Copyright of AI-
Generated Works

Does human prompting count as authorship?

Generally no—and the length or sophistication of the prompt does not change that. The U.S. Copyright Office's position, stated clearly in its March 2023 Rule and reinforced in the January 2025 report on copyrightability, is that a prompt directs a machine. It does not, by itself, express the traditional elements of authorship—the way a human selecting brushstrokes, drafting sentences, or making compositional choices does. So the question is not how good your prompt was. The question is what you did with what came out.

The analysis shifts meaningfully when the human curates, edits, or rearranges AI output in a way that reflects genuine creative choice. Those contributions can be protectable—separately from the underlying AI-generated material. I made this argument in the Harvard International Law Journal: we accept that Jeff Koons is the author, not his dozens of studio assistants, many of whom are sophisticated designers and engineers. Why is AI treated differently? That question does not yet have a clean answer. But it is the right one to keep asking.

What does the U.S. Copyright Office currently say about AI-generated works?

The Office has been consistent, even as the technology has moved fast. In its March 2023 Rule on Works Containing Material Generated by Artificial Intelligence, it stated that copyright protects "only material that is the product of human creativity." The January 2025 report—which was widely anticipated and, frankly, widely misread—did not signal a departure. Far from it. The Office maintained that existing law can resolve questions of copyrightability without new legislation, and that most prompt engineering will not suffice for registration.

What the Office does protect: human-authored portions of AI-assisted works, where the applicant properly identifies and disclaims the AI-generated material. Several high-profile decisions—the partial registration of Zarya of the Dawn, the refusal of the Midjourney-generated imagery within it, and the refusal in Thaler v. Perlmutter—shape the current practical bar. The Copyright Office completed its three-part report series on AI and copyright in 2025. All three installments are worth reading closely—the doctrine has not shifted in favor of AI-generated works, and the Office has been consistent throughout.

How should a SaaS company license AI-assisted derivative works?

Three drafting moves matter most—and I see all three missed regularly in SaaS agreements I review.

First, allocate ownership of outputs clearly: between the SaaS provider, the customer, and the model vendor. Do not assume the default favors whoever holds the account. It often does not. Read the model vendor's terms of service carefully—several have changed materially in the past 18 months.

Second, address training rights separately from output rights. A license to use a tool is not a license to train on customer data. These are different permissions, and bundling them creates exposure on both sides.

Third, build in a provenance representation. The customer needs to be able to track—at the time of the agreement and later—which outputs were AI-assisted, which were human-authored, and which contain third-party inputs. That becomes decision-critical when they want to register, enforce, or sublicense the work. If you cannot answer those questions now, a litigation counterparty will answer them for you.

What does the Anthropic code-leak matter teach about training-data provenance?

The core lesson is structural—and it applies well beyond Anthropic. Even sophisticated model developers can end up with training pipelines where the downstream legal status of an artifact depends on upstream collection decisions made months or years earlier. That is not a hypothetical. It is the fact pattern at the center of Bartz v. Anthropic, Kadrey v. Meta, and Thomson Reuters v. Ross Intelligence. All three are, at their core, provenance problems.

For any company building on foundation models, the takeaway is practical: treat provenance as a first-class contractual and engineering concern, not a footnote. Ask your vendors to document training-data sources, retention policies, and any known infringement claims. Document your own inputs the same way. When provenance is knowable after the fact, disputes resolve faster and settlements cost less. When it is not, you are negotiating blind.

Are style transfers copyrightable if the style reference is copyrighted?

Style itself is generally not protectable under U.S. copyright law—the statute protects specific expression, not an artistic approach or aesthetic. That principle holds when the output is an AI-generated image "in the style of" a named artist. But there are two important caveats.

First, the output can still infringe if it reproduces protectable elements of a specific work—composition, distinctive characters, recognizable scenes. "In the style of" is not a safe harbor if the result captures protected expression rather than just aesthetic feel. Second, the risk profile shifts when the style reference was used in training without authorization. That is a live and unresolved question across the generative-AI copyright litigation docket, including Concord Music Group v. Anthropic. So: style transfer is not categorically safe. The analysis is fact-specific every time, and the litigation is still writing the rules.

Who owns the output when an employee uses a company-issued AI tool?

Start with the employment agreement and any work-for-hire or IP-assignment language. If the output qualifies as a work of authorship at all—which, as discussed above, depends on the level of human creative contribution—those clauses typically assign the human-authored components to the employer. So far, so familiar.

What employment documents rarely address is AI-assisted output specifically: whether a prompt counts as a creative contribution, whether the model vendor's terms impose a license on the employer independently of the employment relationship, and whether outputs generated outside the scope of employment still flow to the company. These are not edge cases. They come up in every AI-use policy review I do. Update your employment agreements and your AI-use policy together—treating them as one system removes most of the common disputes before they start.

Can AI-generated work be registered with the Copyright Office?

Yes, with disclaimers—and the how matters as much as the whether. Applicants must identify AI-generated material in the application and limit their copyright claim to the human-authored portions: the selection, coordination, or arrangement of AI output; human-written text interleaved with AI-generated text; or substantial human editing that rose to the level of creative authorship. The Office has issued registrations on that basis.

Fully AI-generated works, without meaningful human creative contribution, are not registrable. And importantly, most denials trace back to applications that overclaim—not to the underlying work itself. Documentation matters: save your prompts, your drafts, and your reasoning for major creative choices while you are making them. Reconstructing that record after the fact, under examination, is harder than it sounds.