Can You Own and Safely Publish AI-Written Content? What the Courts Actually Decided

The one-paragraph version. A machine cannot be the author of a copyrighted work. A person who uses a machine can be. And a model’s output can still infringe someone else’s work, even when nobody can own the output. Those three statements are stable enough to plan around. Nearly everything else sold as “settled” in this area is not.

Updated September 30, 2026 · About 13 minutes · Checked against court filings, agency documents, and law-firm analyses (source ranking at the end)

My position, up front

Most content teams have their risk ranking backwards.

They worry about being sued over generic AI-assisted articles. On the evidence I could find, that risk is real only for a narrow set of outputs: summaries and retellings of specific, expressive works, and verbatim regurgitation. What they underrate is the quieter loss: publishing AI-heavy work they barely touched, then discovering they have little to enforce when someone copies it.

The state of play in five lines:

• Ownership rule: The Supreme Court declined to review Thaler v. Perlmutter on March 2, 2026. Fully autonomous AI output cannot be registered. That decision does not say how much human input is enough.
• The open question: Allen v. Perlmutter (D. Colo.) is fully briefed. As of the most recent independent status report I found (August 12, 2026), there was no merits decision. Check the docket before citing anything about it.
• Exposure: Every output-infringement ruling I found that let claims proceed was against an AI developer, not a downstream publisher. Publisher risk is a reasonable inference from basic copyright law. It is not yet a holding.
• Contracts: Vendor indemnities exist, but their exclusions for “modified” and “combined” output overlap with normal professional use.
• What to do: Keep two separate records, one for what you made and one for what your output resembles (section 7).

Key insight most coverage misses. Ownership and exposure are two different problems with two different fixes. Editing and documentation strengthen what you can enforce. Neither one protects you if your output copies someone else’s work. Teams that treat “we edited it heavily” as a safety measure are solving the wrong problem.

Correction ledger (for returning readers).

1. The February 2026 version said a November 2025 Cohere ruling covered “stylistic synthesis.” It did not. That case concerned “substitutive summaries,” verbatim copies, and trademark claims tied to hallucinated articles. Fixed.
2. That version described three separate SDNY rulings on outputs; the October 2025 “narrative” and “outline” points come from a single OpenAI opinion. Fixed.
3. It dated the Copyright Office’s brief to January; the filing copy is stamped February 6, 2026, while several secondary sources say January. Discrepancy flagged in section 3.
4. In this revision: the Copyright Office’s count of registered AI-assisted works (“thousands”) could not be re-verified, so it is now attributed only to the Office’s own 2025 report. Softened.
5. In this revision: the vendor-contract analysis is a Westlaw Today piece dated September 9, 2026, republished by its authors’ firm. The plan-tier details (which subscription is covered) do not come from it and are labeled as unverified. Clarified.

1. What Thaler decided, and what the headlines added

Stephen Thaler listed his own AI system as the sole author of an image and disclaimed any human creative contribution. The Copyright Office refused registration. The D.C. Circuit affirmed on March 18, 2025. The Supreme Court denied certiorari on March 2, 2026, without comment.

Two things are true at once, and the coverage often blurred them.

What it settles. In courts that follow the D.C. Circuit, a work with no human author cannot be registered. The government argued the case was a poor vehicle for bigger questions precisely because Thaler disclaimed any human input.

What it does not settle. A cert denial is not a ruling on the merits. The D.C. Circuit itself did not hold that copyright is inherently barred for works made with AI assistance; the line depends on the respective roles of human and machine, as Carlton Fields noted. Copyright lawyer Aaron Moss framed it well: Thaler asked whether nothing is enough; the harder question is whether a lot is.

Evidence grade: strong. Appellate holding plus cert denial, confirmed across several independent sources (Constitution Center, Norton Rose Fulbright, law-firm summaries). I did not pull the Supreme Court docket page itself.

2. What the Copyright Office told Allen, and why it matters more than the headline

Jason Allen entered prompts into Midjourney at least 624 times, picked an image, cleaned it up in Photoshop, upscaled it, and won a state-fair prize. The Office required him to disclaim the AI-generated material and, because he refused, would not register the work as submitted.

That last detail changes how you should read the dispute: the refusal turned on the disclaimer, not on a blanket rule against AI. In the government’s framing, Allen could have registered the human-authored portions.

