AI Output Copyright Liability 2026: What the Courts Actually Said
Three federal rulings in six weeks confirmed that publishing AI-generated content can expose you to infringement claims—even without intent, even without knowing. Here’s the chain of liability, and what you actually need to do before your next publication.
- Human authorship rule is now appellate law. The D.C. Circuit (March 2025) and likely the Supreme Court (cert denied, early 2026) closed this question: a machine cannot be the sole author.
- But you can still own your work. Copyright Office has registered thousands of mixed human-AI outputs. What it needs: evidence of specific expressive choices you made.
- Output infringement is now a real courtroom risk. Three SDNY rulings (Oct–Nov 2025) let claims survive motions to dismiss — summaries, reconstructions, derivative outlines all at issue.
- Allen v. Perlmutter will set the prompting threshold. 624 iterations, first ruling expected 2026. Your iterative AI workflow is implicated.
- Vendor indemnification promises have structured carve-outs. Read the contract. The scenario you need covered is probably excluded.
- The disclaimer option is real and underused. The Copyright Office told Allen he could have registered the human parts. Most teams don’t know this exists.
The Disclaimer Option Nobody’s Using — and How It Changes Everything
Here’s the thing that didn’t make the headlines: the Copyright Office told Jason Allen he could have won.
Not the whole case — not the full registration for the Midjourney-generated image. But in its January 2026 cross-motion filing, the Office said Allen likely could have successfully registered the work had he agreed to disclaim the AI-generated elements. His Photoshop edits. His Gigapixel AI refinements. The selection decisions. All of that — protectable, registrable, his.
He declined. He wants the whole output or nothing.
That’s his call. But the practical implication for anyone who isn’t making a test case out of their creative process is this: the partial-registration path exists right now, under current law, before Allen is decided. Most content teams don’t know it’s an option. They either assume AI outputs are fully protectable (they’re not) or fully unprotectable (also not true). The reality sits in between, and the Office has been quietly processing mixed-work registrations for two years.
“You can use AI as a tool or list it as the sole author — but you cannot do both and claim full copyright protection.”
Editorial synthesis — sources: U.S. Copyright Office Part 2 Report (Jan 2025), Thaler v. Perlmutter, D.C. Circuit (March 2025)What does a registrable AI-assisted workflow actually look like? The Copyright Office Part 2 Report laid out a rough spectrum. Prompts alone don’t qualify — the model determines execution, and that execution is the Office’s concern. But substantial human rewriting does qualify. Creative selection among multiple competing outputs can qualify. Original arrangement of AI-generated material into a larger human-authored structure — that qualifies. You don’t have to throw out the AI. You have to show the expressive choices that were yours.
Practically: keep prompting logs, iteration records, selection rationales. Not to satisfy a detective — to document the decisions that happened. If you reworked an output, save both versions. That’s not paranoia. That’s the evidence that separates “I made this” from “the machine made this and I published it.”
The human authorship standard creates an invisible problem: the outputs that look most polished — the ones that need the least editing — are also the ones with the weakest copyright claim. A content team doing light editing because the AI output was already “good enough” is exactly the workflow the Copyright Office’s criteria penalize. The less you touched it, the less you own it. The outputs that feel like victories are the ones with the weakest protection.
The Human-Authorship Wall Is Now Appellate Law
Stephen Thaler built an AI system — the Creativity Machine — that generated an image with no human prompting, no selection, no editing. He applied for copyright registration and listed the machine as author. He believed that was accurate. The Copyright Office said no. He sued. Lost at district court. Appealed.
In March 2025, the D.C. Circuit affirmed: the Copyright Act of 1976 “requires all eligible work to be authored in the first instance by a human being.” Thaler petitioned the Supreme Court in October 2025. The DOJ recommended denial early 2026. This is settled appellate law. Thaler almost certainly exhausted it.
What matters for anyone using AI as a tool — not as sole creative agent — is what the ruling doesn’t say. The court didn’t hold copyright unavailable for AI-assisted work. It held a machine cannot be the author. Thaler’s case failed on his specific facts: no human made the expressive choices. That distinction is the legal foundation every AI-assisted workflow now stands on.
The Copyright Office Part 2 Report (January 29, 2025) translated the Thaler rule into an operational spectrum. Tier 1 — peer-reviewed legal analysis Prompts alone don’t qualify: while a prompt describes an idea, the AI determines execution. Substantial human rewriting can qualify. Creative selection among competing outputs can qualify. Incorporation of AI content into a larger human-authored work can qualify. The question isn’t whether AI was in your workflow. It’s whether you can point to the specific expressive decisions you made.
How Many Prompts Does It Take? The Allen Case Is About to Answer That
Jason Allen spent days prompting Midjourney. 624 iterations — not accepting whatever the model produced, but directing it, rejecting outputs, refining toward a vision that existed before the first prompt. When Midjourney finally produced something close enough, he took it into Photoshop, fixed it, ran it through Gigapixel AI, and entered the result in an art competition. Won first place. Then tried to register the copyright. Copyright Office said no: too much machine, not enough human.
