Your message does not stay on your device
When you type into an AI chat tool, the text is normally sent over the internet to servers where the model runs. Unlike a calculator or a word processor working offline, the processing happens elsewhere. That means the company running the service can, in principle, receive, log and keep what you send.
What actually happens differs from one service to the next, and it can change when the terms change. The points below describe common patterns, not a rule that applies to every product. The service’s own privacy policy and settings pages are the authority for any specific tool.
What may be stored
Several kinds of information can be kept:
- The conversation itself. Many services save chats to your account so you can return to them. Saved chats may stay until you delete them, or for a set period.
- Account details. Your email address, name, payment information if you pay, and the device or location information that most online services collect.
- Uploaded files and images. Documents, photos and screenshots you attach are inputs too, and can contain far more personal detail than a typed question.
- Usage data. Timestamps, features used and error reports.
Even after you delete a chat from your view, the service may keep a copy for a limited time for security, legal or operational reasons. Deleting from the screen and deleting from the company’s systems are not always the same thing.
Why companies keep it
Reasons usually fall into a few groups. Storing history is a feature: you can pick up where you left off. Logs help the company detect abuse and fix faults. Some services also use conversations to improve their models, which is where privacy concerns are sharpest, because material you typed could contribute to future training. How training data is gathered more broadly is covered in the article on training data.
Some services also let trained staff or contractors review a sample of conversations, for example to check safety or quality. If a policy mentions human review, assume that a person could read some of what you write.
Settings that can limit it
Many tools offer controls, though names and defaults vary. Things worth looking for:
- A switch for model improvement. Some products let you opt out of having your chats used for training. In some, it is on by default, and in others it is off.
- Chat history controls. You may be able to turn history off, delete individual chats or clear everything.
- Temporary or incognito-style chats. These typically do not appear in your history and may be kept for a shorter time, but they are not necessarily invisible to the provider.
- Data export and deletion requests. Depending on where you live, you may have legal rights to see or delete data a company holds about you. Process and scope differ.
- Connected apps and memory features. Some assistants can remember details across chats or link to email, calendars and files. These are convenient, and they also widen what is stored. Review what is switched on.
- Workplace or education accounts. These often follow different rules set by the organisation, which may include its own retention and monitoring.
Settings reduce exposure but do not remove it. A setting that says your chats are not used for training does not necessarily mean they are not stored.
Risks beyond the company’s own policy
Storage creates other possibilities. Any stored data can be exposed in data breaches, and an account with a weak password makes that easier. Using a password manager and two-factor authentication on your AI account is sensible for that reason, since a chat history can reveal a lot about you.
Shared devices and shared accounts are another route. If a family member or colleague can open your account, they can read your history.
There is also the question of what other people’s information you are pasting in. A client list, a friend’s medical details or a colleague’s email is not only your data. Many workplaces have rules about this, and breaking them can have consequences beyond privacy.
Models can sometimes reproduce fragments of material they were trained on, though how likely that is varies and is generally low for any specific item. It is a further reason to avoid typing things you would not want to appear elsewhere.
Habits that help
- Leave out what you do not need to share. Replace names, addresses and account numbers with placeholders when asking for help drafting something.
- Do not paste secrets. Passwords, recovery codes, card numbers and government ID numbers do not belong in a chat.
- Be cautious with health, legal and financial details. If you want help understanding a letter or a report, remove identifying information first.
- Read the settings once. A few minutes on the privacy page of a tool you use daily is worth it. Check again after major updates.
- Use separate accounts or modes for sensitive work where your situation allows.
- Delete what you no longer need, while remembering that deletion may not be instant or complete.
A good test is to imagine the text appearing in an email to a stranger. If that would worry you, rewrite it or leave it out.
Reading the fine print
Privacy policies are long, but a few questions get you most of the way: Is my content used to train models? Can humans review it? How long is it kept? Can I delete it? Who else receives it? Guidance on reading privacy policies without a law degree applies directly. If a policy is vague on those points, take that as a reason for more caution.
Wider habits matter too. The same care you apply to your digital footprint on websites and apps applies to chatbots, because the details you give an assistant add to what companies may know about you. For general principles, see the material on online privacy.