But the offer is narrower than “everything you did by hand is protected”:

Three points deserve more attention than they got.

The offer isn’t a guarantee. The Office invited Allen to file again and explain why his edits rise to authorship. That is an invitation to argue, not a promise of a certificate.

Partial registration of AI-assisted work is normal practice. The Office’s January 2025 Part 2 report sets out the disclaimer approach. It is a routine mechanism, not a loophole.

Allen concerns a 2022 workflow. The Office’s brief reportedly notes that the case does not concern newer image models that allow iterative, plain-English editing. Whatever the Colorado court decides, it will rule on a whole-image, re-roll process. If your team edits regions of an image or passages of text by instruction, the ruling may not map onto what you do.

The pattern that fits every Office decision I looked at

This is my analysis, not a legal holding: the Office draws the line at control over identifiable pieces of the expression, not at effort or iteration count.

• Zarya of the Dawn (comic): human-written text and human selection and arrangement of panels were protected; the individual Midjourney images were not.
• A Single Piece of American Cheese: registered in January 2025 after 35 rounds of inpainting, on the basis of selection, coordination, and arrangement of AI-generated elements.
• Allen: 624 whole-image prompts, no control over any specific region of what came back.

Inpainting lets you direct a change to a particular part of the image. Re-rolling a prompt does not. If that reading is right, the count of prompts is a distraction, and the useful question is: can I point to a specific part of this work and say I decided that, not the model?

Second-order effect. The outputs that need the least editing are the ones with the thinnest ownership claim. A team that ships AI output as-is because it was “already good” has, by the Office’s logic, authored very little of it. Polish that came from the model is not polish you own. This follows from the Office’s reasoning; no court has applied it to a text workflow yet.

3. Where Allen stands today

The complaint was filed September 26, 2024. Allen moved for summary judgment on August 25, 2025. The government’s cross-motion is dated January 2026 in most secondary sources (one law-firm blog says the copy hosted by the Copyright Office is stamped February 6, 2026, so treat the exact date as a small unresolved discrepancy). Allen’s reply came in February 2026, and summary-judgment briefing is complete.

I found no merits decision. The most recent independent tracker I located reported none as of August 12, 2026, and I found no later report through September 30.

My read, low confidence: the case is reviewed under the APA’s deferential “arbitrary and capricious” standard, which makes a clean, sweeping Allen victory harder than headlines imply. A narrow ruling on the disclaimer issue is at least as plausible. The opinion itself is what would resolve this.

4. Can someone sue you over AI output? What the rulings show

Start with the principle. An uncopyrightable work can still infringe another author’s work; the Copyright Act does not require the infringing copy to be protectable itself. AI output that reproduces protected expression can infringe. Thaler changes none of that.

Now the cases, with the defendant listed, because that is the detail most summaries skip.

What the newest ruling changes, and what it doesn’t

Doe v. GitHub is the most recent appellate word, and it is easy to over-read. It closed one route to DMCA statutory damages (up to $25,000 per violation) against developers. It did not say AI output can’t infringe. The court repeatedly separated the DMCA theory from traditional infringement and noted that substantial similarity between an output and a protected work could still support an infringement claim. For a publisher, the practical reading is: the DMCA shortcut got narrower; the ordinary “does this copy protected expression?” question did not.

The real disagreement between sources

The courts are not moving in one direction. The judge who let novel-summary claims proceed had already dismissed news-summary claims, then dismissed a news group’s abridgment claims in the same October opinion. A different judge let “substitutive summaries” of news go forward in Cohere. The likeliest reconciliation is that content matters: highly expressive fiction gets protection at the level of plot and character, factual reporting gets less. But Cohere also involved retrieval of the publishers’ own articles, a different mechanism. Anyone telling you the doctrine on AI summaries is consistent is smoothing over real tension. Authors Alliance, for one, called the Game of Thrones reasoning thin for a ruling at the pleading stage.

What I could not find

I found no reported decision holding a downstream publisher liable solely because it published a model’s output. Nothing I reviewed rules that out. Intent is not required for direct infringement, and no settled framework yet sorts liability among developer, platform, and end user. The exposure is plausible. It is not demonstrated.

