Explore AI
What happens to what you type
The most consequential question about AI, and the one people ask last.
Every hosted AI service works the same way at the level that matters: your text leaves your machine and arrives on someone else's computer.
What happens after that varies enormously — and the difference is between services people treat as interchangeable.
The three arrangements
Free tiers. Usually the most permissive about using your data. On many services, free access is paid for by allowing your conversations to be used to improve future models. Read the terms: this is frequently stated plainly, and frequently not read.
Paid consumer plans. Typically better. Many providers state that conversations are not used for training by default, and offer an opt-out or an incognito mode. Retention periods still apply.
Business and API accounts. Usually the strongest terms — no training on your data, defined retention, sometimes a data-processing agreement. This is often the only arrangement suitable for professional work.
What "we do not train on your data" means
It means exactly that, and no more. It is a policy, not a technical guarantee.
- It can change, and has changed, at several providers.
- It does not mean the text is not stored.
- It does not mean nobody can read it.
- It does not cover third parties in the chain.
None of that makes such promises worthless. It makes them promises.
The uncomfortable part
Metadata, retention and human review are where most of the real risk sits, not in training:
- Conversations are typically retained for a period, even when not used for training.
- Some providers review samples for safety and quality, by people or by automated systems.
- Anything sent to a model may be logged for abuse monitoring, sometimes for months.
- If you use a third-party app on top of a model, your text passes through that company too.
The rule that covers almost everything
Would you be comfortable if this were read by a stranger, and kept for a year? > > If not, do not put it in a hosted model.
That covers client material, medical information, legal documents, unpublished work, other people's personal data, credentials, and anything under an NDA. It is a two-second test and it is right nearly always.
When hosted is simply not an option
Some material cannot be sent anywhere, and the answer is not a better prompt — it is a different architecture:
- Run a model locally. Nothing leaves the machine. See what can your machine run.
- Rent a GPU. You control the machine, and you destroy it when you are done.
- Anonymise first. Strip names and identifiers, then use a hosted model.
The first option is the reason local AI exists at all, and for many people it is the only argument that matters.
A note on regulation
If you are handling personal data in Europe, this is not a matter of preference. Sending personal data to a model provider is processing it, with everything that implies — a lawful basis, a processor agreement, and honesty with the people whose data it is.
That is not legal advice. It is a reason to involve someone who gives it, before the project ships rather than after.
Where to go next
- Should you run it locally?
- What can your machine run?
- How to reach a model — the privacy column, compared