What Should My Startup Ask Its AI Vendor About Training Data and Indemnities?

An indemnity is your AI vendor's promise to step in and pay if someone sues you over the model: the defense, and in many cases the damages. Every major provider now offers one, so the promise itself tells you almost nothing. What matters is three questions. Which lawsuits does it actually cover? What do you have to have done, and be able to prove, to keep it? And does it apply to the plan and the models you actually use?

The email that starts this

It usually arrives from a customer, not a plaintiff. Procurement sends over their standard AI questionnaire, or their lawyer adds an AI clause to your contract, and now you are promising people who pay you that your product's AI will not get them sued. Your first thought is the right one: my model vendor covers this. Sometimes it does. Whether it covers you depends on three questions most founders have never had a reason to ask.

One: which lawsuits does it actually cover?

Picture the two complaints that could land on your desk. In the first, a photographer says the image your product just generated for a customer copies her work. In the second, an author says the model itself should never have been built, because it was trained on his books without permission.

Those are different lawsuits, and vendors treat them differently. A promise to stand behind the output covers the first. It does not necessarily reach the second. Among the big four, OpenAI stands behind the output your use produces. Anthropic's promise expressly reaches the training data too. Google makes them two separate promises with different fine print. The marketing sentence, "we indemnify our customers," sounds identical everywhere. The coverage is not.

Two: what do you have to do to keep it?

Think of the conditions the way you think of an insurance policy that pays only if the smoke detectors were on. Every one of these promises comes with settings you have to leave in place: keep the safety filters on, do not strip out the citation tools, and use data you actually have rights to when you fine-tune.

Microsoft goes furthest. To keep its protection for Azure OpenAI, you must have tested your product for third-party content before any trouble starts, and "the report of results and mitigations must be retained by the customer and provided to Microsoft in the event of a claim." In plain terms, when you ask Microsoft to take over a lawsuit, the first thing it asks you for is a report you were supposed to have written months earlier. Nobody writes that report the day the claim arrives.

So the real question is not whether you are covered on paper. It is whether the person who set up your API calls eight months ago left the filters on, and whether anyone wrote down that they did.

Three: does it cover what you actually run?

These protections generally apply to paid plans and to models the vendor has fully released. Free tiers, betas and previews are carved out. That matters because preview is exactly where your team tries the newest model first, which means you can walk out of coverage by upgrading, and nothing about that moment will feel like a legal decision.

Two more things to check while you are in the document. First, anything you feed the model is your responsibility, not the vendor's. If your fine-tuning data or your uploaded documents cause the problem, the promise does not apply, so you need to know where your own data came from. An acquirer will ask the same question someday. Second, look for a dollar ceiling. Some vendors cap what they will ever pay at an amount tied to your fees, and if you pay the vendor a few thousand a month, a cap like that can be a small fraction of what defending a lawsuit costs. The ceiling usually lives in a different section than the promise does.

The one page that answers everyone

Keep a single page per AI vendor: the plan you are on, the models in production, which safety settings are enabled, the date you last tested your outputs and where that report is saved, and whether the promise has a dollar ceiling. Give one person the job of updating it when the stack changes.

That page takes an afternoon, and it answers a customer's questionnaire, an investor's diligence list, and a demand letter. Few afternoons return more.

Where this is still unsettled

Whether training a model on copyrighted material is lawful is an open question, and courts in different countries are answering it differently. That is exactly why this language keeps moving. Terms you read a year ago may not say what they said. Check the effective date on your vendor's page before you rely on your memory of it.

Frequently asked questions

Do the major AI vendors indemnify their customers for copyright claims?

Yes, all of the largest ones publish some form of intellectual property indemnity for paid commercial use. They differ substantially in scope. Some cover only the output the model generated, some also cover the data the model was trained on, and all of them attach conditions to coverage.

What is the difference between an output indemnity and a training data indemnity?

An output indemnity responds to a claim that a specific generation infringed someone's rights. A training data indemnity responds to a claim that the vendor should not have trained on the underlying material in the first place. They answer different questions, and a vendor can offer one without the other. Google, for example, addresses them in two separate subsections and gives them different exclusions.

Does the indemnity still apply if I use a free tier or a preview model?

Usually not. Vendor indemnities are commonly limited to paid use and to generally available services on a published list, with beta, preview and pre-release offerings expressly excluded. This is a common gap because early-stage teams prototype on exactly those tiers.

What do I have to do to keep my AI vendor's indemnity?

At minimum, leave the safety and filtering features on, do not circumvent citation tools, and make sure you have the rights to any data you use for fine-tuning or retrieval. Microsoft goes further: it requires customers to have run evaluations designed to detect third-party content and to have kept the report, because Microsoft asks for it when you ask it to cover a claim. Treat the conditions as an operational checklist, not fine print.

Can I pass my vendor's AI indemnity through to my customers?

Not automatically, and not for free. You can only give what you have, and the exclusions in your upstream agreement will rarely line up with the ones in your customer contract. The difference between the two sits with you, which is a reason to compare them deliberately rather than accept an AI indemnity clause from a customer's template without reading your vendor terms next to it.