problem with an agreed-upon answer — court rulings so far have been mixed and fact-specific, and the outcome will likely differ by jurisdiction, by whether the model can be shown to reproduce specific protected expression (versus purely statistical pattern-learning), and by whether the plaintiff is a rights-holder with existing licensing markets (news publishers, stock photo agencies) versus one without. The EU AI Act adds a parallel obligation regardless of how the fair use question resolves: general-purpose AI providers must publish a summary of training data and comply with EU copyright law, including respecting text-and-data-mining opt-outs (covered later in this topic).

For AI PMs, the practical implication isn't picking a side in this debate — it's recognizing that your product's underlying training data is a live legal risk you don't fully control, especially if you're building on a third-party foundation model. This risk needs to be actively managed through vendor contracts, not assumed away.

A separate question from training data legality: once a model generates an output, who owns it? In the US, the Copyright Office and courts have consistently applied a human authorship requirement — copyright protects the product of human creativity, and works generated without meaningful human creative control are not eligible for copyright protection at all. The 2023 Thaler v. Perlmutter case affirmed this squarely: an image generated autonomously by an AI system, with no human creative input, could not be copyrighted.

The Copyright Office's 2023 registration guidance draws a more nuanced line for the common case where a human uses AI as a tool: works containing AI-generated material can be registered, but only the human-authored portions are protected, and applicants must disclose which parts of a submitted work were AI-generated. A photo edited with AI-assisted retouching is different from an image produced entirely from a text prompt with no further human creative arrangement — the former likely retains protection for the human's contributions, the latter likely doesn't qualify at all. This is a fact-specific, case-by-case determination, and the line between "AI as a tool a human directs" and "AI as the sole author" is not always crisp.

Other jurisdictions take meaningfully different approaches. The UK's Copyright, Designs and Patents Act (CDPA) Section 9(3) explicitly grants copyright in a "computer-generated work" to "the person by whom the arrangements necessary for the creation of the work are undertaken" — a framework that predates modern generative AI but was designed precisely for this case, and grants ownership even without direct human creative authorship. This creates real divergence: content that's protectable in the UK may be unprotectable (public domain, effectively) in the US.

For AI PMs, this has concrete product implications. If your product generates content for users — marketing copy, images, code — your terms of service need to address ownership clearly, and you can't promise users clean, enforceable copyright over pure AI output in the US. Enterprise customers in particular will ask about this directly, since a competitor could legally reuse unprotectable AI-generated content with no recourse.

As litigation risk has grown, a licensing market for training data has emerged as the more durable alternative to relying on fair use. Major AI companies have signed licensing deals with publishers (OpenAI with The Associated Press, Axel Springer, and others; various deals with academic publishers) and with stock content providers (Shutterstock and Getty Images license both their libraries to AI companies and, separately, market AI-generated content trained on properly licensed data as a lower-risk alternative for commercial customers). For AI PMs building products that generate or rely on content, understanding whether your underlying model's training data is licensed, scraped, or a mix materially affects your own product's risk profile — pass-through risk from your model vendor becomes your risk too.

On the input side, text-and-data-mining (TDM) opt-out mechanisms let rights-holders exclude their content from AI training. The EU's DSM Copyright Directive establishes a TDM exception but allows rights-holders to opt out (commonly implemented via robots.txt directives and the TDMRep protocol, a machine-readable standard for expressing reservation of rights). Whether opt-out signals are reliably honored by crawlers, and what happens to content scraped before an opt-out was added, remain contested practical questions — but the direction of travel across most jurisdictions is toward giving rights-holders more explicit control, not less.

