Skip to main content

What to keep out of AI chatbots, and the privacy settings to check first

Keep passwords, bank details and other people's private information out of AI chatbots. Check their privacy settings, and check anything important against an official source.

KBY Everyday Tech 9 min read Checked 22 November 2026 ChatGPT, Gemini, Copilot and Claude privacy settings, as their makers describe them on 22 November 2026 (personal accounts, on the web and in the phone apps) UK
Two hands typing a message on a smartphone at a desk, with a computer screen behind
Photo: Vladislav Šmigelski on Pexels

The scam reporting numbers and services are for the UK: Report Fraud covers England, Wales and Northern Ireland, and Scotland reports to Police Scotland.

Ask an AI chatbot whatever you like. But keep your passwords, bank details and other people's private information out of it. Leave out medical or legal details you would not want a stranger to read, too.

Before you rely on one, check two settings: whether your chats are saved, and whether the company may use them to improve its AI. Then check anything important somewhere official. Every company in this guide says its chatbot can get things wrong.

A chatbot is a website or app you type (or speak) a question into, and it writes back in ordinary sentences. What you type is called a prompt. If you are new to them, start with our beginner's guide to AI chatbots.

This guide covers OpenAI's ChatGPT, Google's Gemini, Microsoft's Copilot and Anthropic's Claude. Each is described from its maker's own help pages, as they read on 22 November 2026. If a menu looks different, check the maker's page linked below.

Things to leave out of a chat

OpenAI's ChatGPT FAQ says: "Please don't share any sensitive information in your conversations."

Google's Gemini Apps Privacy Hub says people read some chats, including trained reviewers from its service providers. It asks you not to enter confidential information you wouldn't want a reviewer to see, or Google to use to improve its services.

The National Cyber Security Centre (NCSC) is the UK government's cyber security agency. In a 2023 blog post, ChatGPT and large language models: what's the risk?, it says the company running a chatbot can see what you type, and stores it. Stored chats could be hacked, leaked or accidentally made public.

So the NCSC advises against putting sensitive information into a public chatbot. Nor should you ask anything that would cause you problems if it were made public. A health or relationship worry can be sensitive just because of what it reveals about you.

Microsoft's Transparency Note for Copilot gives a list for teenagers. Don't share passwords, your precise location, financial details or health records. Don't share other sensitive personal information about yourself or anyone else.

Some cases deserve a closer look:

  • Codes and passwords. The government's Stop! Think Fraud campaign calls it a warning sign when anyone asks for a one-time passcode, or for your PIN or password in full.
  • Other people. Google asks you to get permission before using people's faces or voices in Gemini.
  • Real names and dates. The Information Commissioner's Office (ICO) is the UK's data protection regulator. Its advice on using AI to help write a data protection request or complaint suggests placeholders such as "[person A]" or "[insert date]" instead.
  • Health and legal matters. Google says not to rely on Gemini for diagnosis, treatment, medical advice or legal, financial or other professional help.
  • Advice from a lawyer. The ICO warns that pasting it in could affect its confidentiality, and the legal protection that can keep it from being shared.

Where your chats go, and the switches that control it

Two separate things can happen to what you type. The chat is saved to your account so you can come back to it: your history. And the company may use it to improve its AI.

The companies call that training their models. The model is the program behind the chatbot, and training is how it learns from examples. Each usually has its own setting.

Some companies keep a copy for a while even after you delete a chat. A chat you rate with the thumbs up or down buttons can be treated differently too.

ChatGPT

OpenAI's page on data controls in ChatGPT says data sharing is on by default for personal Free, Plus and Pro accounts. The setting is called Improve the model for everyone. To turn it off on the web:

  1. Open your account menu.
  2. Select Settings, then Data controls.
  3. Select Improve the model for everyone, turn it off and select Done.

In the iPhone or Android app:

  1. If you see a Home tab, select it, then open the sidebar.
  2. Select your profile icon to open Settings.
  3. Select Data controls and turn off Improve the model for everyone.

Turning it off does not delete your saved chats. OpenAI says voice mode has a separate control, and a chat you rate with a thumbs up or down may still be used for training.

For one conversation you don't want kept, open a new chat, select Temporary, and choose Personalized or Unpersonalized before your first message. OpenAI's page on temporary chat says it stays out of your history and isn't used to improve its models. OpenAI may still keep a copy for up to 30 days for safety purposes.

To delete a chat, select the more options menu (•••) next to it, then Delete. OpenAI says it is then permanently deleted within 30 days, with some exceptions, such as security or legal reasons.