My exposure tiers

This framework is mine, built from the rulings above. It is analysis, not law.

5. Vendor indemnification: read the exclusions, not the press release

Most major providers offer some IP indemnity. The useful question, in the words of a September 2026 analysis by Sterne Kessler’s Sasha Rao and Todd Hopfinger (originally in Westlaw Today, September 9, 2026), is whether coverage survives the way you actually deploy the technology.

What that analysis reports about the terms it reviewed:

• Modified output. Anthropic’s commercial terms exclude modifications by the customer; OpenAI excludes output that has been modified, transformed, or combined with non-OpenAI products; Google limits coverage to certain categories of unmodified output.
• Combinations and integrations. Wrappers, retrieval systems, custom instructions, and product integrations may fall inside “modification” or “combination” exclusions, depending on the contract language.
• Customer inputs. Anthropic excludes “inputs or other data provided by Customer,” and other providers make similar carve-outs. That matters if you paste in a protected article and ask for a summary.
• Knowledge. Anthropic and Google exclude claims where the customer knew or should have known the use was likely infringing.
• Training data. Providers are more willing to cover claims about their own training data than about your prompts.

The authors’ central point is worth repeating: many of these exclusions describe ordinary, intended use, not abuse. A team that edits, integrates, and supplies source material is doing exactly what the exclusions target.

What I did not verify. Some legal explainers say coverage differs by plan (for example, enterprise and API customers versus consumer tiers). The Sterne Kessler piece does not address plan tiers, and I did not confirm each vendor’s current wording. Contract versions also change: the analysis cites Anthropic’s terms as of June 2025 and Google’s as of July 2026. Read the live terms for your plan.

Disclosure. This revision was prepared with Claude, an AI model made by Anthropic. Anthropic’s commercial terms are among those analyzed above. I relied on a third-party law-firm reading of them, not on the vendor’s own description.

6. Non-obvious implication: your safest content is also your least ownable

Put sections 2 and 4 side by side and a tension appears that few teams plan for.

• The content least likely to draw a claim (original analysis and explainers, Tier 4) is also the content where a high-volume, low-touch AI workflow leaves you with the weakest ownership story.
• The content most likely to draw a claim (retellings and summaries of specific works, Tiers 2 and 3) is the content where your vendor indemnity is most likely to be defeated, because you supplied the input.

So the two risks don’t offset each other; they stack in different places. Low-touch generation of generic explainers: low exposure, thin ownership. Tool-assisted summarizing of a specific book or article: real exposure, and probably outside your indemnity. That is why the ledgers below are separate.

7. The two-ledger method (my framework)

Registration helps you enforce rights and gives you procedural presumptions. It is not a shield if someone claims your output copies theirs. So keep two records.

Ledger A: what I made (protects ownership)

• Save the raw model output and the final version, so the human layer is visible by comparison.
• Log the specific decisions that were yours: passages rewritten, sections restructured, elements selected and arranged.
• For high-value work, register the human contribution and disclose AI-generated material following the Office’s disclaimer practice. Timing matters for remedies too: a screenwriter in a July 2026 case registered years late and lost access to statutory damages and fees.
• Caveat: no court has said what proof of human authorship looks like at trial. Moss warns that a brief registration statement may not carry the day. Records are cheap insurance against an unanswered question.

Ledger B: what it resembles (reduces exposure)

• Flag any output that summarizes, retells, or outlines a specific identifiable work (Tiers 2 and 3).
• For anything lyric-, poem-, or passage-like, search a few distinctive strings from the output before publishing.
• Note what source material went into the prompt. Under most indemnities that input is your problem.

8. What would change my mind

The question I cannot answer today: will courts protect prompt-heavy work? The evidence that would settle it is a written Allen opinion, and I could not find one.

9. Action plan

This week

1. Ban one workflow: feeding a specific book, article, or song into a model and publishing the resulting summary, retelling, or outline. It is the highest-value target for plaintiffs on the current rulings, and it likely falls outside your vendor’s indemnity because you supplied the input.
2. Start saving raw output next to the final version for anything you might one day need to enforce.

This quarter

3. Audit your last 90 days by tier, not by date. Tier 1 and 2 items first.
4. Ask counsel four questions about your vendor contracts: Do our edits count as “modification”? Does our retrieval or wrapper count as “combination”? What does our plan tier actually cover? Who bears risk for our own inputs?
5. Choose the small set of pieces worth registering, and register the human layer with an honest AI disclosure.

