Using AI safely comes down to four habits: be careful what you type, check who is really asking, limit what an AI can do on its own, and check what it tells you. None of them needs technical skill. This page pulls the site’s safety pages together in plain words and ends with a checklist.
- Your privacy: what you type goes to the company that runs the AI. Treat it like a postcard, not a sealed letter.
- Scams: a voice or a face on a call is no longer proof. Agree a family code word, and call back on a number you already have.
- Agents: an AI that acts for you is a helper holding your keys. Make it ask before it pays, sends or deletes anything.
- Checking: AI can be confidently wrong, and it tends to agree with you. Check names, numbers, quotes and sources.
- Children: each app sets its own age rule, and parental controls do not show you the chats. Talking matters more.
Your privacy: what happens to what you type
The plain point: what you type into a chatbot is sent to the company that runs it. It may be stored, read by staff, or used to train future versions.
Prefer paper? Print the one-page cheat sheet.
The comparison: treat a chat like a postcard, not a sealed letter. Postcards usually arrive safely. But anyone handling one could read it, so you would not write your bank PIN on it.
- Training. Some apps use your chats to improve their AI unless you switch it off. Meta says this is on by default in its Muse agent.1 Is Meta Muse safe to use? shows the three taps that switch it off.
- Memory. Some apps now remember you from one chat to the next. Anthropic says memory is on by default in Claude.2 So a new chat is not always a blank page, as how to write AI prompts notes.
Where the AI runs matters too. The same model can reach you through the company’s app, a copy you download and run yourself, or a copy run by a cloud company. DeepSeek’s privacy policy says it stores personal data in China.3 A downloaded copy sends nothing back. Is DeepSeek, Kimi or Qwen safe to use? explains the three routes.
“Is it private?” is really four questions: is it kept, is it used for training, can a person read it, and was it yours to share? What happens to what you type takes them one at a time. The last one is fully in your hands.
Switch off training you do not want, check whether memory is on, and keep passwords, ID numbers and other people’s details out of every chat.
What never to tell an AI chatbot
The plain point: leave out anything that could hurt you, or someone else, if a stranger at the company read it or it leaked.
Canada’s privacy regulator puts the rule in a few words: “do not share sensitive or identifiable personal information”, “do not share personal information about other people”, and “never share pictures or personal information of minors”.4 The companies say much the same. Anthropic’s help page for Claude names financial information, health records, passwords and confidential documents,5 and Google asks Gemini users not to enter anything confidential they would not want a reviewer to see.6
Put together, that makes a short list to keep out of every chat:
- Passwords and login codes, including the one-time codes sent to your phone.
- Card and bank numbers. No chatbot needs them to help you budget.
- ID numbers, such as a social security, passport or driving licence number.
- Medical details and health records. Ask the general question, without your name or file attached.
- Work secrets: client files, contracts, unreleased plans, anything your employer has not cleared for AI.
- Other people’s details, and anything at all about children, including their photos.
The comparison: it is like talking on a busy train. You can still ask for directions; you just leave out your PIN and your neighbour’s address. If you need help with something personal, change the names and details first, as the Canadian regulator suggests.
Ask your question without the private details. The answer is almost always just as useful.
How to turn off training and use temporary chats
The plain point: each big chatbot has two switches worth knowing. One stops your future chats being used to train the AI. The other opens a temporary chat that is kept out of your history and not used for training.
- ChatGPT: go to Settings, then Data controls, and turn off “Improve the model for everyone”. For a one-off private chat, open a new chat, select Temporary, and choose Personalized or Unpersonalized before you send your first message (Unpersonalized leaves out your memories and custom instructions). OpenAI may still keep a copy for up to 30 days for safety.78
- Claude: select your name in the menu, then Settings, then Privacy, and switch off “Help Improve our AI models”. For a one-off private chat, start a new chat outside a project and click the ghost icon; an incognito chat is not used for training and is kept for 30 days by default.910
- Gemini: go to Settings & help, then Activity, and turn off Keep Activity, which is on by default for anyone 18 or over. For a one-off private chat, click Temporary chat next to New chat; Google does not use it to train its AI and keeps it for 72 hours.111213
Two limits to know. Turning off training does not delete the chats you already have; OpenAI says you delete those separately.7 And it works going forward: Anthropic says chats already used in training stay in models that have already been trained.9 Menus move, so if a step does not match your screen, follow the help page linked for that app.
