Your AI Vendor Says It Will Defend You If the Output Infringes. What Do You Have to Do to Keep That Promise?

Your product runs on a model from one of the large AI providers. An enterprise customer's contract team asks a direct question: if something your product generates is accused of infringing someone's copyright, who defends the claim? You remember reading that your provider offers protection for exactly that. Before you repeat the promise to your customer, it is worth knowing what the provider's promise covers and what it asks of you.

This post is for founders and operations leads at software companies that build on third-party AI models. It describes how these promises are commonly structured, using four providers' published terms as examples. It is contract practice, not a statement of law, and it does not evaluate any provider's terms or say whether a clause is enforceable.

The short answer

Several major AI providers' business terms now include a promise to defend customers against certain third-party intellectual property claims about AI output. Each promise comes with conditions. Some apply only to paid services or to listed products. The terms address matters such as modifications or combinations, knowledge of infringement, and trademark claims, but their wording and scope differ. Some depend on safety features being switched on, and one provider's documentation says the customer will need to demonstrate compliance when it asks for a defense.

Three points follow for a company that builds on these models:

1. Find the document that actually contains the promise. It is not always the main agreement.
2. Read the conditions as things you have to do and be able to prove, not as fine print.
3. Compare the promise you receive with the promise you make to your own customers. A gap may leave you responsible for a claim your provider does not cover.

Where the promise lives

The output promise is often in a different document from the main contract, and the main contract can point the other way.

• OpenAI. The Services Agreement defines "Customer Content" as "the Input and the Output." Its general indemnity covers claims that the Services infringe, but excludes claims arising from Customer Content, and the customer indemnifies OpenAI for claims related to Customer Content.¹ The output promise is in OpenAI's separate Service Terms, which say OpenAI's indemnification obligations to API customers include claims that the customer's "use or distribution of Output infringes a third party's intellectual property right."² The Services Agreement says that Service Terms indemnity is "not subject to any liability cap."¹
• Microsoft. The "Customer Copyright Commitment" is in the Microsoft Product Terms and works through the customer's volume licensing agreement.³ Additional required mitigations are published on a separate documentation page.⁴
• Google Cloud. The promise is in the Service Specific Terms, as "Additional Google Indemnification Obligations" for Generated Output.⁵ It covers services listed on a separate page, where Google does not provide the service at no charge.⁵⁶
• Anthropic. The promise is in the Commercial Terms of Service and covers claims about the customer's "paid use of the Services ... in accordance with these Terms or Outputs generated through such authorized use."⁷

Before relying on any of these, confirm which version of which document governs your account. A negotiated enterprise agreement can differ from the published terms. Our post on API dependency and vendor terms covers how online terms bind a company that clicks through them.

The conditions that recur

Reading the four sets of terms side by side, the same kinds of conditions appear. The wording differs, and the differences matter.

Safety features left on. Microsoft's commitment requires that the customer not have "disabled, evaded, disrupted, or interfered with the content filters, restrictions in Metaprompts, or other safety systems."³ Google's promise does not apply where the customer disregards, disables, modifies or circumvents "source citations, filters, instructions, or other tools."⁵ OpenAI's does not apply where the customer or its end users "disabled, ignored, or did not use any relevant citation, filtering or safety features or restrictions."²

Required mitigations, and proof. For some Microsoft services, the customer must also implement published mitigations, including a metaprompt (a system message) directing the model to prevent copyright infringement and evaluations designed to detect third-party content in output. The page says the report of results "must be retained by the customer and provided to Microsoft in the event of a claim," and that a customer tendering a claim "will be required to demonstrate compliance."⁴

Knowledge. All four address the customer's knowledge, but not in identical terms. OpenAI refers to knowledge that output was infringing or likely to infringe; Microsoft to modifying, using or distributing output in a manner the customer knows or should know is likely to infringe or misappropriate; Google to creating or using output the customer knew or should have known was likely infringing; and Anthropic to use the customer knows or reasonably should know violates or infringes third-party rights.²³⁵⁸ Google also excludes use of output after the customer receives notice of an infringement claim from the rightsholder.⁵

Modified or combined output. Google's promise covers "unmodified" output.⁵ OpenAI's excludes claims where output was "modified, transformed, or used in combination with products or services not provided by or on behalf of OpenAI."² Anthropic's excludes claims arising from the customer's modifications and from combining the services or outputs with technology or content Anthropic did not provide.⁸ These provisions should be checked against the particular claim; they should not be treated as a single rule that every modification or combination defeats every claim for coverage.

