A marketer publishes an image made from a single prompt, and a rival copies it. Under ai and copyright law, the marketer likely has no copyright claim in the purely AI-generated image.
The rule behind that answer is human authorship. U.S. law requires it, so AI-only output is generally not protected, while work shaped by a person can be. Training on copyrighted works is a separate question, and courts are still working through it case by case, with fair use at the centre.
For the wider policy picture, see our ai regulation explained overview. Businesses with Canadian operations should also read Canada’s AI Rulebook 2026: What New Regulations Mean For Businesses, Developers, And Citizens – Technopolicy.
Copyright in AI-generated output
Output is protected only to the extent a human authored it. A university library guide summarises the position plainly: U.S. copyright law requires human authorship, so what an AI system produces on its own generally falls outside copyright.
Purely AI-generated work
The leading example is Thaler v. Perlmutter. The U.S. Copyright Office had already denied Stephen Thaler’s application in 2019 for lack of human authorship, and Judge Beryl A. Howell of the D.D.C. upheld that denial, according to a Columbia University law student review article. The court granted the Copyright Office summary judgment, reasoning that, in the words of the Copyright Alliance, human authorship is an essential part of a valid copyright claim.
The facts there were extreme. Thaler admitted he had no direct involvement in prompting the artwork, and the court needed only a minimal human connection to decide the case. It did not set the threshold for how much human involvement is enough.
AI-assisted work
Most real projects sit between fully automatic output and fully human work. The Copyright Alliance describes a spectrum with three zones: wholly AI-generated works, human work assisted by AI, and a gray area in between. It adds that protectability should be decided case by case, so no bright-line test exists for how much AI use defeats protection.
That uncertainty matters most to creators. If you rely on AI for drafts but select, rearrange and rewrite the material yourself, your human contribution is the part you can plausibly claim. If you accept raw output as final, expect little or nothing to protect.
Myths versus what the cases support
Several confident claims circulate about AI and copyright. The table sets each against the position supported by the cases and commentary.
| Common belief | What is supported | Planning use |
|---|---|---|
| Paying for the AI tool makes me the copyright owner of its output. | U.S. copyright law requires human authorship, and AI-only output is generally not copyrightable. | Do not build a business model on exclusive rights over unedited output. |
| AI output cannot infringe anyone. | When AI reproduces copyrighted material, traditional infringement standards apply. | Review output for close copies before you publish it. |
| Training on copyrighted books is always infringement. | In Bartz v. Anthropic, training on authors’ books was held fair use because of its transformative nature. | Treat fair use as a live defence for training, not a settled loss. |
| Training is always fair use. | The same court held that Anthropic could not use pirated materials, and plaintiffs in the NYT case argue that training copies can infringe. | Source training data lawfully and do not assume one ruling covers all cases. |
The first and third rows come from a university library guide, and the second and fourth draw on AI Frontiers commentary on the lawsuits.
Training on copyrighted works
Whether training on copyrighted works is legal has no single answer yet. Models are built by scraping the web, using public datasets and entering licensing agreements, as AI Frontiers notes. Each route carries different risk. Fair use is the doctrine that allows third parties to use copyrighted materials without permission in certain circumstances, and it is the main defence for the unlicensed routes.
Where fair use has been accepted
Two decisions favour developers. Appellate judges in Authors Guild v. Google treated Google Books’ copying of entire books to build a searchable database as transformative fair use. In Bartz v. Anthropic, a court found that training on authors’ books was fair use because of the transformative nature of the training, which the library guide describes as the first such decision. That ruling has a limit: the judge held that Anthropic could not use pirated materials.
Where claims are still being tested
The New York Times Company sued Microsoft and others, including OpenAI. The plaintiffs argue that training requires making digital copies of data, and that those copies can themselves infringe. That case is part of nearly 40 U.S. copyright suits against AI companies since 2020, per the June 2025 commentary. The Copyright Alliance goes further, arguing that wholly AI-generated output built on ingested works has the potential to supplant the market for those works, which is the market-harm argument rights holders press against fair use.
For a developer, the practical lesson is that the training copy and the lawfulness of its source both matter. For a creator, the lesson is that a favourable ruling in one case does not remove your ability to object to specific uses.
Ownership, licensing and contracts
Where copyright does exist, the author is the initial owner. The Columbia review article notes that a work-made-for-hire arrangement can shift ownership, since an employer can automatically own the copyright to a work created by an employee. This only helps for the human-authored portion. An employer cannot own a copyright that never arose.
Licensing is the other half of the picture. Taylor & Francis sold access to its research to Microsoft for AI training, according to the library guide. Whether that is lawful depends overwhelmingly on the publishing contract between publisher and author, and most academic authors are not compensated unless a specific agreement requires it. If you sign publishing or platform terms, check whether they allow training use and whether payment is owed.
Teams formalising these controls may also find AI Ethics Certification: How To Choose, Get Certified, And Stay Current In 2026 – Technopolicy useful for building governance skills.
Steps to reduce risk
The cases suggest a short set of habits. They lower exposure but do not guarantee protection or immunity.
- Record human contribution. Because protectability is decided case by case, keep drafts, selections and edits that show a person shaped the final work.
- Avoid pirated sources. The Bartz ruling accepted fair use for training but excluded pirated materials, so use licensed or lawfully obtained data.
- Check output for copying. Traditional infringement standards apply when AI reproduces copyrighted material, so a human should review before release.
- Protect confidential inputs. The library guide advises avoiding confidential, proprietary or restricted data in AI tools.
- Put ownership in writing. Employment and vendor contracts should state who owns what, using work-made-for-hire where employees create the work, and should address training use.
Frequently asked questions
| Question | Answer |
|---|---|
| Did the US Copyright Office reject AI-generated art in the Thaler case? | Yes. It denied the application in 2019 for lack of human authorship, and the district court upheld the denial. |
| Does Bartz v. Anthropic make AI training legal? | No. The court found training on authors’ books to be fair use but held that pirated materials could not be used. |
| How many AI copyright lawsuits are there in the US? | Commentary published in June 2025 counted nearly 40 filed against AI companies since 2020, including the New York Times suit that names OpenAI. |
| Can my employer own work I make with AI? | An employer can automatically own copyright in work an employee creates, but that applies only to material with human authorship. |
