Your Product Uses AI. What Your Customer Contract Must Say
By Nadine Deeb, Esq. · Published July 31, 2026AI product terms may include a provision stating that, as between the parties, the customer owns output.
It reads like a concession. It is meant to close deals. The customer wants to know that what your product produces belongs to them, and you want to say yes.
Here is the part that is easy to miss. A contract can allocate rights between vendor and customer, but it cannot itself create copyright protection in material that lacks the required human authorship. The parties can agree not to assert rights against each other. That agreement is different from a warranty that copyright exists, is registrable, or will be enforceable against a third party.
Those are three separate promises. A clause that reads like all three, but only delivers the first, is the gap worth closing before your next enterprise deal.
What the Copyright Office Actually Says
The rule is human authorship. In the Copyright Office’s words, copyright protects only material that is the product of human creativity.
Applied to generative AI, the Office’s March 2023 registration guidance draws the line at prompting. Under that guidance, when a user supplies only a prompt and the system determines the expressive elements, the Office treats the generated material as lacking human authorship. The Office’s stated reasoning is that prompts function like instructions to a commissioned artist: they identify what the person wants depicted, but the machine decides how. That material has to be disclaimed on a registration application.
That is registration policy. It does not resolve every downstream infringement or ownership question, and it is not a final rule governing every dispute about who owns what.
The federal courts have addressed the outer edge. In March 2025 the D.C. Circuit held that the Copyright Act requires a work to be authored in the first instance by a human being, and affirmed the refusal to register a work submitted as authored solely by a machine.
[NEW] The Supreme Court declined to review that decision on March 2, 2026. A denial of review is not a ruling on the merits, but the D.C. Circuit’s holding stands.
It matters what that case did not decide. The court expressly declined to reach whether the Constitution independently requires human authorship, and it did not decide whether the person who built and operated the system might be the author, because that argument had been waived on the record below. The opinion also says the human authorship requirement does not prohibit copyrighting work made with the assistance of artificial intelligence. It should not be read as settling every AI-assisted workflow.
So the accurate statement is narrower than the one that circulates at conferences. Material generated from a prompt alone is treated as lacking human authorship. Where a human selects, arranges, or meaningfully modifies AI-generated material, that can support a copyright claim, and the protection extends to the human contribution rather than to the AI-generated material itself.
Whether any particular output includes protectable human authorship is, in the Office’s own words, necessarily a case-by-case inquiry. A standard ownership provision generally cannot determine that in advance.
A Concrete Illustration
In February 2023 the Copyright Office reviewed a registration for a graphic novel combining human-written text with images generated by Midjourney.
The Office concluded that the author’s text, and her selection, coordination, and arrangement of the written and visual elements, were protected. It treated the Midjourney images as not the product of human authorship. Because the existing registration did not disclaim the AI-generated content, the Office stated it intended to cancel the certificate and issue a new one covering only the material she created.
The scope of the registered claim was narrower than the original certificate reflected. That is a statement about the registration, not a determination that the registrant held no rights. Registration scope and ownership are different questions, and the Office was deciding the first.
The practical lesson still lands: paper can describe something more expansive than what is underneath it. That is worth remembering when your contract is the paper.
Does your customer agreement promise something about output your product cannot deliver?
We can read your current terms against the issues in this article and tell you which provisions are doing real work and which are not.
What the Ownership Clause Can and Cannot Do
The clause still does real work. It establishes that you will not assert rights in your customer’s output, which is often the customer’s underlying concern.
Depending on the product, governing law, and negotiated risk allocation, counsel may consider whether the clause should do the following. This is a non-exhaustive list of issues, not a formula:
Frame the provision as an allocation between the parties rather than as a transfer of ownership, so that what you are agreeing to is not asserting rights, not warranting that rights exist.
Consider whether to acknowledge that the legal status of machine-generated material is unsettled and that protection may depend on the customer’s own creative contribution.
Consider whether the agreement should avoid promising registrability, exclusivity, or enforceability against third parties, which are three separate promises.
Whether any of this is right for a particular agreement depends on the deal.
Accuracy and Product Disclosures
AI-enabled products can have limitations that matter to customers. Product communications, documentation, and contract terms should describe material limitations accurately and consistently.
AI product terms may address accuracy and other limitations of output. A product-specific explanation of known limitations may help align product communications, documentation, and contract terms with one another. Whether any particular disclaimer is enforceable depends on the governing law, the transaction, and the full contract.
The practical point is consistency. If the contract describes limitations that the marketing contradicts, the inconsistency is the thing that gets quoted back to you.
Allocating the Risk of an Incorrect Output
This is a risk allocation question rather than a copyright question.
Contract provisions that often address that allocation include the limitation of liability, the exclusion of consequential damages, and carve-outs from either. Whether a particular cap or exclusion holds up depends on the governing law, the contracting process, the parties, the product’s use, and any applicable statutory restrictions. That is a question for counsel reviewing your actual agreement, and nothing here should be read as an opinion on enforceability in any state.
As a business-risk observation rather than a legal one: a cap tied to one month of fees may have a materially different risk profile depending on the use case and the potential harm. It is worth pricing the risk you are actually taking.
