AI and Data Licensing · Memo

AI Output Rights and Copyright After the 2023-2025 USCO Guidance

The USCO's 2023-2025 guidance and the Thaler decision have set the structural rule on AI outputs. The application is still open, and I want to walk through what is registerable, what is not, and how to draft around the uncertainty.

The Copyright Office's 2023 statement of policy on works containing AI-generated material, the 2024 report on copyright and AI, and the January 2025 Part 2 report (covering copyrightability) together establish the framework. Pure AI-generated output without sufficient human authorship is not copyrightable. AI-assisted work where the human has selected, arranged, modified, or otherwise exerted creative control over the result may be copyrightable in the human-authored elements. Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023), aff'd in unpublished disposition by the D.C. Circuit, confirms the human-authorship rule at the constitutional level under the Copyright Act, 17 U.S.C. section 101 et seq.

The structural rule is clear. The application is where the practitioners do their work.

What counts as 'sufficient human authorship'

The Copyright Office's guidance has identified a few patterns that the Office has, and has not, treated as sufficient.

• Selection and arrangement. A human who selects and arranges AI-generated elements into a compilation may have a copyright in the selection and arrangement, but not in the underlying elements. The Zarya of the Dawn matter (graphic novel) is the canonical example: the registration was granted for the text and the selection-and-arrangement, but not for the AI-generated images themselves.
• Substantial human modification. If a human takes an AI-generated output and modifies it substantially (redraws, rewrites, recomposes), the modifications can be copyrightable. The threshold is fact-dependent. Light touch-ups do not qualify. Substantial creative intervention does.
• Iterative prompting alone is generally not enough. The Office's position, as articulated in the Part 2 report, is that crafting prompts does not constitute authorship of the output, regardless of how detailed or skilled the prompting is. This is the position that is most contested among commentators, and it may be the position that gets litigated.
• Disclosure obligations. The Office requires applicants to disclose AI-generated material in works submitted for registration. Failure to disclose is grounds for refusal or cancellation of the registration.

The registration strategy

For counsel registering AI-assisted works, the structural approach I take:

1. Identify the human-authored elements. Walk through the work and identify what a human created (text, selection, arrangement, modifications). Document the process. Save drafts, version histories, and the prompts.
2. Identify the AI-generated elements. List the material that was generated by AI and the tools used. The disclosure obligation requires this.
3. Register for the human-authored elements only. The application should expressly disclaim the AI-generated material and claim copyright in the human-authored content. This is the approach that survives Office review under the current guidance.
4. Document the creative process. If the work involves substantial human modification of AI outputs, document the modification. The Office's review focuses on the human contribution, and the documentation supports the registration if challenged.

Contractual treatment of outputs

For agreements involving AI-generated outputs (whether software licenses, professional services contracts, or content licenses), the contractual structure now needs to address copyright in outputs explicitly. The drafting moves:

• Acknowledge the copyrightability question. The agreement should not assume that AI outputs are copyrightable. The acknowledgment is not a concession; it is a posture that protects the licensee from over-claiming.
• Allocate human authorship roles. If the work is intended to be copyrightable, the agreement should specify who is responsible for the human-authorship element and what process that party will follow.
• Assign whatever copyright exists. The agreement should include an assignment clause that transfers whatever copyright interest exists, even if uncertain. The assignment is belt-and-suspenders against the possibility that the law evolves to recognize some interest the parties did not anticipate.
• Address trade-secret treatment. Where copyright protection is uncertain, trade-secret protection under the California Uniform Trade Secrets Act, Civ. Code section 3426 et seq., and the federal Defend Trade Secrets Act, 18 U.S.C. section 1836, can fill some of the gap. The contractual treatment should preserve trade-secret status where applicable.

The output-infringement question

The other side of the output-rights question is when an AI output infringes someone else's copyright. The case law is developing across multiple matters. The Doe v. GitHub action raised the question for code outputs. The Andersen v. Stability AI action raised it for image outputs. The NYT v. OpenAI complaint includes specific allegations of verbatim or near-verbatim reproduction of NYT content in model outputs. The legal theories include direct infringement (the output is itself an infringing copy), contributory infringement (the model enables the user to infringe), and vicarious infringement (the model provider benefits from infringement it has the right and ability to control).

For licensees of AI tools used in commercial work product, the practical posture is that some outputs may infringe and the licensee should not assume otherwise. The vendor indemnity discussed in my SaaS IP indemnity memo is the contractual safety net. The licensee's operational discipline is the practical safety net: do not use AI outputs that visibly reproduce identifiable third-party works without separate clearance.

What I would flag to clients

Three points I make sure clients understand before relying on AI-generated material in a commercial workflow.

1. The copyright status of pure AI outputs is currently 'no copyright' under USCO guidance and Thaler. A competitor can copy the output and sell it without infringement liability to the licensee. If exclusivity matters for the business model, the licensee needs to add a human-authorship layer or rely on a different IP regime.
2. Disclosure on registration is mandatory. The licensee should not file a registration that claims more than the law currently allows. Mis-statement of authorship in a registration is a federal-level problem.
3. The law is moving. The Copyright Office may issue further guidance. The Ninth Circuit and Federal Circuit may issue rulings in the pending matters. The drafting position should be conservative enough to survive a tightening of the rules, with mechanics that allow the parties to capture additional protection if the rules loosen.

The USCO's posture on prompts

One area where I would not stake a client's exclusivity: the prompts themselves. The Office has taken the position, repeated in the Part 2 report, that the prompts are like instructions to an artist, and that the artist (the model) is the one doing the creative work that ultimately produces the output. Commentators have argued that highly developed prompts are themselves authorship, and there is a plausible academic case for that view. The current Office position is the opposite, and counsel registering work should not assume the academic case will prevail in the near term.

The summary, for counsel managing an AI-assisted creative pipeline: register the human-authored elements, disclose the AI contribution, document the process, assign whatever copyright exists, and do not rely on AI-generated material alone for exclusivity. The drafting should anticipate that the law will tighten before it loosens, and the agreements should be drafted to survive either path.

Practical clauses I include in current drafts

A few specific drafting moves that I include in 2026 agreements where AI-assisted output is in scope.

• A representation that the party producing the deliverable will maintain a record of which elements were AI-generated and which were human-authored, and will provide that record on request.
• A clause that limits the licensee's right to register copyright in any output without the licensor's prior consent, where the licensor has a continuing interest in the source materials.
• A clause that addresses what happens if a third party claims that an output infringes that party's copyright. The default is to route the claim through the indemnity discussed in my SaaS IP indemnity memo; the explicit language reduces argument later.
• A clause that preserves trade-secret status by classifying the prompt library, the training pipeline, and the model outputs as confidential information unless and until publicly disclosed.

None of this resolves the underlying copyright question. The drafting just makes sure the parties have allocated the risk explicitly rather than leaving it to a court to decide in 2027 or 2028 based on whichever theory has carried the day in the appellate decisions by then. Counsel who treat output rights as a footnote in an AI deal are the ones who will be calling later about whether their client owns what they thought they owned.

Sergei Tokmakov, Esq., CA Bar #279869. This memo is attorney commentary on legal questions and is not legal advice. Reading it does not create an attorney-client relationship. Past matter outcomes depend on facts and the responding party; nothing here is a prediction of result.