Copyright &
Generative AI
Two practical essays on the rules that matter now—what human authorship actually requires, and how to structure ownership and training rights in commercial agreements.
What Counts as Human Authorship When You Use AI
Where the U.S. Copyright Office draws the line on human authorship in AI-assisted work.
In short: the prompt does not count. Selection, arrangement, and substantive editing of AI output can count. The burden is on the applicant to describe their human contribution clearly enough that the examiner can find it.
The U.S. Copyright Office has now reviewed and ruled on more than a hundred applications involving AI-assisted works. The practical bar for human authorship is no longer a matter of speculation—we can see, application by application, where the line falls and what kind of human contribution clears it.
The current U.S. position.
The Copyright Office set out its framework in two documents. First, the March 2023 Statement of Policy: Works Containing Material Generated by Artificial Intelligence, which clarified that the human-authorship requirement applies to works incorporating AI-generated material and that applicants must disclose AI use. Second, the January 2025 Report on Copyright and Artificial Intelligence, Part 2: Copyrightability, which restated and refined the position: the existing copyright framework can resolve AI-authorship questions without legislative change, and the dispositive question is whether a human exercised creative control over the expressive elements of the work.
That last phrase—creative control over expressive elements—is doing all the work. It is what separates a prompt (not enough) from a curated arrangement of AI outputs (potentially enough). And it is what the Office is testing every time it reviews an application.
Why the prompt does not get you there.
The Office's reasoning, drawn most directly from its review of Kris Kashtanova's Zarya of the Dawn and Jason Allen's Théâtre D'opéra Spatial, is that prompting is more like instruction than authorship. The human formulates a request; the model generates the expressive output; the human cannot reliably predict or determine what the output will look like. Even with extensive iterative prompting, the relationship between the prompt and the output is too attenuated to constitute the kind of creative control copyright protects.
Allen made this argument as forcefully as anyone has. He told the Office he ran more than 600 prompt iterations and edited the final image in Photoshop to get the result he wanted. The Office rejected the application as to the Midjourney-generated portions. The Photoshop edits, had he properly disclaimed the AI material and described the edits with specificity, might have been registrable on their own. He did not, and the application failed.
What does count.
The most direct guidance comes from the Zarya of the Dawn registration, which the Office issued in partial form. Kashtanova received copyright on the text she wrote and on the selection, coordination, and arrangement of the AI-generated images within the comic. The images themselves were disclaimed.
That selection-coordination-arrangement framework is not new—it comes from compilation copyright doctrine and Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991)—but it gives applicants a usable template. A human-authored selection of which AI outputs to use, in which order, with which interstitial human-authored material, can constitute a protectable compilation.
Substantive editing of AI output is the other clear path. The Office has explained that if a human takes an AI-generated image and meaningfully modifies it—not by changing a color or running a filter, but by adding, removing, or rearranging expressive elements—those modifications can be registrable. The protectable claim is to the human's contribution, not to the underlying AI material.
How to apply this in practice.
Three tests, applied in order.
First, what did the human do that the model did not? If the answer is "wrote a detailed prompt," the human contribution is not registrable. If the answer is "selected from many generated outputs based on aesthetic judgment, arranged them, and added human-authored text," the human contribution likely is. The test looks at the contribution, not the difficulty.
Second, can the human contribution be described to a third party with specificity? The application requires the applicant to identify AI material and disclaim it, and to describe the human-authored material the applicant is claiming. If you cannot point to specific selections, arrangements, or edits and explain why each was a creative choice, the application is likely to be rejected for vagueness even if the underlying contribution would otherwise qualify.
Third, would the work look meaningfully different if a different human had made the same creative choices? This is the Bleistein v. Donaldson Lithographing Co., 188 U.S. 239 (1903), test—copyright protects creative expression that bears the mark of a particular human's choices, not generic output. If the answer is yes, you have something worth claiming. If no, the contribution is likely below the threshold of originality.
A note on registration practice.
Most denials trace back to applications that overclaim rather than to the underlying work itself. Applicants who clearly identify and disclaim the AI material and limit their claim to specific human contributions—text, selection, arrangement, editing—are far more likely to receive a registration, even a partial one.
Documentation matters and should be created while the work is being made, not reconstructed later. Save prompts and iterations. Save intermediate edits. Keep notes on the creative choices behind selection and arrangement. When the application asks the applicant to describe the human contribution, the answer should be specific and grounded in contemporaneous record-keeping.
Where this is heading.
The Office has now published all three parts of its AI series. Part 3, released in May 2025, addresses AI training and infringement directly. Its most consequential finding for the copyrightability question is what it rejects: the analogy between AI training and human learning. The Office concluded that humans retain only imperfect impressions of works they have experienced, filtered through their own personalities, histories, and memories. Generative AI training involves the creation of perfect copies, analyzed nearly instantaneously, producing a model that creates at superhuman speed and scale. The two processes are different in kind, not degree—and the Office said so directly.
