Is AI Safe? Personal Data and Reliability

Are you hesitating to use an AI because you worry about your data, or because you fear it will get things wrong? These concerns are reasonable, and a few simple steps can reduce them. This guide applies to all AIs: none is considered safer or riskier than another.

Don’t Type Just Anything

The CNIL (the French authority that protects personal data) gives clear advice.

• Avoid entering personal data as much as possible.
• Never share confidential information: personal data, company or government data, or information covered by a secret.
• In consumer online tools, what you type may be used to train the AI. So don’t enter anything confidential, personal, or strategic.
• For sensitive data (people, contracts, finances): don’t enter it, or anonymize it. Anonymizing means removing anything that could identify someone.

Replace Details with [Brackets]

Before sending your text, replace each personal detail with a word in brackets. The AI works with the template, and you add the real details yourself afterward. The method for writing a good request is explained in Write a Good Prompt.

Example request Write a letter disputing an invoice. The recipient is [supplier name], the invoice number is [invoice number], the disputed amount is [amount], and the reason is [one-sentence explanation]. Polite but firm tone.

Know the Settings of Each AI

Names and menu paths change regularly. The labels below are the ones cited by official help pages and by the CNIL. The CNIL also explains these settings on its page Opting Out of Your Data Being Reused.

AI	Refuse use of my conversations for training	Delete a conversation	Temporary conversation
ChatGPT	Settings, “Data controls” (the CNIL calls it “Gestion des données” in French), then turn off “Improve the model for everyone” (“Améliorer le modèle pour tous” according to the CNIL) and confirm with “Done.”	“•••” menu on the conversation, then “Delete.” “Delete all chats” also exists and cannot be undone. The conversation disappears right away; its permanent deletion from the systems is scheduled within 30 days, with some exceptions.	“Temporary chat,” “Temporary” button. It stays out of your history and is not used to improve the models. A copy may be kept for up to 30 days for security.
Claude	Your name, “Settings,” “Privacy,” then the toggle under “Help Improve Claude” (the CNIL says to turn off “Aider à améliorer Claude”). Data already used for training is not erased.	On the web, hover over the conversation and click “⋮.” It disappears from your history immediately and from internal systems within 30 days. To delete in bulk: “Recents,” “View all,” tick the boxes, then “Delete Selected.”	“Incognito chat” (ghost icon): no history, no memory for Claude. Kept for 30 days for security.
Copilot	According to the CNIL: account settings, “Privacy” tab, then turn off “Model training on text” and “Model training on voice.”	In the “Chats” list on the left, “…More,” then “Delete.” For your whole history: Microsoft privacy dashboard, then “Delete all activity history.”	See the official help: Copilot.
Gemini	According to the CNIL: “Activity” tab, then “Turn off and delete activity.” This may delete past conversations, and future ones will no longer be saved.	You can delete your conversations at any time from Activity in the Gemini apps.	Temporary chat, with a personal Google account only (not a work or school account). It does not appear in Activity and is not used to train Google’s models.
Mistral	Website: according to the CNIL, “Settings,” “Preferences” tab in the “Le Chat” section, then uncheck “Allow use of your interactions to train our models.” Mistral’s help page describes another path (“Admin Console,” “Privacy” section). Mobile: “Settings,” “Data and account control,” then turn off “Enable data sharing.”	Three vertical dots to the right of the title, “Delete,” then confirm. Cannot be recovered; some data may remain in the systems depending on the privacy policy.	See the official help: Mistral.
Perplexity	The “AI data retention” setting is on by default (Free, Pro, Max); turning it off lets you opt out. Another help page calls it “AI Data Usage,” under Settings.	From your library (perplexity.ai/library), three dots next to the thread. To clear everything: three dots at the top right, “Delete all threads” (cannot be undone).	Incognito sessions, which expire after 24 hours. See the official help: Perplexity.

On ChatGPT, turning off training does not delete your history, and it only applies to new conversations. On Gemini, even with activity turned off or in a temporary chat, Google still uses conversations to generate responses and, with human reviewers, to protect its services.

Understand Why an AI Can Be Wrong

According to the CNIL, a generative AI can produce inaccurate answers that still sound plausible. These are called “hallucinations.” The risk is higher for recent events, because these tools are not always connected to up-to-date knowledge bases. Trusting an answer too much, without checking it, can lead to poor decisions.

Check Any Important Answer

• The final decision is yours: step back and question the answers.
• Cross-check with an official source.
• Choose a tool that cites its sources, as the CNIL recommends. The full method is in Find Reliable Information with AI.
• For health, money, or legal matters, consult an official website or a professional.

Example request Explain [topic] to me. Cite your sources with a link for each important piece of information. If you are not sure about a point, say so clearly instead of guessing.

To find the AI that suits your needs, you can try the free questionnaire.

Good Habits

• Don’t enter personal data or confidential information; replace it with [brackets].
• Check your AI’s settings and delete conversations you no longer need.
• Verify any important answer with an official source.
• Ask the AI for its sources and read them.
• Always reread before sending or using a text: you make the final call.