The US Supreme Court declined on March 2 to hear Thaler v. Perlmutter, and it gave no reasons. That silence leaves AI-generated work copyright where the D.C. Circuit put it, which is no human author, no copyright. If your firm sells images, copy or code that an AI helped make, the job this quarter is a paper trail showing which parts a person decided.
Key points
- What did the Supreme Court decide on AI-generated work copyright? Nothing on the merits. By declining Stephen Thaler’s petition, the Court left standing a March 2025 D.C.
- Why plan around a case the Court wouldn’t hear? Because a pass is still an outcome. The appellate ruling says a person has to author the work, and we don’t expect a higher court to revisit it soon.
- How much human input is enough? Nobody can give you a bright line. Reed Smith’s review of the Copyright Office’s practice points to two refusals.
- Which contract clauses protect what you sell? Four, and none needs a lawyer to explain it to your team.
What did the Supreme Court decide on AI-generated work copyright?
Nothing on the merits. By declining Stephen Thaler’s petition, the Court left standing a March 2025 D.C. Circuit decision that affirmed the Copyright Office’s refusal to register an image Thaler said his own system made without human help. Copyright, the appeals court held, requires authorship “in the first instance by a human being.”
The petition’s denial isn’t an endorsement. It does remove the near-term chance of a change from the top.
Thaler applied in 2018 and listed his software as the sole author. He also argued the machine was his employee, so the work belonged to him as work made for hire. The appeals court rejected that, reasoning that the authorship requirement applies to every eligible work, including those made for hire. That is the whole holding. It covers work made autonomously, which is rarer in business than the headlines suggest.
Why plan around a case the Court wouldn’t hear?
Because a pass is still an outcome. The appellate ruling says a person has to author the work, and we don’t expect a higher court to revisit it soon. Anyone betting on a quick reversal is betting on a different case, different facts and a long wait.
The lazy reading is that nothing changed, so nothing needs doing. We read it the other way. Before the petition, you could tell yourself the question was open at the top. It’s now closed enough to plan around, and planning costs a few clauses and a folder. Finding the gap in the middle of a dispute costs a great deal more. And the people who would need to produce the evidence, an agency account lead or a freelancer on another contract, are the least motivated to find it for you after the invoice is paid.
How much human input is enough?
Nobody can give you a bright line. Reed Smith’s review of the Copyright Office’s practice points to two refusals. In one, images were not protected despite hundreds or thousands of descriptive prompts. In the other, a work was refused despite more than 624 text prompts.
The same sources say human selection, arrangement or substantive modification can support protection for those human parts.
Picture a marketing team that asks an agency for 40 product images. The agency generates them from text prompts and emails over a zip file. On the cases above, the team may own very little it can stop a competitor from copying. Now suppose its designer composes, retouches and combines those images into a campaign. The arrangement and the edits are the part it can credibly defend. The zip file is a lunch order. The campaign is the meal.
Which contract clauses protect what you sell?
Four, and none needs a lawyer to explain it to your team. Most of the exposure sits in agency and freelancer contracts nobody has reread since AI tools arrived. Your own counsel should draft the final wording, since we are describing a practice and not giving legal advice.
- Disclosure. Require suppliers to say, per deliverable, whether generative AI produced any part of it.
- Human contribution. Ask for a short written note of what a person selected, arranged or changed.
- Warranty. Make the supplier warrant that it has the rights it is passing on, and say plainly that this excludes anything it can’t own.
- Your own files. Keep prompts, drafts and edit history for anything you may want to register or enforce.
A usable record is three lines per deliverable. Who chose the concept, what the designer or writer changed after the tool produced its draft, and the date. Attach the file versions before and after. It takes ten minutes when done as the work happens, and it can’t be reconstructed honestly two years later when a dispute starts. Memory of who did what fades fast, and a supplier who has since lost the account has little reason to help you rebuild it.
Treat the clause list as a ranking too. Disclosure comes first, because you can’t judge a deliverable until you know what made it. The warranty comes last, since a promise from a supplier is only as good as its willingness to pay when the promise fails.
If you already have an AI use policy, our piece on staff privacy policy for ChatGPT Health shows how to write rules staff will follow, and the record-keeping step belongs in the same document. For the video side of the problem, our look at Disney’s copyright fight with ByteDance over Seedance covers who carries the risk when generated media goes into a campaign.
Does the US rule hold in Canada?
Unknown. We haven’t checked for a Canadian court or Canadian Intellectual Property Office ruling on it, and a US decision on a US statute doesn’t bind anyone here. Canada has no federal AI statute to fall back on either, as our explainer on the stalled Canadian AI regulation lays out.
Our earlier look at Canada’s AI regulation reset covers how the rules were meant to arrive.
Other countries differ too. The UK takes a separate route through computer-generated works, which Pennington’s review says has never been tried on modern generative AI. A work can be protected in one country and unprotected in another, so ask your own counsel before assuming Canada matches.
Is a paper trail worth it for every asset?
No. Competitors rarely copy a product photo, enforcement is expensive, and nobody needs a paper trail for every internal slide or stock thumbnail. It is worth keeping for assets you’d hate to see copied, such as a brand mark, a signature illustration or a code library you license to customers. For those, a note is cheap insurance.
It gets worth it for the things you’d hate to see copied, like a brand mark, a signature illustration or a code library you license to customers. For those, a note in a folder is cheap insurance. Thaler’s case involved a fully autonomous system, and most business use is mixed. That mixed middle is untested, and an appeals ruling on AI-assisted facts could move the line either way. We’d also back off if courts started treating prompt-only output as protectable, or if your business runs on speed rather than exclusivity.
Until then, pick the three assets you’d least like to lose and ask the agency this week what a person changed in each.
Frequently asked questions
Can I copyright something an AI made?
Not on the current US position if a machine made it entirely without human authorship. Protection may reach the human-authored parts, such as selection, arrangement and substantive edits.
Did the Supreme Court rule that AI work can’t be protected?
No. It declined to hear the case, which leaves the lower-court ruling in place without adding its own reasoning.
What should a small business do now?
Add an AI disclosure clause to supplier contracts and keep dated records of the human decisions behind anything you may need to enforce.
Archive entry dated March 3, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
