Who owns a picture that a machine drew? That question sounds simple until money, reputation, and a decade of creative work sit on the answer. In 2026, the rules around AI and copyright have moved from think-piece territory into courtrooms, registration offices, and compliance departments. The picture is clearer than it was two years ago, but it is far from settled. Here is what actually matters, what remains open, and how to protect yourself either way. (Zuckerberg Chan CZI Biohub AI Biomedical Shift 2026: Explained) (Copyright and Artificial Intelligence)
The Foundation: Humans Still Need to Be in the Loop
The single most important principle has not changed: copyright protects human creativity, not machine output. The U.S. Copyright Office made this explicit in its January 2025 report on copyrightability, concluding that material generated purely by AI cannot be registered. Prompts alone, the office said, do not supply authorship. What can be protected is the human contribution around the output: selection, arrangement, editing, and creative modification. (Copyright)
That position has teeth in court. In Thaler v. Perlmutter, the D.C. District Court ruled in August 2023 that a work generated solely by AI without human authorship could not be registered, and the D.C. Circuit affirmed the decision in March 2025. For anyone building a business on AI-generated content, the practical takeaway is blunt: if a human cannot point to a meaningful creative decision, there is probably nothing to own.
This matters more than it sounds. A company that fills a product catalog with unedited model output may find it has no enforceable rights when a competitor copies the text wholesale. Ownership is not automatic, and registration requires disclosure of AI-generated material.
Registration in Practice: Disclose, Then Claim What Is Yours
The Copyright Office’s registration guidance is where theory meets paperwork. Applicants must disclose the use of generative AI in a work and identify which portions are human-authored. A graphic novel with AI-assisted illustrations can still be registered if the human author’s text, layout, and creative arrangement carry the expressive weight.
Common mistakes show up again and again here. People list “AI tool” vaguely instead of describing what the tool produced. They claim the whole work when only the human edits are protectable. Or they skip registration entirely and assume the timestamp on a cloud document proves authorship. It does not.
What works in practice is documentation. Keep drafts, prompt logs, edit histories, and dated versions. If a dispute arrives, the person who can show a creative trail wins far more often than the person who can only show a final file. This is the same discipline that protects any creative project, and it is now the difference between owning an asset and holding a file.
The Training Question: Fair Use Is a Spectrum, Not a Switch
The harder fight is not about output. It is about input. Whether AI companies can train models on copyrighted books, articles, and images without a license has produced a patchwork of rulings, and the results resist easy summary.
In Bartz v. Anthropic, the Northern District of California held in June 2025 that training on lawfully acquired books was transformative fair use, while the acquisition of pirated library copies was not excused. A reported $1.5 billion settlement followed in September 2025. In Kadrey v. Meta, the court ruled for Meta on fair use for training in June 2025, and plaintiffs appealed. Thomson Reuters v. Ross Intelligence went the other way in February 2025, with summary judgment rejecting fair use for a non-generative legal research tool, though that precedent is narrow and should not be stretched to cover general-purpose models.
| Case | Court | Core holding | Status |
|---|---|---|---|
| Thaler v. Perlmutter | D.C. District, then D.C. Circuit | No copyright for purely AI-generated work | Affirmed March 2025 |
| Bartz v. Anthropic | N.D. Cal. | Lawful training transformative; piracy not excused | Settled, reported Sept 2025 |
| Kadrey v. Meta | N.D. Cal. | Training on books was fair use | On appeal |
| Thomson Reuters v. Ross | Federal court | No fair use for legal research AI | Narrow precedent |
| NYT v. OpenAI/Microsoft | S.D.N.Y. | Unauthorized training claims | Ongoing |
The pattern is not “training is legal” or “training is theft.” It is that how a model was built matters enormously. Lawfully obtained data, transformative purpose, and no market substitution push toward fair use. Pirated sources and outputs that compete directly with originals push the other way. For companies licensing content to AI developers, that distinction is the negotiating lever. For developers, it is a procurement rule: know where your corpus came from.
Europe Draws a Different Line
The European Union has taken a regulatory route rather than waiting for judges. The EU AI Act, Regulation (EU) 2024/1689, entered into force in August 2024. Article 53 imposes transparency duties on providers of general-purpose AI models, including a requirement to publish a sufficiently detailed summary of training content. Those obligations applied from August 2, 2025.
Alongside it sits Article 4 of the EU’s Digital Single Market Directive, the text and data mining exception with an opt-out for rights holders, in force since 2021. The combination gives European creators something American creators do not yet have: a formal mechanism to say “not my work” before training begins, backed by disclosure rules.
The contrast is worth stating plainly. The U.S. leans on litigation and registration practice. The EU leans on transparency and opt-out. The UK, by contrast, still has no general mining exception beyond non-commercial research under the 2014 rules, and government consultations between December 2024 and February 2025 produced no settled policy. A company operating across all three markets has to plan for three different compliance postures, not one.
What This Means for Content Creators
For writers, illustrators, and photographers, the 2026 landscape rewards a few unglamorous habits.
First, register your best work and describe your process honestly. A registration that overstates human authorship is worse than no registration, because it can be challenged and invalidated.