Third-Party Claims Arising From Output
Separate from accuracy: what if an output resembles someone else’s protected work and the customer faces a claim?
Some customers may request indemnity for third-party intellectual-property claims associated with output. Whether to offer that protection, and on what conditions, requires product-specific and deal-specific analysis.
If an indemnity is offered, the scope, exclusions, customer obligations, and verification mechanisms should be deliberate and clear. Conditions a customer cannot satisfy, or that you cannot verify, create their own problems.
Rights to Use Customer Data
Terms are often silent here, or grant a permission broad enough that the customer’s own counsel will strike it.
Three questions tend to shape the clause. Is the permission directed at customer inputs, at outputs, or both? Is it on by default with an opt-out, or off by default with an opt-in? And does the customer have the authority to grant it?
That last question is the one founders tend not to anticipate. A customer’s ability to grant rights in inputs or outputs may depend on its agreements with its own users and on the nature of the data.
Whether a particular arrangement carries obligations under privacy law is a separate analysis. The answer may depend on the data involved, the parties’ roles, applicable privacy law, notices and consents, and the customer’s underlying agreements. That deserves its own treatment and is not resolved here.
Changes to the Underlying Model
Your product may depend on a model you do not control. Providers deprecate versions, adjust behavior, change pricing, and revise their own terms.
Depending on the product and the customer’s dependency on it, counsel may consider whether the agreement addresses what you are committing to, whether that is a model, a capability, or a service level; what notice is given before a material change in behavior; and what happens if an upstream provider’s change forces a change to your own terms.
Silence is not flexibility. It means the question arrives with no agreed answer.
Human Review and Acceptable Use
Where a product is not intended to be the sole basis for a consequential decision, the agreement and the user experience may both address the role of human review. Contract terms alone do not cure regulatory, product, consumer-protection, or tort exposure, and terms that assume oversight while marketing promises full automation create an inconsistency worth resolving.
An acceptable use provision can identify applications the business is not willing to support. That is partly risk management and partly a filter.
Before Your Next Enterprise Deal
Read your current terms against these questions:
Does the ownership provision imply something about copyright that the law may not supply? Does the accuracy language describe your actual product? Is the liability cap sized to the harm an incorrect output could cause in the use you are selling into? Does the agreement address rights in customer data, and can the customer grant what is being asked? Does it address changes to the underlying model?
These are cheaper to work through before they are embedded in signed agreements than afterward.
Frequently Asked Questions
No, and that overstatement causes real problems. Under the Copyright Office’s registration guidance, material generated in response to a prompt alone is treated as lacking human authorship. A work containing AI-generated material can still support a copyright claim where a human selects or arranges it creatively, or modifies it enough to meet the originality standard, and the protection extends to the human contribution. Whether a given output qualifies is a case-by-case inquiry.
Not necessarily. It can establish that you will not assert rights in the customer’s output, which is a real and useful allocation between the parties. The concern is if it is read as a warranty that protectable copyright exists, is registrable, or is enforceable against third parties. Those are different promises.
No. In March 2025 the D.C. Circuit affirmed the refusal to register a work submitted as authored solely by a machine, holding that the Copyright Act requires a work to be authored in the first instance by a human being. The court did not decide whether the person who built and operated the system could be the author. [NEW] The Supreme Court declined to review the decision on March 2, 2026.
When seeking U.S. copyright registration for a work containing AI-generated material, the Copyright Office directs applicants to disclose that material and describe the human author’s contribution. Material generated by AI that is more than de minimis should be excluded from the copyright claim. If an applicant discovers that a pending application or existing registration omitted relevant AI-generated material, the Office’s guidance describes procedures for correcting the record. The appropriate approach is fact-specific.
The Copyright Office’s guidance indicates that an applicant should correct a pending application or the public record for an existing registration when required disclosure was omitted. In a later infringement case, inaccurate application information can also create registration issues under the Copyright Act, but the consequences depend on the statutory requirements and the facts, including what the applicant knew. This is registration guidance, not a general business disclosure rule.
Some customers may request it. Whether to offer that protection, and on what conditions, requires product-specific and deal-specific analysis. If an indemnity is offered, the scope, exclusions, customer obligations, and verification mechanisms should be deliberate and clear.
Terms are generally easier to work through before they are embedded in signed agreements. Revising a provision already given to many customers can mean amendments or inconsistency across a contract base.
Current as of July 31, 2026. Copyright Office guidance and legal developments concerning AI may change.
Disclaimer: This article is for general informational purposes only and is not legal advice. It does not create an attorney-client relationship with Accord & Shield Legal, PLLC. Laws, regulations, Copyright Office practices, and the facts of AI-enabled products and customer relationships can change and may apply differently to a particular business, contract, product feature, data set, customer use case, or jurisdiction. Do not rely on this article as a substitute for advice about your specific agreement or circumstances. Consult qualified counsel before adopting or revising customer-facing terms.
If your product uses AI, counsel can help assess how your customer agreement, product disclosures, privacy practices, and risk allocation fit your particular offering.
Does Your Product Use AI?
If your product generates output for customers, the terms you ship with it decide what you have actually promised. We can read your current agreement against the issues in this article — ownership, disclosures, risk allocation, data rights, and model change — and tell you which provisions are doing real work.