A conversation is not a private thought
Typing into a chatbot feels like talking to yourself or to a trusted advisor. Technically, though, you are sending text to a company’s service. What happens to that text depends on the provider, the product, your account type and your settings, so no general statement is true of every assistant.
This article covers AI chatbot privacy in general terms. It describes what typically can happen and offers cautious habits. It cannot tell you what a particular service does today; for that, its own privacy documentation is the authority.
What may be stored
Depending on the service, any of the following might be kept:
- The text of your prompts and the replies, often saved as a conversation history so you can return to it.
- Files and images you upload, along with anything the assistant extracts from them.
- Voice recordings or transcripts, if you speak to the assistant.
- Account details, such as your name, email address, payment information and subscription status.
- Technical and usage data, such as device type, approximate location from your network address, timestamps and which features you use.
- Feedback, such as ratings and comments you submit about answers.
- Memory or personalisation notes, on assistants that remember preferences or facts between conversations.
Retention periods vary. Some services keep conversations until you delete them, others delete after a set time, and some offer modes where history is not saved. Even when you delete a conversation from view, backups or legal obligations may mean copies are removed on a delay, so check the stated policy.
Who might see it
Several parties can potentially have access, in different circumstances.
- The provider’s systems, which process the text in order to respond.
- Employees or contractors, who in some services may review a sample of conversations for safety, quality or abuse prevention. Providers usually describe when this happens in their policies.
- Other businesses, such as hosting or analytics providers working for the company.
- Authorities, where the law requires a company to hand data over.
- Anyone who gains access to your account, which is why account security matters. Strong, unique passwords and two-factor authentication help, and a password manager makes that easier.
- People you share with. Some tools let you share a conversation through a link, and anyone with that link may be able to read it.
- Attackers, if the provider suffers a data breach. A company holding stored conversations is a target like any other.
Training: the question many people ask
A common worry is whether your conversations will be used to improve the model. Practices differ. Some services use conversations for improving products unless you opt out, some do not by default, and business or education plans often have different terms from free consumer accounts. Settings may exist to control this, and their names and defaults change.
It is also worth knowing that models do not usually work like a database that repeats earlier users’ chats on demand. But the possibility that fragments of training material could surface in output is an area of ongoing discussion, and it is a good reason not to put truly sensitive information in at all. The way such material shapes models is explored under training data.
Details that are sensible to keep out
A useful rule is to avoid typing anything you would be uncomfortable seeing in a leaked document. In practice that usually means leaving out:
- passwords, passcodes, security answers, recovery codes and API keys
- full payment card numbers, bank account details and government ID numbers
- detailed health, legal or financial information tied to your name
- other people’s private details, such as a friend’s medical situation or a customer list
- confidential work material, unless your employer has approved the tool for that purpose
- precise home addresses and travel plans
- private photos, especially of children or of other people who have not agreed
You can often get the help you need by changing the details. Replace names with initials or roles, round numbers, remove dates and places, or describe the situation in general terms. A message such as “a colleague keeps missing deadlines; how should I raise it?” does not need names.
Settings worth checking
Look for options such as:
- turning off chat history or using a temporary or incognito-style chat
- opting out of having conversations used to improve models
- viewing, exporting and deleting your history
- managing or switching off memory features
- controlling shared links
- reviewing which apps or plug-ins the assistant is connected to, since linking email or files gives the assistant more to see
Not every service offers all of these, and settings can change after updates, so revisit them from time to time.
Local versus remote processing
Where the model runs also matters. Services that run on remote servers must receive your text, while on-device AI may process some requests without sending them anywhere. A local label is worth verifying and is not a guarantee.
Legal protections
Some places have data protection laws that give people rights to access, correct or delete personal data, and require companies to explain what they collect. What applies depends on where you live and where the company operates. Reading the provider’s privacy policy, even briefly, can tell you what it says about retention, sharing and training.
A short routine
- Assume anything you type could be stored.
- Strip out identifying details before you paste.
- Check the history and training settings once, and revisit occasionally.
- Protect the account itself.
- Use a different tool, or none, for matters that are truly sensitive.
Used with these habits, assistants can stay helpful without becoming a record of things you never meant to share.