Second, treat AI as a tool inside a human workflow, not a replacement for it. The more visible the human choices, the stronger the claim. An illustrator who uses a model for texture and then reworks it by hand has a defensible position. One who publishes raw output does not.
Third, watch the licensing market. Content deals with AI developers increasingly hinge on provenance. If you can prove your archive was never scraped from a pirate site, you are a cleaner counterparty.
Fourth, mind the platform economics. Rights disputes are reshaping app stores and marketplaces at the same time, and fee fights can change what your work earns even when ownership is clear. The ongoing dispute over the Apple scotus 27 percent app store fee 2026 shows how distribution rules and intellectual property collide, and creators should read those terms as carefully as they read a license. (Apple private cloud compute Google cloud 2026: Explained)
What This Means for Businesses Using AI
Companies deploying generative tools face a different set of traps. The most common is assuming that a vendor’s indemnity clause covers everything. It usually covers copyright claims about outputs, not claims about training data, and it rarely covers the scenario where your own team fed copyrighted material into a tool.
A workable internal policy has four parts. Document which tools are approved and for what. Prohibit pasting third-party copyrighted text into unapproved models. Keep human review in the loop for anything published. And record the provenance of training or reference material used in-house.
There is also a strategic angle. Firms that treat AI output as a commodity often discover they cannot defend their brand assets. Those that layer human authorship on top, through editing, curation, and original framing, end up with something both defensible and more valuable. The same logic applies to hardware bets that depend on software ecosystems, as seen in the shifting calculus around the Tesla website robotaxi semi roadster three hardware bets 2026, where intellectual property and platform strategy are increasingly hard to separate. (Tesla SpaceX Merger December 6 AGM Barclays 2026: Explained)
The Open Questions Nobody Has Settled
Honesty requires naming what is still unresolved. The Copyright Office’s pre-publication report on generative AI training, released in May 2025, suggested that training can be fair use in some cases but not others, with market dilution as a key concern. Whether that report was finalized, revised, or withdrawn under the current administration is unclear. No federal AI-copyright statute had passed as of late 2026. The New York Times case against OpenAI and Microsoft, filed in December 2023, continues, and its outcome could reshape licensing norms overnight.
The EU’s general-purpose AI code of practice, drafted through 2025, has its own adoption timeline to verify. And appellate outcomes in Kadrey and related matters remain pending. Anyone who tells you the law is settled is selling something.
How to Build a Defensible Position Now
Practical steps, in order of impact:
- Audit your content pipeline and list every AI tool in use, including shadow tools employees adopt quietly.
- Register your most commercially valuable human-authored works, with accurate AI disclosure.
- Keep provenance records for source material, prompts, and edits.
- Add a human-review checkpoint before anything is published or licensed.
- Review vendor contracts for training-data indemnities, not just output claims.
- Track the EU transparency obligations if you sell into Europe.
None of this guarantees a win in court. It does mean you are not the easiest target in the room.
Conclusion
The 2026 rules on AI and copyright come down to a single idea with many applications: human creativity is what gets protected. Courts have said it about authorship, the Copyright Office has said it about registration, and the EU has built transparency rules around the same instinct. Training disputes remain genuinely unsettled, and the fair use picture depends heavily on how data was obtained. Creators who document their process and businesses that govern their tools well will navigate the next round of rulings far more comfortably than those waiting for certainty that is not coming. The law is moving. The people who move with it, deliberately, keep their assets.
Frequently Asked Questions
Can I copyright something I made entirely with AI?
No. Under current U.S. practice, purely AI-generated material is not copyrightable, and prompts alone do not count as authorship. You can protect the human parts: your selection, arrangement, edits, and creative direction. The Copyright Office report from January 2025 and the Thaler v. Perlmutter rulings both point the same way.
Does training an AI model on copyrighted books count as fair use?
Sometimes. Courts have split. In Bartz v. Anthropic, training on lawfully acquired books was held transformative, but using pirated copies was not excused. Kadrey v. Meta went in favor of fair use for training, and that ruling is on appeal. The answer depends on how the data was obtained and whether the output substitutes for the original.
What does the EU AI Act require from AI companies?
Article 53 requires providers of general-purpose AI models to publish a sufficiently detailed summary of their training content, among other transparency duties. Those obligations applied from August 2, 2025. Separately, the EU’s text and data mining exception includes an opt-out that rights holders can exercise.
Do I have to tell the Copyright Office that I used AI?
Yes, when you register a work that contains AI-generated material. You must disclose the AI use and identify the human-authored portions. Overstating your contribution can invalidate the registration, which is worse than registering a narrower claim.
Is my company liable if an employee pastes copyrighted text into an AI tool?
It can be. Vendor indemnities typically cover claims about outputs, not about training data or about material your own staff supplied. The practical fix is a clear internal policy: approved tools only, no pasting third-party copyrighted work, and human review before publication.
What is the biggest mistake creators make with AI content?
Publishing raw model output and assuming it is an asset. Without meaningful human contribution, there may be nothing to own and nothing to enforce. The second biggest mistake is skipping documentation. Drafts, edit histories, and prompt logs are what win disputes, and they cost almost nothing to keep.