In August 2025, Allen moved for summary judgment — arguing that 624 prompts aimed at a specific pre-existing creative vision satisfy the originality threshold. In January 2026, the Copyright Office filed its cross-motion, holding its position that iterative prompting — however detailed — doesn’t transfer the AI’s expressive choices to the human directing them.
Two things are true simultaneously: Allen’s approach represents the highest-investment human-direction case yet to reach this question, and the Copyright Office still says it’s not enough. Cross-motions are fully briefed. Ruling expected in 2026. Every content team with iterative AI workflows is waiting for the same answer.
The Allen disclaimer option and the output-infringement rulings create a situation that neither source addresses independently: a content team that publishes AI output without a copyright registration has no protection against infringement claims and no registration to leverage defensively if sued. The standard advice — “AI outputs can’t be copyrighted, so don’t bother registering” — ignores that registration of the human-contribution layer both (a) establishes a defensible ownership claim and (b) creates the evidentiary record that distinguishes your workflow from the full-AI scenarios courts are finding problematic. The copyright is not just about what you own. It’s about what you can prove you made.
The Output Infringement Problem: What the Three SDNY Rulings Actually Mean
Most AI copyright guides lead with: can you protect your output? Wrong question. The more operationally urgent one is: can someone else sue you over your output — even one you had no reason to believe was infringing?
Three federal rulings between October and November 2025 showed that output-based infringement claims can survive motions to dismiss. Low bar — not a merits determination, fair use arguments remain active ahead of summary judgment. But the claims are in court. They survived initial challenge. And the content types that survived are specific.
Output that reconstructs plot, character, or narrative structure from a recognizable source survived motion to dismiss on substantial similarity grounds. The court found that AI outputs can be “substantially similar” to source material without any intent from the publisher.
An AI-generated structural outline or summary of a specific protected work — one that reproduces the creative selection and arrangement of the source — can constitute infringement independent of verbatim reproduction.
Output that replicates the distinctive stylistic elements of a specific author’s work, not just generic style, survived at the pleadings stage. Fair use is the remaining defense — but it has to be litigated.
| Output Type | Legal Status (Oct–Nov 2025) | Evidence Level | ⚠ Adversarial Column — What This Doesn’t Resolve |
|---|---|---|---|
| Narrative reconstruction of recognizable source | Survived MTD on substantial similarity — SDNY | Moderate | MTD survival ≠ merits win. Fair use arguments not yet decided. No final ruling on liability. |
| Derivative outline or structural summary | Claim viable at pleadings stage — SDNY | Moderate | Applies to outputs that reproduce creative selection of source, not all summaries. Scope of “structural similarity” undefined. |
| Stylistic synthesis of specific named author | Survived MTD — Cohere, SDNY Nov 2025 | Directional | Generic style not at issue. Applies specifically to output mimicking distinctive, identified authors. Breadth of “distinctive” unresolved. |
| Human-edited AI-assisted work (documented) | Registrable under partial disclaimer — Copyright Office | Strong | Requires contemporaneous documentation of human expressive choices. Undocumented workflows don’t qualify retroactively. |
| Prompt-only AI output (no human edit) | Not registrable — Copyright Office, D.C. Circuit | Strong | Thaler settled for fully autonomous AI. Iterative prompting (Allen) still unresolved. Ruling expected 2026. |
The Vendor Indemnification Trap
Every major AI vendor — OpenAI, Google, Anthropic, Adobe — has announced some version of copyright indemnification. The announcements are real. The coverage is narrower than the announcements suggest.
Standard carve-outs across vendor indemnification terms as of early 2026 typically exclude: outputs where the user provided infringing inputs, outputs used in ways the vendor couldn’t anticipate, outputs modified post-generation by the user (which is basically any professional workflow), and claims arising from the user’s specific prompting choices. Read the actual terms. Not the press release.
The scenario where you’re most at risk — an AI-generated summary or reconstruction of copyrighted material that went out without a human review pass — is frequently the scenario that falls into a vendor carve-out. The user provided the source material as context. The user directed the summary. The user published it. That’s often on the user’s side of the indemnification line, not the vendor’s.
“Vendor indemnification covers what the model did. It generally doesn’t cover what you did with what the model produced.”
Editorial synthesis — sources: OpenAI Terms of Service (Jan 2026), Adobe Generative AI Terms (Nov 2025), Google AI Terms (Dec 2025)This isn’t an attack on vendor indemnification. It’s a description of what it is. These programs are designed to cover training-data claims and foundational liability — not every downstream publishing decision. Know the difference before assuming you’re covered.
The Complication Worth Naming
Everything above implies that the legal risk landscape is clarifying and tightening. That’s mostly right. But the thesis-complicating evidence: most output-infringement claims are nowhere near resolution, and the surviving-MTD victories for plaintiffs are the beginning of a long evidentiary fight, not the end.
