Will an LLM provider train on your data?
The question that stalls more AI projects than any technical objection. Here is the actual answer, and the part vendor policy pages leave out.
Answered in short
5 things that decide this
- 01OpenAI, Anthropic and Google do not use API-tier requests to train their models by default. This is stated in each provider's API data usage policy, not something you have to negotiate.
- 02Consumer products are a different agreement. ChatGPT's free and Plus tiers, and Gemini's consumer app, may use your conversations for training unless you turn that setting off yourself.
- 03Training is the wrong thing to worry about first. Retention windows, subprocessors and breach notice terms are where business data actually gets exposed, and providers vary a lot on all three.
- 04Your enterprise agreement overrides the public policy page. If a sales rep tells you something the contract doesn't say, the contract wins.
- 05A data processing addendum, not a blog post, is the document that has to say no training, a stated retention period and named subprocessors.
Two different products get treated as one question
Someone on the team read that ChatGPT trains on your conversations. That's true for the consumer product, and it gets carried straight into a conversation about calling OpenAI's API from a piece of internal software. Those are different agreements with different providers of the same company.
OpenAI's API platform states that inputs and outputs sent through the API are not used to train its models unless you opt in. Anthropic's commercial terms say the same for Claude's API and console usage. Google states the equivalent for the Gemini API through Vertex AI or AI Studio's paid tier.
The consumer apps run a separate policy. ChatGPT's free and Plus tiers may use chat content for training unless a user disables it in settings. Gemini's consumer app carries a similar default. Pasting client data into a browser tab is a different tier than calling the API. That tab is where the training question is still live.
Training was never the real risk
Get the training question settled and a harder set remains. How long is a request kept on the provider's servers, and who else can see it while it sits there.
Providers keep API requests for a retention window even when they don't train on them, usually for abuse monitoring. That window is a stated number of days in the policy, not zero. Ask for it directly, and ask whether it can be shortened or turned off for your account.
Subprocessors are the second gap. Your prompt may pass through a cloud host, a monitoring tool or a fraud-detection vendor before it reaches the model. A provider's data processing addendum has to name those subprocessors, not describe only its own handling.
Breach notice terms come third. If a subprocessor is compromised, the contract should state how fast you're told. A no-training clause with no retention limit and no named subprocessors reads as a policy page, not a guarantee.
Related questions
01Does turning off training also stop data retention?
No. Those are separate settings. Opting out of training does not change how long the provider keeps the request, whether that's a consumer toggle or an API default. Retention is a separate policy, usually stated in days. It applies whether or not the data is ever used for training.
02Is a self-hosted or open-weight model automatically safer?
It removes the provider's servers from the picture, which removes their retention window and subprocessor chain. It does not remove your own responsibility for access control, logging and who inside your organisation can query the model. Self-hosting trades one set of risks for another; it does not delete the category.
03What should be in the contract before we send business data to an LLM API?
A stated no-training default for the tier you're using, a named retention period, a list of subprocessors, and breach notice terms with a timeframe. If any of those four is missing or vague, ask for it in writing before the integration goes live.
04Does this apply to fine-tuning a model on our own data?
Fine-tuning is a separate, opt-in action you take deliberately, not something that happens by sending normal requests. The question here is about the default handling of ordinary API calls. Fine-tuning has its own data-handling terms, usually stricter, because you're deliberately building a model that has learned from your data.
Related
- What to ask an AI vendor →The fuller vendor-diligence checklist, of which this is one item.
- Does sending client data to an LLM waive privilege →The sharper version of this question for firms bound by privilege.
- Self-hosted vs managed LLM →What you trade by keeping the model off a provider's servers entirely.
- Shadow AI →What happens when staff skip this question and paste data into a free tool anyway.