Set the training switch once in every chatbot you use, and open a temporary or incognito chat for anything you would rather not keep.
Scams, fake voices and fake videos
The plain point: AI did not invent new scams. It made old ones cheaper and more convincing.
A voice can be copied from a very short clip. Microsoft researchers showed a model that copies a voice from about three seconds of recording.14
Faces are no safer. In early 2024 a finance worker at the engineering firm Arup joined a video call where every “colleague” was a deepfake, an AI-made copy of a real person. He then sent about US$25.6 million.15
The comparison: a scam is a play in three acts. Act one is a trusted face or voice. Act two is a reason not to check: hurry, or secrecy. Act three is a payment that cannot be undone, such as gift cards, crypto or a wire transfer. AI made act one cheap. It did nothing to acts two and three, and that is where you stop it.
- Agree a family code word. Say it out loud. Never post it or write it in a chat.
- Hang up and call back on the number saved in your phone, never one the caller gives you.
- Treat hurry plus secrecy as the alarm, whatever the voice sounds like.
- Judge a message by what it asks for, not by its grammar.
How AI fraud works shows seven common scams. Deepfake self-defence and the printable household guide set up the code word. How to spot AI-generated images, video and voices and cloning a voice or a face cover the rest.
Do not try to hear or spot the fake. Follow a rule instead: code word, callback, and a pause before any payment.
Agents: an AI that acts for you
The plain point: a chatbot gives advice. An agent acts: it opens websites, fills in forms, sends emails or buys things for you.
The comparison: a chatbot is a friend on the phone. An agent is a helper holding your house keys and bank card. The helper is honest, but it also follows any note it finds lying around.
Those notes are called prompt injection: hidden instructions in a web page, email or document that try to redirect the agent. It reads them as words, just like your request, and cannot reliably tell them apart. Meta, which makes the Muse agent, says this is unsolved.16
| Action | Think of it as | The rule |
|---|---|---|
| Readingpages, files, inbox | Think of it as: looking through the window | Rule: Usually fine. Keep private records out of reach. |
| Changingediting, deleting | Think of it as: moving the furniture | Rule: Only with a backup you have actually tried. |
| Sendingemail, payments, posts | Think of it as: posting a letter you cannot get back | Rule: A person says yes, every time. |
- Give it the fewest keys the job needs, and switch off any “always allow” for payments and sending.
- Keep a browsing agent away from your valuables: a separate browser profile, not signed into your bank or main email.
- Watch what it does, not what it says. Its commentary comes from the same AI that may have been tricked.
Agent guardrails, prompt injection, letting an AI browse for you and permission tiers go deeper.
Advice you can ignore. Actions have already happened. Anything an agent cannot undo should wait for your yes.
Checking what AI tells you
The plain point: AI can say something false with total confidence, and it leans towards agreeing with you.
It guesses because guessing pays: training and testing reward a confident guess over “I don’t know”.17 The comparison: a student in a test where a blank scores zero learns never to leave a blank.
It also bends when you push. In one study, pushing back moved a wrong answer to a right one in 43.52% of cases, but moved a right answer to a wrong one in 14.66%. So an AI changing its mind because you objected is not proof it was wrong.18 Think of an eager new employee who will not contradict the boss, whether the boss is right or not.
Skipping the check has cost people. In 2023 a US judge sanctioned two lawyers who “submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT”.19
- Open every source it gives you. How to check an AI citation takes a minute.
- Check names, numbers, dates and quotes. Why does AI hallucinate? explains why errors gather there.
- Ask neutrally, and keep your own view back until it has answered. See how AI influences you back.
- Find the original before sharing a screenshot. How to check a viral screenshot gives five checks.
Use AI like a clever friend who is sometimes confidently wrong and rarely argues. Check anything you will repeat, sign or act on.
Children and families
The plain point: each chatbot sets its own minimum age, and parental controls change settings rather than show you the chats.
The age rules differ. Anthropic’s consumer terms for Claude say 18; other services allow teenagers with conditions.20 OpenAI says linking a parent’s and a teen’s accounts does not show the parent the chats.21
The comparison: controls are a fence round the garden, not a camera in the house. They shape where a child can go, not what was said.