Rights to what you put in. Microsoft conditions its promise on the customer having sufficient rights to use the input, including data used to customize the model.³ Google excludes cases where the customer lacks the necessary rights to data used to customize or retrain a model, or to customize the output.⁵ OpenAI excludes cases where the customer or its end users lacked the right to use the input or fine-tuning files.² Anthropic excludes claims arising from inputs or other data the customer provided.⁸

Trademark claims. All four exclude claims that the output, as used in trade or commerce, violates trademark rights.²³⁵⁸ Anthropic also excludes claims based on practicing a patented invention contained in an output.⁸

Your end users. OpenAI's exclusions refer to what the customer "or Customer's End Users" knew or did.² If your customers' users can switch off a safety feature or paste in material they have no right to use, their conduct may affect your own coverage.

How a claim is handled

Indemnities usually include a procedure, and missing a step can cost coverage. Anthropic's terms are an example: the indemnified party "must promptly notify" the other and "reasonably cooperate in the defense"; the indemnifying party "will retain the right to control the defense," including the choice of counsel; and the obligations are excused if late notice or a failure to cooperate "materially prejudices the defense."⁹ If a customer sends you a claim letter about output, the clock under your provider's terms may already be running.

The terms change

These promises are revised. Microsoft's mitigations page gives customers, in most cases, six months from publication to implement newly required mitigations; for new services, features or models, new requirements take effect at or after launch.⁴ Anthropic's update provision generally requires 30 days' notice, permits immediate changes for legal or regulatory reasons, and states that changes are not retroactive.¹⁰ OpenAI's Service Terms page showed an update dated October 16, 2026, and Google's Service Specific Terms were last modified October 12, 2026, when read for this post.²⁵ Decide who at your company watches for changes, and how a change reaches the people who answer customer questionnaires. Our post on switching AI vendors covers what else to track when you depend on a provider.

The gap between what you receive and what you promise

This is the part a founder controls. Your enterprise customer may ask you for an indemnity against claims that your product's output infringes. If you agree, compare your promise with your provider's, clause by clause:

• Scope. Does your promise cover output your product transforms, edits or combines with your own features and data? Several providers' promises exclude modified or combined output, and a software product often does both.
• Conditions. Do you require your customers to keep the same safety features on, and to have the rights to what they put in? If your provider's coverage depends on your end users' conduct, your contract with your customer may need to say so.
• Exclusions. Do you exclude trademark claims, and uses your customer knew were likely to infringe, as your provider does?
• Procedure. Will your customer notify you quickly enough for you to notify your provider on time?
• Caps. How does your promise sit against your own liability cap? Our post on liability caps covers how caps and carve-outs interact.

Where your promise is broader than the one you receive, you may assume obligations your provider's indemnity does not cover. Your ultimate exposure also depends on your customer agreement's limitations and exclusions and any other applicable protection. That can be a sensible commercial choice, but it should be a deliberate one. Our posts on AI product customer contract terms and on enterprise SaaS contracts and DPAs cover the customer side of the same negotiation, and our post on your AI provider's acceptable use policy covers another place where upstream terms reach your customers.

Questions to settle before you repeat the promise

These are questions to take to your contracts, not conclusions about any particular agreement:

1. Which document contains the output promise, which version governs our account, and is the service we use covered on the plan we pay for?
2. Which safety features and mitigations does coverage depend on, are they switched on, and could we show it on the date of a claim?
3. Does our product modify, transform or combine output in a way the exclusions describe?
4. Do our customer terms pass the relevant conditions to our customers and their users?
5. What exactly do we promise our customers, and where is it broader than what we receive?
6. Who receives claim letters, and how fast can a notice reach our provider?
7. Who watches the provider's terms for changes?

Our AI vendor contract questions cover the same ground from the buyer's side, and our post on AI-related questions in investor diligence explains why investors ask about indemnities.

What this post does not cover

This post does not address who owns AI output or whether it can be protected by copyright; see our post on who owns AI-generated code and content. It does not say whether any output infringes, whether any provider's clause is enforceable, or how a provider would respond to a particular claim. It does not cover insurance. Providers other than the four named here may structure their promises differently, and a negotiated agreement can differ from published terms.

How we can help

Accord & Shield Legal reviews and drafts SaaS agreements and customer contract terms for technology companies in Arizona, California and Texas. To better understand the scope of your matter, you can schedule an initial consultation. The consultation is evaluation-only; no legal advice is provided. Representation begins only after an engagement agreement is signed and any required retainer is received. Please do not send confidential or sensitive documents before we confirm that we can assist.

---

This article is general information, not legal advice. Reading it does not create an attorney-client relationship. Vendor terms change, and whether a provider's promise applies to your company depends on your agreement and your facts; consult an attorney about your situation.

Book an Initial Consultation

Schedule an initial consultation to discuss the scope of an AI vendor or customer contract question about output indemnity. Bring a short, nonconfidential description of the situation and any customer deadline.

Please do not send contracts, documents or confidential information until the firm has agreed in writing to represent you.

Schedule a consultation Call (623) 239-2682