Fair use is a real defense. Transformative use — an AI output that does something genuinely new with source material rather than reconstructing it — has survived copyright scrutiny before. The court in the SDNY cases found claims viable, not that fair use fails. Authors Guild v. Google (2d Cir. 2015) established that large-scale reproduction can be transformative if the use serves a different purpose than the original. Whether AI outputs that summarize or restructure text cross the transformative line is, right now, genuinely unanswered.
The honest version of the risk picture: it’s elevated and getting more defined, not settled. The output-infringement claims surviving motions to dismiss are a signal, not a verdict. Teams that treat this as resolved in either direction — “we’re definitely fine” or “we’re definitely liable” — are both wrong. The legal framework is mid-construction. You’re building workflow on active scaffolding.
What’s Still Coming in 2026
Allen v. Perlmutter ruling. Expected sometime in 2026 — could land before you’re done reading this, could be six months out. The ruling will define whether iterative human direction over AI outputs confers copyright in the AI’s expressive choices, or only in the distinctly human contributions layered on top. That’s the line every AI-assisted creative workflow is waiting for.
Congressional action: at least three AI-copyright bills had bipartisan support as of early 2026. None passed. The likelihood of a legislative resolution before judicial resolution on the core questions is low. Courts will define the operative standard first.
More output-infringement cases. The SDNY rulings opened a door. More claims will follow, more defendants will face discovery, and the evidentiary record of what “substantially similar” means in the AI output context will start to fill in. By the end of 2026, we’ll have significantly more data on what content types actually proceed past summary judgment.
What this means for your next publication decision
The output-infringement risk isn’t a future problem — it’s an inventory problem. The pieces your team published without a human review pass in 2024 and 2025 are the exposure. Not the next piece. The question isn’t whether to use AI. It’s whether you can demonstrate, for pieces already published, that a human made the expressive choices that differentiate them from the source material they were trained on or prompted with.
What you do: Audit the last 90 days of AI-assisted content. For each piece, ask: does this reconstruct, summarize, or structurally mirror a recognizable protected source? If yes — that’s the output-infringement risk category. For new work, implement a documentation protocol before generation, not after. Prompting logs, iteration records, edit diffs. Not because anyone’s asked for them. Because the moment they become relevant, it’s too late to create them.
Here’s what’s going to stop you: The documentation overhead feels disproportionate to the perceived risk — until you’ve read a motion-to-dismiss brief with your company’s content as exhibit A. The teams that built documentation protocols in 2024 are not the teams facing discovery in 2026.
Stop doing this: publishing AI-generated summaries or reconstructions of specific named works without a human review pass that genuinely engages with the output’s relationship to the source. “We used AI and reviewed it” is not a review. Asking whether the output reconstructs structure, narrative, or creative selection — that’s a review.
The registration question nobody’s asking yet
Your organization probably has an IP registration protocol that predates generative AI. It almost certainly doesn’t address partial-disclaimer registration for mixed human-AI works. The Copyright Office has been processing these registrations quietly for two years. The Allen case may set a prompting threshold — but it won’t resolve the disclaimer option, which the Office has made clear is already available.
What you do: Map your AI-assisted content workflows against the Copyright Office’s registrability spectrum. For any work where documented human expressive choices exist — substantially rewritten outputs, creative selection among competing generations, original arrangement of AI-generated material — establish a partial-disclaimer registration process now, before the Allen ruling changes the landscape in either direction. If the ruling goes against Allen, the partial-disclaimer path becomes the primary registration vehicle for AI-assisted work. Get the protocol in place first.
Here’s what’s going to stop you: The Copyright Office’s partial-disclaimer process requires contemporaneous documentation that most teams aren’t generating. Retroactive documentation doesn’t satisfy the standard. The workflows that qualify need to be identified before the content is published, not after a claim arrives.
Stop doing this: treating vendor indemnification terms as a blanket coverage analysis without reading the carve-outs. The specific scenarios in the SDNY rulings — user-directed AI summarization of protected source material — fall into vendor carve-outs in most current indemnification programs. Confirm which side of the carve-out your workflows sit on before a claim forces the analysis.
The Allen ruling will not end the uncertainty — it will shift where the uncertainty sits. If Allen wins, the prompting-threshold question becomes the new battleground: how many iterations, how specific a pre-existing vision, how substantial the human selection? Every vendor will recalibrate their indemnification terms. Every content team’s “we iterated extensively” claim becomes a legal argument. If Allen loses, the partial-disclaimer path becomes the operative playbook for all AI-assisted content teams. Either outcome means the same operational conclusion: document the human choices now, before the ruling makes that documentation legally consequential.
The framework is being built in real time. Three rulings in six weeks moved the infringement question from theoretical to judicially viable. The disclaimer option is real and underused. The vendor indemnification you’re relying on has carve-outs your team hasn’t read.
What the courts haven’t done: given anyone a clean answer. That’s still coming. But “it’s still being decided” isn’t the same as “it hasn’t started yet.”
It started.