Google Gemini

Google's page on managing your Gemini history says History is on by default if you are 18 or over. By default, history older than 18 months is deleted automatically. To turn History off on a computer:

  1. Go to gemini.google.com.
  2. Click Settings and help, then History.
  3. Near the top, click On, then Turn off or Turn off and delete history.

In the Gemini phone app, tap Menu, then your profile picture or initial, then Gemini history, and do the same.

Even with History off, Google says your chats are saved with your account for around two months. It says this is to provide the service, process feedback and protect its users and the public.

Training is a separate choice. The same page says the settings that let Google train AI models on your chats and recordings are off by default. They can only be turned on while History is on.

With training off, the Privacy Hub says, people can still review your chats to improve Google's services in other ways. Reviewed chats are separated from your account and kept for up to three years, even if you delete them.

For a temporary chat, click Temporary chat next to New chat at gemini.google.com. Google says these are not used to train its AI models, but are kept with your account for around two months.

Microsoft Copilot

Microsoft has offered an updated Copilot app since 18 August 2026. Its page on your privacy controls and choices covers that version, used with a personal Microsoft account.

Its page about your activity history says prompts, responses and file contents in the app aren't used to train foundation models. Those are the large models Copilot is built on.

Copilot keeps your chat history. To delete a chat, select it in the Chats list on the left, then select More (...) and Delete. For your whole history, Microsoft points you to its privacy dashboard.

Other choices are in Settings, then Personalization:

  • Turn off Saved memories to stop Copilot remembering details from your chats. To delete what it already saved, select Manage, then Delete all memories.
  • Turn off One shared experience to stop your Copilot chats shaping what you see on Bing, Edge and MSN.
  • Turn off Allow ads personalization if you see ads in Copilot. Microsoft says personalised ads might use your chat history and saved memories.

Microsoft's older privacy FAQ covers the older app. It says older chats not moved over by 31 December 2026 will be permanently deleted. You move them by signing in to the updated app.

Claude

Anthropic's page is my data used for model training? covers the Free, Pro and Max plans. Anthropic says it uses your chats to improve its models if you allow it, or if a chat is flagged for a safety review.

To check or change the setting on a computer:

  1. Select your name in the settings menu.
  2. Select Settings, then Privacy.
  3. Use the switch under Help Improve our AI models.

On a phone, go to your name, then Settings, then Privacy.

If you allow it, Anthropic says it may keep your chats for up to five years, with the details that link them to you removed. If you turn it off, it won't use your chats in future training. Anything already in a trained model stays there.

Its page on how long it stores your data says a deleted chat leaves your history straight away. It is deleted from Anthropic's systems within 30 days.

For one conversation you don't want kept, use an incognito chat: click the ghost icon at the top right when you start a new chat outside a project. Anthropic says these aren't saved to your history or used for training, but are kept for 30 days.

When the answer sounds right but isn't

Each maker warns of this. OpenAI's page does ChatGPT tell the truth? calls this a hallucination. Examples include wrong dates, made-up quotes and sources that don't exist, often stated with confidence.

Anthropic's page on incorrect or misleading responses says Claude can write things that look correct but are very mistaken. Microsoft says Copilot can make up content that sounds reasonable but is wrong. Google says Gemini may produce inaccurate information, including about people.

Before you act on an answer about your health, money or rights:

  1. Ask where it got the answer, then open the links yourself. ChatGPT and Copilot show links when they draw on the web.
  2. Read the original page, not just the summary. Anthropic says the original may hold details Claude left out.
  3. Check facts on the official site. The ICO says AI can make up laws and dates, and get contact details wrong.
  4. Ask a plain question, not a leading one. The NCSC says these tools are often gullible with leading questions.

Scams that use AI, and how to answer them

Criminals use these tools too. The government's Stop! Think Fraud campaign says advances including AI make a fake email, advert or website look very convincing.

The NCSC warned in 2023 that chatbots could help criminals write convincing phishing emails in their targets' own language. Phishing means messages that trick you into handing over details or money.

The campaign describes AI voice cloning. Fraudsters fake the voice of someone you know or trust, claim an emergency and ask for money.

It also explains deepfakes: digitally created or altered images, videos and audio that copy a real person. That could be a celebrity, someone you know or you. Its page on staying alert to fraud mentions social media investment adverts with deepfake celebrity endorsements.

What Stop! Think Fraud advises:

  • Agree a safe phrase with close family and friends. If a caller claiming to be one of them can't say it, hang up.
  • If you feel pressured, hang up. Call back on a number you know is genuine, such as the one on the back of your card.
  • Call 159 to reach your bank, or 101 for the police.
  • Don't trust the number on your screen. Fraudsters can fake it, which is called spoofing.
  • Be wary of anyone asking for a one-time passcode, cryptocurrency or gift cards.

