Who reads what you type

Three questions, asked of every tool

1. 1Is my input used to train the model?
2. 2Is it stored, and for how long?
3. 3Can a human being read it?

These are separate questions with separate answers, and companies answer the first one loudly and the other two quietly. "We do not train on your data" is not "we do not keep your data." A provider can truthfully say the first while retaining your conversations for thirty days and allowing staff to review flagged ones.

The pattern, as of 2026

This is the shape it has held for a while. Verify it for your specific tool, because it changes.

Consumer free tiers. Frequently used for training by default. Usually there is a switch in settings, and it is usually not the first thing you see. In some jurisdictions the default is reversed by law.

Consumer paid tiers. Varies by company, and this surprises people. Paying does not automatically opt you out. Several products have trained on paying subscribers' conversations, with a setting available to decline.

Business, team and enterprise tiers. Generally covered by a contract that says your content is not used for training. If you handle other people's information for a living, this is what you are actually buying.

API access. Across the major providers, API inputs are generally not used for training by default. Retention for abuse monitoring is common — often around thirty days — and zero-retention arrangements exist for customers who ask and qualify.

Local models. Nothing leaves your machine, provided the app around it is not sending telemetry. Check the settings once.

The sentence people miss

Somewhere in most privacy policies is a line saying that a sample of conversations, or conversations flagged by automated systems, may be reviewed by human beings for safety and quality.

This is not sinister — it is how abuse gets caught and how models get fixed. But it means the honest mental model is not "a machine reads this and forgets." It is "a machine reads this, and occasionally a person might."

How to check, in five minutes

• Open Settings. Look for Data controls, Privacy, or Improve the model. Read every toggle, including the ones already on.
• Open the privacy policy and use your browser's find function. Search for: *train*, *improve our models*, *human review*, *retention*, *delete*.
• For a paid work tool, find the trust or security page rather than the consumer policy. They often say different things.
• Take a screenshot with the date visible. Terms change, and a dated screenshot is the only record you will have of what you agreed to.

If you cannot find a clear answer in five minutes, treat that as the answer.

Rules that do not depend on trusting anyone

Do not paste what you could not email. Passwords, API keys, a customer database, unredacted medical or legal files. Not because a leak is likely, but because the cost of being wrong is far larger than the effort of not doing it.

Redact by role. The model does not need the name. "The patient," "the supplier," "the employee" produces the same quality of answer as "Mrs Adeyemi." This one habit removes most of the risk from most of the work.

Remember deletion has limits. You can usually delete a conversation. If data has already gone into a training run, it cannot be extracted from the trained model afterwards. Deletion going forward is real. Deletion going backwards is not.

Check whether it is even your decision. Your employer may have a policy. Your country may have a law — GDPR in Europe, the DPDP Act in India, LGPD in Brazil, POPIA in South Africa. Handling someone else's personal data usually shifts this from a preference to an obligation.

None of this requires paranoia. It requires knowing which of three promises a company is making, and reading the one sentence where they say it.