So the talk matters most: it can be wrong; it is not a friend; keep personal details out; and companion apps, built to feel like a relationship, are different. Google itself tells parents Gemini “should not be treated as a friend”.22 AI chatbots and kids has each service’s rules; AI companions covers the research.
Set up the controls together, then keep talking. A child who expects a calm reaction comes back when a chat feels wrong.
When not to use AI at all
The plain point: AI makes things worse when the effort was the point, when you cannot check the answer, or when a mistake lands on someone who trusted you.
The comparison: a calculator is great for the sums, and useless if the point was learning to add.
- Learning. The struggle is how the skill forms.
- Messages whose value is that you wrote them: a condolence, an apology.
- Health, money or legal decisions. Prepare with it; decide with a professional.
- Anything confidential. Once sent, it cannot be taken back.
Is the effort I am skipping a side task, or the task itself? If this is wrong, who pays? When not to use AI walks through nine cases; using AI at work covers the job version.
A safety checklist
Ten habits. Most take a minute once.
- Never type passwords, bank details, ID numbers or other people’s private information.
- Find the training setting in the AI you use most, and decide on purpose (steps above).
- Check whether memory is on.
- Agree a family code word, out loud, starting with the oldest relatives.
- Hang up and call back on a saved number when anyone asks for money or secrecy.
- Never pay in a hurry by gift card, crypto or wire transfer.
- Make agents ask before they pay, send, post or delete.
- Open every source an AI gives you before you repeat it.
- Look up the age rule for each app your child uses, and keep talking.
- Ask “who pays if this is wrong?” before relying on an answer.
Questions people ask
Is it safe to tell ChatGPT personal things?
Only what you would be comfortable with a stranger at the company reading. Unless you switch off “Improve the model for everyone”, OpenAI can use your new chats to train its models, so for anything sensitive, turn that off or use a temporary chat, and leave out names and numbers.7 For health, money or legal worries, a chatbot can help you prepare questions, but a professional should make the decision.
Can AI companies read my chats?
Sometimes, yes. Google says human reviewers, including staff at its service providers, read a subset of Gemini chats, and Anthropic says a small number of staff may review Claude chats you allow it to use for training, after removing your account details.135 Chats flagged by safety systems can be looked at whatever your settings, so the safest habit is still to leave private details out.135
Does deleting a chat delete it?
Not at once, and not always everywhere. OpenAI removes a deleted chat from your view straight away and schedules it for permanent deletion within 30 days, with exceptions for security and legal reasons; Anthropic says deleted Claude chats leave its back-end within 30 days; and Google keeps Gemini chats already seen by human reviewers for up to three years, even after you delete your activity.232413 And chats already used for training stay in the models already trained, Anthropic says of Claude.9
Where to go next
- “I want the full security picture.” Open the AI Security hub.
- “I want to protect my parents.” Print the household guide together.
- “I use AI at work.” Read AI at work: a beginner’s guide.
- “I am new to all of this.” Start with AI explained for complete beginners.
- “I want to think more clearly.” Try critical thinking and the Living with AI hub.
- Related: AI Chatbots and Mental Health.
- Related: Using AI to Budget.
- Related: AI for Older Adults.
- Related: Asking AI About Your Health.
- Related: How to Plan a Trip With AI.
- Related: Is AI Dangerous? The Real Risks, Sorted by How Sure We Are.
- Related: Is ChatGPT Safe to Use? Privacy, Accuracy and Scams.
- Related: What Is AI Ethics?.
- Related: Is This Safe to Paste Into ChatGPT? Free Privacy Checker.
- Related: Common Ways People Misuse AI, and What to Do Instead.
- Related: Turn Off AI Training in ChatGPT, Claude and Gemini.
- Related: Pasted Something Private Into ChatGPT? What to Do Now.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. A familiar voice calls in a hurry asking for money. What does the page say to do?
- Listen closely for signs that the voice is fake
- Hang up and call back on the number saved in your phone
- Pay a small amount first to see if it is real
Show the answer
Answer: Hang up and call back on the number saved in your phone. The page says not to try to spot the fake, but to follow a rule: code word, callback on a saved number, and a pause before any payment.
