What to tell an AI assistant, and what to keep to yourself

An AI assistant feels like a private conversation. It is closer to a postcard addressed to a large company: useful, usually harmless, and not the place for your PIN.

Where your words go

Whatever you type is sent to the company that runs the assistant. The makers say so themselves. In alphabetical order, and at the time of writing:

• ChatGPT. OpenAI’s help pages say it may use your content to train its models, and that you can switch this off with a setting called “Improve the model for everyone”, under Settings, then Data controls.
• Claude. Anthropic’s privacy pages say chats are used to improve Claude if you choose to allow it, through a “Model Improvement” option in its privacy settings, and that conversations flagged for safety review may also be analysed. If you allow training, Anthropic says it may keep your data in de-identified form for up to five years.
• Copilot. Microsoft released an updated Copilot app on 18 August 2026. Its help page for that app says that, when you are signed in with a personal Microsoft account, “Prompts, responses, and your file contents when using the Microsoft Copilot app aren’t used to train foundation models.” It does not say how long conversations are kept. Microsoft’s older privacy FAQ, which it now says applies only to the previous version of the app, says conversations there are used for AI training unless you opt out, that some are subject to automated and human review, and that they are stored for 18 months by default.
• Gemini. Google’s privacy page says chats are used to improve its services, including its AI models, while a setting called “Keep Activity” is on. It says a subset of chats is read by human reviewers, and that reviewed chats can be kept for up to three years even if you delete your activity.

Google puts the conclusion more plainly than we could: “Please don’t enter confidential information that you wouldn’t want a reviewer to see or Google to use to improve our services, including machine-learning technologies.” Microsoft’s privacy FAQ for the older Copilot app says much the same: “You shouldn’t provide any confidential or sensitive personal data that you would not want Microsoft to use for purposes explained in this FAQ and in the Microsoft Privacy Statement.”

Names and menus change. The durable version: each assistant has a setting that controls whether your chats are used for training. Look under Settings for data or privacy controls.

What the official bodies advise

The UK’s National Cyber Security Centre (NCSC) says queries are visible to the organisation providing the service, are stored, and “will almost certainly be used for developing the LLM service or model at some point”. (An LLM, or large language model, is the technology behind these assistants.) Stored queries, it adds, could also be hacked or leaked. Its advice is “not to include sensitive information in queries” and not to submit anything “that would lead to issues were they made public”.

Canada’s privacy commissioner gives similar guidance for individuals: limit what you share about yourself, do not share personal information about other people, alter identifying details, delete your chat history where you can, and review your privacy settings.

The never list

Keep these out of every AI chat, whatever the settings say:

• Passwords, PINs and one-time codes.
• Bank, card and pension account numbers.
• Your National Insurance number (the UK’s equivalent of a US Social Security number), or your passport or driving-licence number.
• Full medical records with your name on them.
• Other people’s private information: a friend’s diagnosis, a family dispute with real names.
• Anything confidential from work.

No assistant needs any of these to give you a good answer.

The fine list

Context makes an answer useful, and most of it costs you nothing:

• Your rough age and general level of fitness or experience.
• The town you are travelling to and what you enjoy doing.
• Your interests, your budget in round numbers, what you have already tried.
• The text of a letter with the names and reference numbers taken out.

“I’m in my sixties, visiting Lisbon for four days, I like architecture and dislike queues” gives away nothing and gets a far better answer than “plan a trip”.

Getting help with a private document anyway

Suppose a pension provider sends you a baffling letter. You have two safe routes.

Remove or replace. Before pasting, take out names, addresses, account numbers and reference numbers. Swap them for “Person A” or “[account number]”.

Describe instead of paste. “My pension provider says I must choose between two options by a deadline; one mentions drawdown. What questions should I ask them?” often does the job with no document at all. This is general information, not financial advice: the decision itself belongs with you and a regulated adviser.

Housekeeping

Delete old chats. The makers named above all let you delete conversations. Anthropic says a deleted conversation is removed from its back-end systems within 30 days.

Use the temporary mode for sensitive questions. At the time of writing, OpenAI offers “Temporary Chat”, Anthropic “Incognito” chats and Google “Temporary chats”. Each maker says these are not used to train its models. They are not instantly erased, though: OpenAI says temporary chats may be kept for up to 30 days for safety purposes, Anthropic says incognito chats are kept for 30 days by default, and Google says 72 hours.

Think before you connect. Assistants increasingly offer to link to your email, files or calendar. Google’s page says connected apps let Gemini access emails, files, events and photos. Useful, and a much bigger door than one pasted paragraph. Connect only what you will use.

Lock the account itself. Your chat history sits behind one login. Give it a strong password and two-step verification, as set out in Passwords and two-factor authentication, made simple.

A word on companion apps

Some apps offer an AI “friend”. The US Federal Trade Commission (FTC), which opened an inquiry in 2025 focused on children and teenagers, notes that such chatbots are generally “designed to communicate like a friend or confidant”, which may prompt some users to trust and form relationships with them. Whatever your age, it is a product made by a company, so the same rules apply: the never list holds here too.

This week

1. Find the training setting in your assistant and set it the way you want it.
2. Delete any old chat that contains something from the never list, and change any password you once typed into one.
3. Practise on a real letter: strip the names and numbers, paste, and ask for a plain-English summary.
4. Turn on two-step verification for the AI account.

For where this fits in the bigger picture, start with AI after 50: the complete guide.

Updated 6 October 2026: added how long Anthropic says it keeps incognito chats.

Updated 2 November 2026: Copilot’s privacy terms for the updated app Microsoft released in August 2026 (not used to train foundation models); the 18-month and opt-out details now apply only to the older app.