What AI Vendors Guarantee: Indemnities and Disclaimers

AI vendor guarantees are mostly copyright indemnities with conditions, paired with accuracy disclaimers. What three published terms say.

Guide: AI Tools and IndustryLast verified: October 20265 min read

An AI "warranty" from a software vendor is usually two separate things: a disclaimer that outputs are provided "as is," and a contractual promise to defend the customer if a third party claims the output infringes intellectual property. Microsoft (September 2023), Google Cloud (October 2023), and Anthropic's commercial terms (effective June 17, 2025) each make a version of the second promise, each with conditions, and none promises that an answer is correct.

Key takeaways

  • The common vendor guarantee is an intellectual property indemnity, not a promise of accuracy or fitness for a purpose.
  • Each indemnity depends on conditions such as using built-in safety filters, holding rights to the input, and not intending to infringe.
  • Anthropic's commercial terms disclaim accuracy and tell customers to check factual claims in outputs before relying on them.
  • Terms change, so the date of the document being read matters as much as its content.
  • Only the contract the customer actually signed or accepted controls; announcements and summaries do not.

Two promises that are easy to confuse

When a buyer asks what a vendor guarantees about an AI product, the answer sits in two places in the contract. The first is the warranty disclaimer. Anthropic's commercial terms, for example, state that the services and outputs are provided "AS IS" and "AS AVAILABLE," that Anthropic does not warrant outputs to be accurate, complete, or error-free, and that factual assertions in outputs should be independently checked before they are relied upon. A customer must pass that notice on to its own users.

The second is the indemnity. An indemnity is a promise to defend the customer and pay certain costs if a third party brings a specific kind of claim. In the AI context the claim is almost always that a model's output, or the data used to train the model, infringes copyright or another intellectual property right. An indemnity says nothing about whether an output is true, safe, or useful. A customer can be fully covered against a copyright claim and still be left with a wrong answer.

What three vendors have published

Three documents show how the commitment is built. They are summarized as written on the dates shown, not as a ranking, and each vendor has several products with different terms.

Vendor and documentWhat it coversStated conditions or exclusions
Microsoft, Copilot Copyright Commitment, announced September 7, 2023; expanded and renamed Customer Copyright Commitment on November 15, 2023Defense and payment of adverse judgments or settlements for commercial customers sued over copyright in use of listed Copilot services or their output; the November expansion added Azure OpenAI ServiceCustomer must use the built-in guardrails and content filters, must not try to generate infringing material, and must have the rights to any input provided
Google Cloud, generative AI indemnification, October 13, 2023Two indemnities: claims over training data used to create the models, and claims that generated output infringes; the output indemnity applied to a listed set of productsOutput coverage applies if the customer did not intentionally create or use output to infringe, and is following responsible AI practices
Anthropic, Commercial Terms of Service, version effective June 17, 2025Defense and indemnity against third-party claims that the customer's paid use of the services, including outputs, infringes intellectual propertyExcludes claims arising from the customer's modifications to outputs, combination with third-party technology, the customer's own inputs, and use the customer knows or should know infringes; patents in outputs and trademark use of outputs are also excluded

Reading the conditions

The conditions matter more than the headline. A customer who turns off a content filter, feeds in material it has no license to use, or edits an output so that the edit creates the infringement may fall outside the promise. Google's announcement tied the output indemnity to a customer following responsible AI practices and not intentionally infringing, and Microsoft tied its commitment to use of the required safeguards.

Scope is the other limit. Each promise applies to named products and, in most cases, to paid or commercial plans. Free consumer versions are generally governed by separate terms, and features released as previews or betas are often treated differently from general availability products. The practical step for a buyer is to find the exact product name in the vendor's current terms and read the indemnity clause for that product.

A checklist for reading an AI vendor guarantee

The following checklist is an illustrative reading aid, not a legal test.

  • Identify the document: name, version, and effective date, and whether the customer accepted that version.
  • Find the disclaimer of warranties and note whether accuracy is excluded.
  • Find the indemnity and note whether it covers outputs, training data, or both.
  • List the conditions that must be met, including required filters, rights to inputs, and no intent to infringe.
  • List the exclusions, such as modified outputs, combined technology, patents, and trademarks.
  • Check which products and plan types the indemnity names.
  • Check any cap on liability elsewhere in the agreement, since an indemnity can still be limited by it.

Where consumer warranty law fits

None of the documents above is a consumer product warranty. They are business contracts between a vendor and a customer. A shopper buying a physical product with AI features is covered by a different body of rules, and a short explainer on that subject is linked in the related reading below. Neither the vendor indemnities nor consumer warranty rules guarantee that a particular AI output is correct, which is why the accuracy disclaimers remain standard.

This article is general information, not legal advice. A lawyer who has read the actual agreement is the right source for questions about coverage in a specific dispute.

FAQ

Does an AI indemnity mean the vendor guarantees correct answers?

No. The indemnities reviewed here address intellectual property claims by third parties. The Anthropic commercial terms separately disclaim accuracy and ask customers to verify factual claims in outputs.

Do the indemnities apply to free versions?

The documents reviewed refer to commercial customers, paid use, or named products. Free or consumer versions are typically covered by different terms, so the product's own terms must be checked.

Can the terms change after a customer signs up?

Often yes. Vendors publish new versions with new effective dates, and Microsoft's commitment itself was expanded two months after it was announced. The version the customer accepted is the one to keep a copy of.

Why does a vendor require using its content filters?

Microsoft says its commitment applies as long as the customer used the guardrails and content filters built into the products. The filters reduce the chance of infringing output, and the condition ties the vendor's risk to their use.

Last verified: October 2026

Sources

  • Copilot Copyright Commitment: Ensuring customers can use Copilots without worrying about copyright claims, Microsoft On the Issues, September 7, 2023 (updated November 15, 2023)
  • Protecting customers with generative AI indemnification, Google Cloud Blog, October 13, 2023
  • Commercial Terms of Service, Anthropic, version effective June 17, 2025

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics