How to stop AI tools training on your prompts

The question people ask is "does OpenAI/Google/Anthropic train on my data?" — as though it were a property of the company.

It mostly is not. It is a property of the tier you are on. The same vendor will train on your prompts on the free plan and contractually promise not to on the paid one. Here is what each says, in its own terms, as of 17 July 2026.

The short answer, by tier

Google: the free tier is the trap

This is the one that catches people, because "free API quota" feels like a technical detail rather than a privacy decision.

Google's Gemini API terms are explicit. On Unpaid Services — which expressly includes Google AI Studio and the unpaid Gemini API quota — Google uses the content you submit and the generated responses to develop its products and machine learning technologies. The terms also state that human reviewers may read, annotate and process your API input and output, with the data disconnected from your account first.

The terms then give the plainest instruction any vendor offers on this subject: "Do not submit sensitive, confidential, or personal information to the Unpaid Services."

On Paid Services, the same document says the opposite: Google does not use your prompts or responses to improve its products, and logs them only for a limited period to enforce its usage policy.

The fix is the fix nobody likes: enable billing. There is no setting. The tier is the setting.

One extra catch worth knowing: if you use Grounding with Google Search, Google stores prompts and output for thirty days for that feature — including on paid quota.

Anthropic: consumer plans have a switch

For Claude Free, Pro and Max — and Claude Code used from those plans — your chats are used to improve the models only if you turn that on. It is a setting in your privacy preferences.

Two details are worth knowing:

• Incognito chats are excluded even if you have model improvement enabled. That is a hard exclusion, which makes it the reliable option for a one-off sensitive conversation.
• Safety-flagged conversations are a carve-out. If a conversation is flagged for safety review, it may be used regardless of your setting.
• Feedback is separate. Pressing thumbs up/down stores the whole conversation for up to five years, de-linked from your user ID. That is a deliberate act, but it is a bigger one than the button implies.

Anthropic's commercial products — the API and Claude for Work — are governed by separate commercial terms, not the consumer policy above.

OpenAI: not trained, but not invisible

OpenAI's developer documentation is unambiguous that data sent to the API has not been used to train its models since March 2023 unless you explicitly opt in.

But that is a statement about training, not about storage. By default, OpenAI generates abuse-monitoring logs that may contain your prompts and responses. Excluding your content from those logs means applying for Zero Data Retention or Modified Abuse Monitoring, which are subject to approval and additional terms — not a toggle in settings.

The distinction that matters: training ≠ retention

Every vendor above will tell you, accurately, that they do not train on your paid API data. All three also retain some of it.

• OpenAI: abuse-monitoring logs by default; ZDR requires approval.
• Google (paid): prompts and responses logged for a limited period for policy enforcement.
• Google (search grounding): thirty days, paid or not.

None of this is sinister — it is how you catch people generating abuse material — but "we do not train on your data" and "we do not keep your data" are different sentences, and only one of them is usually being said.

If your actual requirement is this text must not leave the building, none of these controls deliver it. That is a job for a local model, where the question does not arise.

What to actually do

1. If you are on a free API tier, assume you are the product. Google says so outright. Do not paste client data into AI Studio.
2. Enable billing before you send anything confidential. On Google this is the entire control.
3. On Claude consumer plans, check your model improvement setting — and use Incognito for anything genuinely sensitive rather than trusting a global toggle.
4. Don't confuse "no training" with "no logging". They are different promises, and only the first is easy to get. On OpenAI, excluding your content from abuse-monitoring logs means applying for Zero Data Retention — approval and extra terms, not a checkbox. Google's paid terms log prompts for a limited period for policy enforcement and describe no way to opt out. Assume "not logged" is the harder ask everywhere, and check rather than presume.
5. If it truly cannot leave your machine, run it locally. Everything else is a promise, and promises have terms.