Skip to content

Glossary

Opt-out

An opt-out is a setting telling a provider not to use your conversations to train its models, which answers one of the three questions about your material and leaves the other two untouched.

In plain terms

A box you tick, usually buried two levels into settings, saying do not learn from this. It is worth ticking and it is smaller than people think. Your conversations may still be stored, may still be readable by somebody at the provider under defined circumstances, and may still appear in an export. The setting governs training and nothing else.

01

Why it matters

It takes about five minutes, it is often not the default, and it is the one privacy control most people can actually exercise without involving anybody. That combination makes it worth doing on the first day rather than the day somebody asks. It is also the control most likely to be mistaken for a complete answer, which is how organisations end up satisfied about training while material sits in logs for a period nobody established.

02

How it works

The setting sits with the provider and applies from the moment you change it, which means anything sent before is governed by whatever was in force then. That is the part worth knowing early: switching it on protects the future, and the past is a separate conversation about retention and deletion.

Business agreements frequently exclude training by default, so the setting matters most on consumer plans and least on the arrangements organisations sign. Checking which applies to your account is faster than reading anything, and it frequently turns out that the answer differs between two people in the same team.

It says nothing about retention or access. How long material is kept, who at the provider may look at it and under what circumstances, are governed separately and documented separately. A product that does not train on your inputs may still hold them for a defined period, and both facts are usually published.

Where a provider offers no such control at all, that is itself the answer to a question. It is not necessarily disqualifying for public material, and it does mean anything covered by a client agreement should go somewhere else, because you cannot promise what you have not been given the means to promise.

The setting, and everything it leaves untouched

The setting, and everything it leaves untouchedWhat ticking the box settles is real, specific and narrow: your material is not used to improve a model, from now, on this account. What it leaves alone is most of what people assume it covered. Retention runs on its own schedule. Access by staff at the provider is governed separately and usually documented. Anything sent before the setting changed was governed by what applied then. And everything derived from your material, the logs and transcripts and search indexes, sits outside it entirely. None of that argues against ticking it. It argues against the sentence that so often follows, which is that the privacy question has therefore been dealt with.Data controls [x] Improve the model for everyone1 off since you changed it, not before2 Retention .................. see the terms3 Who at the provider may read it .. see the terms Exports, logs, search indexes .... see the terms41The box, usually worded asimproving the model ratherthan as training, which iswhy it is easy to walk past.2It applies from the momentyou set it. Anything sentearlier was governed bywhatever was in force then,and deletion is a separaterequest.3Untouched by the box, andthe answer an auditor asksfor first.4Also untouched: everythingderived from your materialcarries some of it.
What ticking the box settles is real, specific and narrow: your material is not used to improve a model, from now, on this account. What it leaves alone is most of what people assume it covered. Retention runs on its own schedule. Access by staff at the provider is governed separately and usually documented. Anything sent before the setting changed was governed by what applied then. And everything derived from your material, the logs and transcripts and search indexes, sits outside it entirely. None of that argues against ticking it. It argues against the sentence that so often follows, which is that the privacy question has therefore been dealt with.
03

Seen in the wild

  • Open the data controls of an assistant you already use and check whether training is switched off, which most people discover it is not.

    ChatGPT
  • Compare the setting on a personal account against the terms of a business agreement for the same product, which frequently differ.

    Claude
  • Run a model on your own machine, where the question does not arise because nothing was sent anywhere to be learnt from.

    Ollama
04

Common misconceptions

People assume

Opting out means our conversations are not kept.

In fact

It means they are not used for training. Retention is a separate setting with a separate answer, and a provider can honour one entirely while holding your material for a defined period under the other. Both are documented, and reading the retention answer takes the same five minutes.

People assume

It applies to everything we have already sent.

In fact

Generally it applies from the moment you set it. Material sent earlier was governed by whatever was in force at the time, which is why the setting is worth changing before a pilot rather than after one. Where earlier material matters, deletion is the separate request to make.

05

Telling them apart

Opt-out vs Data privacy

Opt-out

One control, covering one question: whether your material is used to train a model.

Data privacy

The whole question, which also covers how long material is kept and who at the provider can see it.

Using the control is worth doing and answers a third of it. Reporting it as the answer is the mistake, because retention and access are entirely untouched by the box you ticked.

06

Questions

Where do I find it?
Usually in a data controls or privacy section of settings rather than anywhere prominent, and often phrased as improving the model or helping us get better rather than as training. It is worth looking for the wording rather than the word, because products describe the same setting quite differently.
Do we need it on a business plan?
Check rather than assume. Business agreements commonly exclude training by default, which is one of the clearer reasons to be on one, and it is still a documented fact rather than a general principle. Confirm it in the agreement you signed rather than in an article about the company.
Does it apply to everyone on our team?
It is a per-account setting, so one person changing it does nothing for a colleague using the same product. On team and business plans training is often excluded by the agreement the organisation signed rather than by a setting anyone has to find, and some providers also expose a control at administrator level. Either of those holds across an organisation in a way one person ticking a box does not.
07

Key takeaways

  • It governs training only, and leaves retention and access exactly as they were.
  • It generally applies from the moment you set it rather than retrospectively.
  • Business agreements often exclude training already, so the setting matters most on consumer plans.
  • A provider offering no such control has answered a question you needed answered.
  • It is five minutes of work and a third of an answer.
09

Tools that use this

  • ChatGPT

    Data controls where the setting is usually not the default.

  • Claude

    A personal setting against a business agreement, which often differ.

  • Ollama

    Running locally, where the question does not arise.

Last checked July 2026

All glossary terms