Our guide to fake delivery texts shows how to report a scam text to 7726.

If you think you have been caught out

If you have lost money, tell your bank as soon as possible, says Stop! Think Fraud's page on reporting fraud. Use a number you know, such as the one on your card, or call 159.

Then report it to Report Fraud, online or on 0300 123 2040. In Scotland, call Police Scotland on 101. If anyone is in immediate danger, call 999.

To report a suspicious call, text 7726 with the word Call and the caller's number. Report a fake website to the NCSC, and a fake advert to the Advertising Standards Authority.

If you gave a password away, the NCSC's page on what to do if you've shared sensitive information says to change it on every account that uses it. Our guide to two-step verification adds a second lock to your important accounts.

Common questions

Do people at these companies read my chats?
Some do. Google says a subset of Gemini chats is reviewed by people, including trained reviewers from its service providers, after the chats are separated from your account. OpenAI's ChatGPT FAQ says it reviews conversations to improve its systems and check they follow its policies. Anthropic says Claude conversations flagged for a safety review may be used to improve how it detects misuse. That is why the makers ask you to leave sensitive details out.
If I delete a chat, is it gone straight away?
It leaves your list straight away, but not every copy goes at once. OpenAI and Anthropic say deleted chats are removed from their systems within 30 days, with exceptions for security or legal reasons. Google says Gemini chats that people have already reviewed are kept for up to three years even if you delete them.
How can I tell whether a call from a relative is a cloned voice?
You may not be able to tell by listening. Stop! Think Fraud suggests agreeing a safe phrase with close family and friends; if the caller can't give it, hang up. Then call the person back on a number you know is genuine.
Does turning off training stop a chatbot saving my chats?
No. Saving your history and using chats to train AI are separate settings. OpenAI says turning off Improve the model for everyone doesn't delete or hide your saved chats, and Google keeps Gemini history and training as separate choices. Use a temporary or incognito chat, or delete the chat, if you don't want it kept in your history.

Sources

Every setting, menu path and figure in this guide comes from these pages, read on 22 November 2026. Menus and prices change with updates: check the source before you rely on one.

  1. What is ChatGPT? (OpenAI)
  2. Data controls in ChatGPT (OpenAI)
  3. Temporary chat in ChatGPT (OpenAI)
  4. Deleting and archiving chats in ChatGPT (OpenAI)
  5. Does ChatGPT tell the truth? (OpenAI)
  6. Gemini Apps Privacy Hub, Last updated 17 November 2026 (Google)
  7. Manage and delete your history in Gemini Apps (Google)
  8. Use Gemini Apps (Google)
  9. Microsoft Copilot for individuals: your privacy controls and choices (Microsoft)
  10. Microsoft Copilot for individuals: your activity history (Microsoft)
  11. Transparency Note for Microsoft Copilot (for individuals) (Microsoft)
  12. Privacy FAQ for Microsoft Copilot (Microsoft)
  13. Is my data used for model training?, 16 March 2026 (Anthropic)
  14. How do I change my model improvement privacy settings?, 3 August 2026 (Anthropic)
  15. How long do you store my data?, 1 July 2026 (Anthropic)
  16. Use incognito chats (Anthropic)
  17. Claude is providing incorrect or misleading responses. What's going on?, 16 March 2026 (Anthropic)
  18. ChatGPT and large language models: what's the risk?, Blog post, published 14 March 2023 (NCSC)
  19. Phishing scams: If you've shared sensitive information, Version 2.0, reviewed 5 September 2022 (NCSC)
  20. Using AI to help write your data protection request or complaint, Published 4 November 2026 (ICO)
  21. Stop! Think Fraud: how to stay safe from scams (Stop! Think Fraud (UK Government))
  22. How to spot phone fraud (Stop! Think Fraud (UK Government))
  23. Recognise the tactics (Stop! Think Fraud (UK Government))
  24. Stay alert to fraud (Stop! Think Fraud (UK Government))
  25. Why you might be more at risk than you think (Stop! Think Fraud (UK Government))
  26. Reporting fraud (Stop! Think Fraud (UK Government))

Spotted something wrong or out of date? Tell us and we will correct it.

About this page

Dates

Published

Corrections

No corrections are listed for this page. Some review sweeps corrected pages without naming each one; those are in the corrections log.

Keep reading

More Everyday Tech guides

All Everyday Tech guides
Share this guideXLinkedInFacebookEmail

Was this useful?

Reports go to the site owner and are used to correct pages. Corrections are listed at /corrections.

All Everyday Tech guides

The KBY Brief

A short email about twice a week with new runbooks, configuration traps and fixes from KBY Technologies.