FAQ

Can I copyright AI-generated content? Not if no human authored it. You can protect human contributions to a work that includes AI material, such as your own writing and your selection and arrangement. In the Office’s view, prompting alone has not been enough.

Did the Supreme Court say AI art can’t be copyrighted? No. It declined to hear Thaler, leaving in place a ruling about work with no human involvement. How much human input suffices remains open.

Can I be sued for publishing AI-written content? In principle, yes: infringement does not require intent. In practice, every output ruling I found that let claims proceed was against an AI developer, and I found no decision holding a downstream publisher liable for publishing output alone.

Does my AI vendor’s indemnity cover me? Possibly, but exclusions for modified output, combinations, and customer inputs may overlap with normal use. Check the current contract for your specific plan.

Do I have to disclose AI use when registering a work? Per the Office’s approach, applicants must disclaim more than a de minimis amount of AI-generated material. The disclaimer can be brief.

Did the GitHub Copilot ruling say AI code is safe to use? No. Doe v. GitHub narrowed one DMCA theory. The court expressly left open whether AI-generated code infringes copyright.

Glossary

• De minimis: too small to matter legally. The Office asks applicants to disclaim AI material beyond this threshold.
• Certiorari (cert): the Supreme Court’s discretionary review. A denial is not a ruling on the merits.
• Substantial similarity: the test for whether a work copies protected expression.
• Indemnity: a vendor’s promise to cover certain third-party claims, subject to exclusions.
• CMI (copyright management information): attribution, notice, and license data that DMCA §1202 protects.

Limitations of this analysis

• Directly read in this update: the Doe v. GitHub coverage (National Law Review), the Sterne Kessler indemnification analysis, and several secondary sources on Allen and Thaler.
• Carried over, not re-read: the brief-specific points about Allen (the disclaimer offer, upscaling, and the note on newer models) come from an earlier direct reading of the Office’s brief and from secondary sources; verify pinpoint citations before relying on them.
• Not read in full: the Stein, McMahon, Munich, Santora, and Doe v. GitHub opinions themselves. Those rows rest on legal commentary.
• Motions to dismiss test whether a claim is plausible, not whether it wins.
• Vendor contracts change often; the plan-tier comparison is unverified.
• The exposure tiers and the “control over identifiable pieces” pattern are my analysis, not holdings.
• Some trackers I encountered were low quality or overstated rulings. I excluded any claim of a merits ruling in Allen that I could not corroborate.

Sources, ranked by strength

Primary documents

1. U.S. Copyright Office Review Board, Théâtre D’opéra Spatial decision letter
2. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 2025)
3. U.S. Copyright Office and DOJ, Response and Cross-Motion for Summary Judgment, Allen v. Perlmutter (D. Colo.)
4. Docket, Allen v. Perlmutter, 1:24-cv-02665 (CourtListener)
5. Doe v. GitHub, Inc., No. 24-7700 (9th Cir. Sept. 16, 2026), opinion on Justia

Expert legal analysis

6. Aaron Moss, Copyright Lately: March 2026 analysis and July 2026 roundup
7. Sasha S. Rao and Todd M. Hopfinger, Sterne Kessler, The Hidden Limits of AI Indemnification (Westlaw Today, Sept. 9, 2026)
8. Ashley M. Robinson, Bradley, 9th Circuit on GitHub Copilot (National Law Review, Sept. 29, 2026)
9. Bird & Bird, GEMA v. OpenAI
10. Loeb & Loeb, Advance Local Media v. Cohere
11. Baker Donelson, Thaler cert denial; Carlton Fields, Thaler; Constitution Center, cert denial

Case-tracking and commentary (weaker)

12. ChatGPT Is Eating the World, on Stein’s Oct. 2025 opinion and Cohere
13. Authors Alliance, critique of the Stein ruling
14. Allen v. Perlmutter status tracker (Aug. 2026)

This article is editorial analysis of public court documents and regulatory filings. It is not legal advice. Consult qualified IP counsel about your specific workflows.

Related on BestPrompt.art: Human Authorship in AI-Assisted Work · AI-Generated Content Best Practices