For now, the practical guidance is unchanged. The prompt does not count. The selection, arrangement, and substantive editing of AI output can count, if the human contribution is real and the application describes it clearly. The framework rewards specificity, and rewards practitioners who treat the human-authorship requirement as a creative constraint to work with, not an obstacle to argue around.
Licensing AI-Assisted Work: A Drafting Guide for SaaS Companies
Three drafting moves that decide ownership, training rights, and provenance in commercial AI-assisted agreements.
The standard MSA-plus-DPA-plus-Order-Form structure does not handle AI-assisted output well. Three explicit drafting moves fix the most common failure modes.
Most SaaS commercial contracts in circulation today were drafted before generative AI became a routine part of customer workflows. The standard MSA-plus-DPA-plus-Order-Form structure handles software licensing, data processing, and confidentiality competently. It does not handle AI-assisted output well, because the questions it needs to answer—who owns what the model produces, what the vendor may do with the customer's data to train future models, and how the customer can later prove which work was AI-assisted—were not the questions the templates were built to answer.
1. Allocate output ownership explicitly
The Problem. Most SaaS contracts include a vendor IP clause and a customer data clause. Neither anticipates AI-generated output, which is neither vendor IP nor customer data in the traditional sense. Three potential claimants, no clear allocation.
The Drafting Move. Address output ownership as its own contractual term, separate from vendor IP and customer data. Where the customer is to own outputs, include a present assignment ("Vendor hereby assigns to Customer all right, title, and interest in and to AI-Generated Outputs"), not a future-tense covenant. The default in most model-vendor terms—OpenAI, Anthropic, Microsoft, Google—is that the end user owns outputs subject to use restrictions. The SaaS provider's contract with its customer should reflect those upstream constraints, not paper over them.
2. Separate training rights from output rights
The Problem. A license to use a tool is not a license to train on the data the tool consumes. Whether the standard "to provide the Services" license covers using customer prompts or outputs to train future models is genuinely unsettled, and most customers do not realize they may have consented to model-training rights by signing a routine SaaS agreement.
The Drafting Move. Address training-data use as an explicit, opt-in term. The clause should specify whether the vendor may use customer prompts, customer-supplied context, and customer-generated outputs to train, fine-tune, or evaluate models, and the answer should be different for each category. For most enterprise customers, the right default is no training-data use without separate written consent.
The stakes here are not hypothetical. In Thomson Reuters v. ROSS Intelligence, decided in February 2025, the District of Delaware held that ROSS's use of Westlaw's headnotes to train a competing AI legal research tool was not fair use—precisely because the training purpose was the same as the original product's purpose. The absence of explicit training rights is a litigation surface, not a neutral gap.
The U.S. Copyright Office's Third Report on AI, released in May 2025, found that when a model is trained on a narrow or small dataset, the model's weights are more likely to retain—or memorize—protectable expression from the works at issue. The risk of training-data infringement is not uniform; it scales with how targeted the training is. The contract should reflect that distinction.
3. Build in a provenance representation
The Problem. When a customer later wants to register, enforce, license, or audit AI-assisted work, the customer needs to know which outputs were AI-generated, which were AI-assisted, and which were entirely human-authored. Standard SaaS contracts do not require the vendor to provide that information.
The Drafting Move. Include a provenance representation in the vendor's commitments. The clause should obligate the vendor to surface or make available, on request, sufficient information for the customer to identify which outputs in a given workflow were AI-generated, which model was used, and what context or training data was incorporated. Pair the representation with a customer obligation to maintain its own records of how AI-assisted output is used downstream. Provenance is a chain; it breaks at the weakest link.
Sequencing across the standard SaaS contract stack
- MSA: output ownership; training-data use of non-personal customer information; provenance representation; model-vendor flow-through language.
- DPA: training-data use of personal data, with the appropriate lawful basis or contractual prohibition; sub-processor obligations for any third-party model provider.
- Order Form: specific opt-ins or opt-outs that depart from MSA defaults, signed at the time the customer makes that decision rather than negotiated into the MSA.
A note for SaaS founders
The three moves above are not just risk-mitigation for legal departments—they are commercial differentiators. A vendor whose contract handles output ownership, training-data use, and provenance with specificity and confidence will close deals faster than a vendor whose contract is silent or evasive on the same questions. The customers are reading these clauses. The clauses should be readable.
And the courts are watching, too. In Kadrey v. Meta, Judge Chhabria closed his fair use analysis with a pointed observation: if AI products are capable of generating the income they do, the companies developing them will figure out a way to compensate copyright holders for use of their materials. That is not a ruling. But it is a signal—and founders who address it contractually are better positioned than those who leave it to litigation.
This post is for general informational purposes only and does not constitute legal advice or establish an attorney-client relationship. For advice specific to your situation, please book a fit call or contact info@ambartlaw.com.