On the output side, content provenance standards are emerging to address a related but distinct problem: as AI-generated content becomes harder to visually distinguish from human-created content, there's growing demand — from regulators, platforms, and rights-holders alike — for a reliable way to label what's synthetic. The C2PA (Coalition for Content Provenance and Authenticity) standard, backed by Adobe, Microsoft, and major camera and AI companies, defines a cryptographically verifiable metadata standard ("Content Credentials") that can travel with an image or video, recording whether and how it was AI-generated or edited. Several jurisdictions are moving toward legal requirements for labeling synthetic content, particularly for political and commercial contexts. For AI PMs, adopting provenance standards proactively is increasingly a trust and compliance feature, not just a nice-to-have.

Copyright isn't the only IP exposure. Right of publicity — a state-level (in the US) legal doctrine protecting a person's name, image, likeness, and voice from unauthorized commercial use — is squarely implicated by any product feature involving voice cloning, face-swapping, or likeness generation. Unlike copyright, right of publicity doesn't require the person to have "authored" anything; it protects their identity itself. Tennessee's ELVIS Act (2024) specifically extended state right-of-publicity protections to cover voice, explicitly targeting AI voice cloning — a preview of where more states are likely headed. Any product allowing users to generate content in the likeness or voice of a real person (celebrity, private individual, or a user's own likeness reused without consent) carries meaningful legal exposure, independent of any copyright question.

Trademark risk arises when AI-generated content incorporates brand names, logos, or trade dress in ways that could confuse consumers about source or endorsement — a code-generation tool producing a logo that resembles a protected mark, or a marketing copy generator producing text that implies a false brand partnership. This risk is usually lower-stakes than copyright or right of publicity but is worth explicit output filtering for products generating visual or brand-adjacent content.

Deepfake-specific legislation is a fast-emerging category distinct from general IP law, driven primarily by two harm categories: political disinformation (several US states now regulate AI-generated political ads near elections) and non-consensual intimate imagery (a growing number of jurisdictions have criminalized AI-generated non-consensual explicit content, with federal legislation like the US TAKE IT DOWN Act targeting this specifically). These laws often carry criminal, not just civil, liability, and platform-level obligations (takedown requirements) are increasingly common alongside individual liability.

For AI PMs, the practical takeaway is that any feature touching a real person's likeness, voice, or identity needs its own dedicated risk review, separate from the standard training-data copyright analysis — the legal frameworks, the affected stakeholders, and the mitigation strategies (consent flows, identity verification, output filtering) are meaningfully different.

Given the unsettled state of the law, AI PMs need a practical risk-management playbook rather than waiting for legal certainty that may be years away. Start with vendor indemnification: most major model providers (OpenAI, Microsoft, Google, Adobe) now offer IP indemnification for enterprise-tier customers, contractually agreeing to defend and cover damages if their model's output infringes a third party's copyright — but these commitments typically come with real carve-outs. Coverage is often limited to enterprise/paid tiers (not free or consumer tiers), can exclude outputs from prompts deliberately designed to elicit infringing content, and rarely covers every downstream use case. Read the actual contract, not the marketing page, and understand exactly what's excluded.

Build content filtering for outputs that closely resemble known protected works — increasingly standard for image and code generation products, where output can be checked against known copyrighted material before being served to a user. Adopt provenance and watermarking (C2PA or equivalent) proactively rather than waiting for a legal mandate; it's cheaper to build in from the start and increasingly expected by enterprise customers and platforms alike.

Design your terms of service to clearly and accurately allocate IP risk and ownership between your platform and your users — don't promise users clean copyright ownership over pure AI output you can't legally deliver, and be explicit about what portions of a mixed human/AI work retain protection. For any product built on your own fine-tuned models, maintain data provenance records: what data was used to train or fine-tune, under what license or opt-out status, so you can answer questions from customers, auditors, or in the event of litigation.

Finally, this is genuinely one of the fastest-moving areas in AI product work — new rulings, new state and national legislation, and new licensing norms are arriving constantly. Build a recurring review cadence with legal counsel rather than treating IP risk as a one-time launch checklist item, and design your architecture (content filtering, provenance tagging, consent flows) to be adjustable as the legal landscape shifts, rather than hard-coded to today's rules.