Does the paid or business version of an AI tool train on your data?

Paid API and enterprise tiers are generally excluded from training by default under providers' current terms, while free tiers (consumer apps and some free API plans) often use your chats unless you opt out. Most people never change the default, so the out-of-the-box behavior is the one that actually decides whether your data is used. Policies change, so confirm the current terms for the exact plan you use. The price you pay won't tell you on its own.

If a model has already trained on my data, can I have it removed?

Usually not in any direct way. Opting out is forward-looking: it stops future training runs from using your data, but it does not reach back into a model that already learned from it. That gap is why an opt-out is a weaker protection than it sounds once the data has been sent.

What is a zero data retention agreement, and does it stop training?

Zero data retention (ZDR) means the provider does not keep your inputs and outputs once the response is returned, which also removes the stored copy that could feed a later training run. It is typically offered to enterprise or API customers on request, sometimes with narrow carve-outs required by law or to review the most severe safety violations, under each provider's current terms. Because it is granted per contract, ask for it in writing rather than assuming a business plan includes it.

Answers

Which AI tools don't train on your data?

The APIs and business plans of OpenAI, Anthropic, Google and Microsoft do not train on your data: all four commit to this in their enterprise and paid API terms. Consumer plans are different. Consumer ChatGPT uses your conversations for training by default, with an opt-out in the data controls; Anthropic's consumer plans ask you to choose. And training is only one of three separate promises worth checking, alongside retention and human review.

Training, retention and human review are different promises

When people ask whether an AI trains on their data, they usually mean three things at once:

Training: your inputs are used to update model weights. Once that happens it cannot be undone, and fragments of your text can, in principle, surface in someone else's output.
Retention: your inputs sit on the provider's servers for some period, trained on or not. Retained data can be breached, subpoenaed, or read.
Human review: provider staff or contractors may look at conversations, typically for abuse monitoring or when something gets flagged.

A provider can honestly promise no training while retaining prompts for weeks and reserving the right to have humans read flagged ones. The three promises are made separately, per plan, often in different documents. Read the training, retention and review terms for the specific plan you are on.

Where the major providers stand

Consumer ChatGPT, meaning the free and paid personal plans, uses conversations to improve models by default; you can turn this off in the data controls. Anthropic's consumer Claude plans put the choice in front of you at signup and ask you to pick.

Business and enterprise plans are a different contract. ChatGPT Team and Enterprise, Claude's business plans, and the equivalent Google and Microsoft offerings state in their terms that customer data is not used for training. The same holds for the paid and enterprise APIs of the major providers: under the providers' current API terms, inputs and outputs are not used for training by default. (Free API tiers can differ. Google's free Gemini API tier, for example, may use prompts to improve its products.)

This promise is contractual, not technical. Nothing in the architecture prevents training on what you send; a paragraph in the terms does. Paragraphs get rewritten.

Why "just opt out" is not the whole answer

Opting out of training does not opt you out of retention. Your prompts still land on the provider's infrastructure and still sit there for whatever window the retention policy sets. The opt-out changes what happens to the data, not whether it exists.

Terms change. The training question is answered by a document the provider can update, and providers have moved the line before: a default you accepted one year can be changed or reversed the next. A promise you have to re-verify every quarter is a thin base for a data policy.

The wrong account gets used. The no-training commitment attaches to a workspace, not a person. An employee who pastes a client contract into their personal ChatGPT account, because the company login was on the other machine, has just put that contract under consumer terms. One consumer login is enough to undo the policy for everything typed into it.

Data that never left cannot be trained on

The stronger position is to stop tracking toggles and shrink what there is to protect. A client's name that never reached the provider cannot be trained on, retained, reviewed or breached, whatever the terms say this quarter. That is data minimisation, and GDPR expects it from you anyway: send the least personal data the purpose needs. It is the same principle that decides whether ChatGPT can be used in a GDPR-compliant way.

There are two common ways to get there. You can run the model on your own hardware so nothing leaves at all, or you can mask sensitive values before they reach a cloud model. The trade-offs (cost, model quality, upkeep) are real, and we compare them in self-hosted LLM or masked cloud AI.

How Cedreon handles it

Cedreon is an AI workspace for sensitive data, developed in Switzerland, built on the masking approach. Before any text leaves for the model provider, it is screened: names, IBANs, addresses and other identifiers are found by pattern matching plus layered best-effort machine detection, and replaced with placeholders. The originals stay in an encrypted vault on Cedreon's side, and the reply is filled back in for the signed-in user. The provider never sees the real values, so its training and retention terms apply to masked text. You chat with models from Anthropic and OpenAI and can upload Word, Excel and PDF files, with answers pointing to the source passage.

Masking pseudonymizes rather than anonymizes, and detection is best-effort. So the honest claim is reduced exposure; we do not claim perfection. That is why every exchange also lands in an append-only, hash-chained audit log: a firm can show afterwards exactly what left the house and what was replaced first. Admins and users can add plain-language guardrails that block, warn or flag on every chat. To see the masking on your own text, Try Cedreon free.

Common questions

Does the paid or business version of an AI tool train on your data?: Paid API and enterprise tiers are generally excluded from training by default under providers' current terms, while free tiers (consumer apps and some free API plans) often use your chats unless you opt out. Most people never change the default, so the out-of-the-box behavior is the one that actually decides whether your data is used. Policies change, so confirm the current terms for the exact plan you use. The price you pay won't tell you on its own.
If a model has already trained on my data, can I have it removed?: Usually not in any direct way. Opting out is forward-looking: it stops future training runs from using your data, but it does not reach back into a model that already learned from it. That gap is why an opt-out is a weaker protection than it sounds once the data has been sent.
What is a zero data retention agreement, and does it stop training?: Zero data retention (ZDR) means the provider does not keep your inputs and outputs once the response is returned, which also removes the stored copy that could feed a later training run. It is typically offered to enterprise or API customers on request, sometimes with narrow carve-outs required by law or to review the most severe safety violations, under each provider's current terms. Because it is granted per contract, ask for it in writing rather than assuming a business plan includes it.