2. If an AI changes its answer after you push back, what does that show?
- Not much: it leans towards agreeing with you, so the change is not proof it was wrong
- That its first answer was definitely wrong
- That it has now checked a trusted source
Show the answer
Answer: Not much: it leans towards agreeing with you, so the change is not proof it was wrong. The page explains that AI tends to bend when you push, so changing its mind because you objected is not proof either way.
3. Does switching off the training setting delete the chats you already have?
- Yes, every old chat is wiped at once
- Yes, but only on paid plans
- No, it works going forward and old chats are deleted separately
Show the answer
Answer: No, it works going forward and old chats are deleted separately. The page says turning off training does not delete existing chats, and chats already used in training stay in models already trained.
- 1Meta Help Center, How to manage your Muse data, “Updated: 2 weeks ago”, read at source 1 Oct 2026“This setting is on when you first use Muse.” ↑
- 2Anthropic, Claude’s memory works everywhere, and you decide what’s in it, 25 Aug 2026, read at source 16 Sep 2026“Memory is on by default”. ↑
- 3DeepSeek, Privacy Policy, last updated 10 Feb 2026, read at source 1 Oct 2026“we directly collect, process and store your Personal Data in People’s Republic of China.” ↑
- 4Office of the Privacy Commissioner of Canada, Your privacy and AI chatbots, date modified 6 May 2025, read at source 5 Oct 2026“do not share sensitive or identifiable personal information”; “do not share personal information about other people”; “never share pictures or personal information of minors”; “alter personal details, for example, use aliases for names, change any identifiable details”. ↑
- 5Claude Help Center, I would like to input sensitive data into my chats with Claude. Who can view my conversations?, 22 May 2026, read at source 5 Oct 2026“we also recommend being thoughtful about sharing highly sensitive personal details such as: Financial information (SSN, credit card numbers, bank account details); Health records or medical information; Passwords or private login credentials; Confidential business or personal documents”. ↑a“We automatically de-link your data from your user ID (like your email address) before any review.” and “Access is limited to a small number of personnel involved in model training.” ↑b“If our safety classifiers flag your conversations, they may still be used to improve our internal trust and safety models, detect harmful content, enforce our policies, or advance our safety research.” ↑c
- 6Google, Gemini Apps Privacy Hub, read at source 1 Oct 2026“Please don’t enter confidential information that you wouldn’t want a reviewer to see or Google to use to improve our services”. ↑
- 7OpenAI, Data controls in ChatGPT, “Updated: 7 days ago”, read at source 5 Oct 2026“Select Settings. Select Data controls. Select Improve the model for everyone, turn it off, and select Done.” ↑a“Does turning off model training delete my chats? No. Delete a chat or your account separately when you want to remove content.” ↑b“When Improve the model for everyone is off, your new conversations won’t be used to train OpenAI models.” ↑c
- 8OpenAI, Temporary chat in ChatGPT, “Updated: 23 days ago”, read in a browser 13 Oct 2026“Open a new chat.” “Select Temporary.” “Choose Personalized or Unpersonalized before you send your first message.”; “Unpersonalized: Does not use memory, custom instructions, or plugins.”; “A temporary chat is not used to improve OpenAI models while it remains temporary”; “OpenAI may keep a copy of a temporary chat for up to 30 days for safety purposes.” ↑
- 9Anthropic Privacy Center, How do I change my model improvement privacy settings?, 4 Aug 2026, read in a browser 13 Oct 2026“Select your name from your settings menu and click the drop down button”, “Select Settings”, “Select Privacy”, “Under Help Improve our AI models, click the button to toggle it off/on”. ↑a“Your data will still be included in model training that has already started and in models that have already been trained”. ↑b“Your data will still be included in model training that has already started and in models that have already been trained”. ↑c
