The Thaler cert denial and human authorship

On 2 March 2026 the Supreme Court denied certiorari in Thaler v. Perlmutter. No argument, no opinion, no merits ruling — the Court simply declined to take the case, which leaves the D.C. Circuit's decision in place. For anyone producing generated work commercially, that is the end of a five-year wait for a definitive answer, and the answer is the one the Copyright Office has been giving all along: copyright protects human authorship, and a work with no human author cannot be registered.

The practical consequence for a production shop is smaller than the headline suggests, and the reason it is smaller is worth understanding properly. This case was never about the work most people actually make.

What was actually decided

Stephen Thaler applied to register a visual work titled A Recent Entrance to Paradise, naming a machine he built — the "Creativity Machine" — as the author, and claiming the copyright himself as the machine's owner. The Copyright Office refused the application on the ground that the work lacked human authorship. Thaler challenged the refusal, lost in the district court, and lost again on appeal: the D.C. Circuit affirmed that the Copyright Act requires an author who is a human being.

The shape of the case matters more than the result. Thaler litigated it on the premise that the machine created the work autonomously, without human creative involvement. That was not an incidental detail — it was the argument. He was asking the courts to recognise a non-human author, not to recognise his own contribution as sufficient.

So the holding that now stands is precisely as broad as that framing, and no broader. A work with no human author is unregistrable in the United States. That is it.

Why a cert denial is weaker than it sounds, and binds you anyway

Two things are true at once and people tend to hold only one of them.

A denial of certiorari is not a decision on the merits. It expresses no view about whether the D.C. Circuit was right. It does not create Supreme Court precedent, and in principle a different circuit could reach a different conclusion in a future case, which is exactly the kind of split that would bring the question back.

And also: the D.C. Circuit is the circuit that matters most here, because appeals from the Copyright Office's registration decisions run through the District of Columbia. Combined with an Office whose registration practice already required human authorship, the operating reality for anyone filing is settled even though the constitutional question technically is not. Waiting for more certainty before adjusting your contracts is now waiting for something with no expected arrival date.

The four questions this does not answer

This is where most commentary overreaches. The decision resolves the zero-human case and leaves the interesting ones open.

How much human contribution is enough. Thaler stipulated to none. The case therefore says nothing about the assisted work almost every commercial shop actually produces — a generated element that a person then selects, edits, composes, cuts or scores. The line between "not enough" and "enough" is drawn by Copyright Office practice and by future litigation, not by this ruling.

Whether training on copyrighted material is lawful. Completely separate line of cases, separate doctrine, separate outcomes. A work being unregistrable says nothing about whether the model that made it was trained permissibly, and the reverse holds too.

Who owns the output as a contractual matter. Platform terms from major providers assign the user contractual rights in outputs, and that continues to function. What a contract cannot do is manufacture a copyright that statute does not recognise — which is why a blanket "all deliverables are original works and all copyright is hereby assigned" clause can be a promise you have already broken. Usage rights in creator contracts covers how those grants are structured in practice, and what a usage right actually is is the short definition.

What happens outside the United States. Jurisdictions differ meaningfully. The UK, for one, has a long-standing statutory provision for computer-generated works with no human author, attributing authorship to the person who made the arrangements necessary for the work's creation — a different starting point entirely. If you licence internationally, the US answer is one input, not the answer. Legal and licensing considerations for business AI content is the wider survey.

What changes in a shop on Monday

If you were already treating raw model output as unowned, nothing changes, and that is the correct reaction. If you were treating the question as open pending a Supreme Court ruling, three things become due.

Fix the warranty language. The originality warranty in most creative services agreements promises something you cannot deliver for purely generated material. The honest replacement is a carve-out stating that the producer does not warrant that AI-generated portions qualify for copyright protection, and that the status varies by jurisdiction. It is a harder conversation than the boilerplate and a much cheaper one than the alternative. What to tell clients about using AI is the framing for having it before signature rather than after delivery.

Put the distinctiveness in the human layer. This is a production instruction, not a legal one. A shop recognisable for a model's house style has nothing to defend, because that style is available to everyone with an account. A shop recognisable for its writing, its editorial rhythm, a named voice and a consistent type system has authored material sitting inside every deliverable. What makes an AI print design legally protectable works through the still-image version of that build.

Keep a production record while you work, not afterwards. Model name, version, date, prompt reference, and a one-line note on what a human changed. Reconstructing it a year later from a folder of flattened exports is not realistic. Generations stay in your history with the prompt attached, and finding something you made before pulls one back by description rather than filename, which is the difference between a record and an archive. On the machine-readable side, what a Content Credential actually records is worth reading before assuming provenance metadata proves authorship — it does not, and it was never designed to.

None of this is legal advice, and the standards are still moving. Take counsel in your jurisdiction before building an enforcement strategy on any of it.

The part that is genuinely new

For five years, "the law is unsettled" was a defensible reason to defer the paperwork. It has stopped being one. The specific question everyone was pointing at — can a machine be an author — now has a stable answer, and the answer removes the excuse without removing the underlying difficulty.

What is left is the harder, more useful work: knowing which parts of your own deliverables carry human authorship, being able to describe them in a sentence, and structuring contracts that do not promise more than that. Shops that did this two years ago are unaffected by the news. Shops that were waiting have a backlog.

FAQ

Does this mean everything I generate is unprotected?

No, and this is the most common misreading. It means a work with no human authorship is unregistrable. Most commercial deliverables are not that — an ad built from twelve generated clips carries a human-authored script, shot selection, order, timing and sound design, and those are contributions a claim can be built on. The generated clips individually remain unprotected; the structure containing them is a different question.

Could a different court still rule the other way?

In principle. A cert denial creates no Supreme Court precedent, so another circuit could reach a different conclusion and put the question back on the docket. In practice, appeals from registration refusals run through the D.C. Circuit and the Office's own practice is unchanged, so the operating position is stable even if the doctrine is not formally nationwide.

If the output is unprotected, can anyone just use my ad?

They can use the parts nobody owns. They cannot use your registered trademark, your script, your recorded voice, your product photography, or your account identity — and platform enforcement processes for impersonation and duplicate creative run on policy rather than statute, often faster than any legal route. Building a mark that survives clearance is where most of the real leverage sits.

Does training my own LoRA change the analysis?

Not directly. A fine-tune changes what the model produces; the output is still model output. Where it matters is upstream: a LoRA trained on drawings you made means those drawings are human-authored works you own, and that ownership sits with the source material rather than with anything generated downstream of it. Licence terms for selling prompts and LoRAs covers what you can actually grant when you sell one.