- 10Claude Help Center, Use incognito chats, “Updated over a week ago”, read in a browser 13 Oct 2026“When starting a new chat with Claude outside of a project, you’ll see a ghost icon in the upper right corner of your screen”; “Click the ghost icon to enable incognito mode.”; “Incognito chats are not used for training.”; “they are retained for either 30 days (default), or longer in accordance with your organization’s custom data retention setting”. ↑
- 11Google, Manage & delete your activity in Gemini Apps, read at source 5 Oct 2026“If you’re 18 or over, Keep Activity is on by default.”; the steps are “Click Settings & help”, then “Activity”, then “Turn off” or “Turn off and delete activity”. ↑
- 12Google, Use Gemini Apps, read at source 5 Oct 2026“Next to New chat, click Temporary chat”; the feature needs “A personal Google Account”. ↑
- 13Google, Gemini Apps Privacy Hub, last updated 24 Sep 2026, read at source 5 Oct 2026“Temporary chats and chats you have when Keep Activity is off are retained with your account for 72 hours”; “Temporary chats are not used to train Google’s AI models.” ↑a“A subset of chats are reviewed by human reviewers (including Google’s trained service providers) to help improve Google services.” ↑b“Even if your Keep Activity setting is off or you use temporary chats, Google still uses your chats to respond to you and help protect Google, our users, and the public, including with help from human reviewers.” ↑c“If you delete your activity, past chats already reviewed by service providers aren’t deleted because they aren’t connected to your Google Account. Instead, they’re retained for up to 3 years.” ↑d
- 14Microsoft Research, VALL-E, Jan 2023, read at source 10 Sep 2026“only a 3-second enrolled recording of an unseen speaker”. ↑
- 15CNN Business, report on the Arup case, 16 May 2024, read at source 17 Sep 2026“He subsequently agreed to send a total of 200 million Hong Kong dollars — about $25.6 million.” ↑
- 16Meta AI Research, How We Built Safety Into Muse, 8 Sep 2026, read at source 16 Sep 2026“Prompt injection remains an open problem in the industry”. ↑
- 17Kalai et al., Why Language Models Hallucinate, 4 Sep 2025, read at source 1 Oct 2026“the training and evaluation procedures reward guessing over acknowledging uncertainty”. ↑
- 18Stanford researchers, SycEval, Feb 2025, read at source 1 Oct 2026“Progressive sycophancy, leading to correct answers, occurred in 43.52% of cases, while regressive sycophancy, leading to incorrect answers, was observed in 14.66%.” ↑
- 19Mata v. Avianca, S.D.N.Y., Opinion and Order on Sanctions, 22 June 2023, read at source 16 Sep 2026.↑
- 20Anthropic, Consumer Terms of Service, effective 8 October 2025, read at source 18 Sep 2026“You must be at least 18 years old”. ↑
- 21OpenAI Help Center, Managing parental controls in ChatGPT, read at source 18 Sep 2026“Linking accounts does not give a parent or guardian access to a teen’s conversations”. ↑
- 22Google, Guide your child’s Gemini Apps experience, read at source 18 Sep 2026.↑
- 23OpenAI, Chat and file retention in ChatGPT, “Updated: 13 days ago”, read at source 5 Oct 2026“ChatGPT removes it from your account view immediately.” and “OpenAI schedules it for permanent deletion from its systems within 30 days, unless the chat was already de-identified and disassociated from your account or OpenAI must retain it longer for security or legal obligations.” ↑
- 24Anthropic, Privacy Policy, effective 10 Sep 2026, read at source 1 Oct 2026 (as cited on this site’s ChatGPT vs Claude vs Gemini page): deleted conversations are “automatically deleted from our back-end within 30 days”.↑
Meta Help Center, How to manage your Muse data, read 1 Oct 2026 · Meta AI Research, How We Built Safety Into Muse, read 16 Sep 2026 · Anthropic, memory announcement, read 16 Sep 2026, and Consumer Terms, read 18 Sep 2026 · DeepSeek, Privacy Policy, read 1 Oct 2026.
Microsoft Research, VALL-E, read 10 Sep 2026 · CNN Business, Arup report, read 17 Sep 2026 · Kalai et al., Why Language Models Hallucinate, read 1 Oct 2026 · SycEval, read 1 Oct 2026 · Mata v. Avianca, read 16 Sep 2026 · OpenAI, parental controls, and Google, Gemini guide for parents, both read 18 Sep 2026.
Read dates are those recorded on the linked detail pages of this site. The comparisons on this page (postcard, busy train, three-act play, helper with your keys, test-taker, eager employee, garden fence, calculator) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK