To use AI chatbots safely, do three things before anything else: keep passwords, ID numbers, health records and other people's private details out of your chats, switch off model training in the assistant's settings (or use its temporary or private chat mode for sensitive questions), and check every important answer against a real source before you act on it.
Below are eight concrete steps, with the exact menu names in ChatGPT, Claude, Gemini and Grok, what each private mode really does, and the newer risks from chatbots that act for you. Settings come from each company's help pages as of October 24, 2026.
How to use AI chatbots safely: the short checklist
- Decide what you will never type into a chatbot.
- Turn off model training, or confirm it is off.
- Use the private chat mode for one off sensitive questions.
- Delete old chats and review shared links.
- Secure your account with multifactor authentication or a passkey.
- Treat answers as drafts and verify anything that matters.
- Give agents and connected apps the least access possible.
- Download assistants only from official sources, and set up teen controls if your child uses one.
Step 1: Know what never to share
Every major assistant stores your chats for a while, even in private modes, so the safest data is data you never send. The Office of the Privacy Commissioner of Canada's advice on AI chatbots is a good baseline: limit what you share about yourself, avoid sensitive or identifiable details, swap real names for aliases, don't share other people's personal information, and never share pictures or personal information of minors.
The companies agree. Google's Gemini Apps Privacy Hub asks you not to enter confidential information you wouldn't want a human reviewer to see, and xAI's (now SpaceXSI) consumer FAQ tells Grok users not to share any personal or sensitive information.
In practice, keep these out of every chat:
- Passwords, recovery codes, API keys and one time codes
- Government ID, passport and social security numbers
- Full bank card or account numbers
- Medical records tied to your name
- Confidential work files
- Other people's private messages or photos
If a document contains any of this, redact it first: swap names for "Person A" and numbers for placeholders.
Step 2: Turn off model training
By default, consumer chats can be used to improve future models. Here is where the switch lives.
| Assistant | Where the training switch is | What changes when it is off |
|---|---|---|
| ChatGPT | Settings, Data controls, "Improve the model for everyone" | New chats are not used for training but stay in your history |
| Claude | Settings, Privacy, "Help improve Claude" | Chats are kept for 30 days instead of up to 5 years |
| Gemini | "Keep Activity" setting | Future chats leave your Activity and are kept for 72 hours |
| Grok | Settings, Data ("Improve the Model") on grok.com; Settings, Data Controls in the app | New conversations are not used for training |
A few details matter here:
- ChatGPT. OpenAI's data controls page says sharing is on by default for Free, Plus and Pro personal accounts, and the setting syncs across web and mobile. Its model training page adds a catch: a thumbs up or down can send that whole conversation to training even after you opt out.
- Claude. Anthropic's retention page says that with training on, chats may be kept in a format with your identity removed for up to five years. With it off, Anthropic keeps the standard 30 day period, which it set out in its consumer terms update.
- Gemini. Google says Keep Activity also controls whether a sample of your chats goes to human reviewers. Chats that reviewers have already seen are separated from your account and kept for up to three years, even if you later delete your activity.
- Grok. xAI's FAQ says feedback you choose to send may still be used for training after you opt out, and that in some regions outside the EU and UK, people using Grok without logging in have no opt out at all. Grok inside X follows X's own settings, not grok.com's.
For how the four compare beyond privacy, see our ChatGPT vs Claude vs Gemini vs Grok comparison.
Step 3: Use private mode for sensitive one offs
Each assistant has a mode for chats you don't want saved. They are useful, but none of them is truly off the record.
| Assistant | Mode | Saved to history? | Used for training? | How long it is kept |
|---|---|---|---|---|
| ChatGPT | Temporary Chat | No | No | Up to 30 days for safety |
| Claude | Incognito chat (ghost icon) | No | No | 30 days by default |
| Gemini | Temporary chat | No | No | 72 hours |
| Grok | Private Chat (ghost icon) | No | No | Deleted within 30 days |
Claude's incognito help page notes that incognito chats skip memory, only work outside projects, and can't be reopened once closed. On ChatGPT, saving a temporary chat turns it into a regular chat.
Pros and cons of private modes
Pros
- Keeps sensitive questions out of your history on a shared device
- Keeps the chat out of training on all four assistants
- Stops the chat from shaping memory and future answers
Cons
- The company still stores the chat for days or weeks for safety reasons
- You lose the conversation when you close it, so save anything you need first
- Some features are missing (Claude's incognito chats, for example, can't be used inside projects)
People miss the safety exception. Google says it still uses temporary chats to protect users and the public, including with human reviewers, and Anthropic says chats flagged by its safety classifiers can be used for trust and safety work even when training is off. Our guide to whether Claude reports chats to police explains how that review works across all four labs.
Step 4: Delete old chats and check shared links
Deleting a chat shortens how long it exists. Anthropic says a deleted Claude chat leaves your history at once and its back end systems within 30 days. xAI says the same 30 day window for Grok, with exceptions for safety, security and legal reasons.
Shared links need their own cleanup, because anyone with the link can read the chat. In July 2025 OpenAI removed a "Make this chat discoverable" checkbox after shared ChatGPT conversations, some of them deeply personal, showed up in Google results. OpenAI's security chief Dane Stuckey said the feature "introduced too many opportunities for folks to accidentally share things they didn't intend to," as The Register and Ars Technica reported. ChatGPT's data controls let you manage shared links; review them a few times a year.
Step 5: Lock down your account
Protect your chatbot account like your email. OpenAI's MFA help page covers adding an authenticator app and other second factors, and OpenAI also supports passkeys. Whichever assistant you use, turn on multifactor protection and never reuse a password.
Step 6: Verify answers before you act on them
Chatbots can sound certain and still be wrong. OpenAI's September 5, 2025 research post defines hallucinations as "plausible but false statements" and gives an example: asked for the title of one of its own researchers' PhD dissertation, a widely used chatbot gave three different answers, all wrong. The paper behind it (arXiv 2509.04664) argues that training and scoring reward guessing over admitting uncertainty, so the problem has not gone away as models improved.
What that means for you:
- Ask for sources, then open them. A link the chatbot names is not proof until you have read it.
- Use built in checks. Gemini's double check feature runs a Google search and highlights statements in green (similar content found) or orange (different or no content found).
- For health, legal, tax or money decisions, use the chatbot to prepare questions, then take them to a qualified professional.
Step 7: Give agents the least access possible
Assistants that browse websites, read your email or act on your behalf add a new risk called prompt injection. OpenAI's November 7, 2025 explainer describes it as hidden instructions in a web page, document or email that try to trick the AI into doing something you didn't ask for, such as sharing your files with an attacker. OpenAI calls it an open research problem.
OpenAI's tips work for any assistant:
- Connect only the apps and accounts a task actually needs, and use logged out mode for simple research.
- Read every confirmation prompt carefully before approving a purchase or a sent email.
- Watch the agent when it works on a sensitive site such as your bank.
- Give narrow instructions. "Reply to the email from my landlord" is safer than "handle my inbox."
Step 8: Avoid fakes and set up teen controls
In its May 2023 security report, Meta said it had found around 10 malware families posing as ChatGPT and similar tools since March of that year, including malicious browser extensions in official web stores, and had blocked more than 1,000 malicious links. Install assistants only from the company's own site or verified app store listing, and distrust extensions promising free access to paid models.
If a teenager in your home uses ChatGPT, OpenAI's parental controls, launched September 29, 2025 and also covered by Reuters, let a parent link accounts with the teen's agreement. Linked teen accounts get extra content protections, and parents can set quiet hours, turn off voice mode, memory and image generation, and opt the teen out of model training. OpenAI says trained staff review signs of acute distress and can alert parents, but also that guardrails are "not foolproof," so talking with your teen still matters.
Which assistant is safest for privacy?
There is no single winner. Claude's 30 day retention with training off is short, but turning training on means up to five years. Gemini keeps chats only 72 hours with Keep Activity off, yet chats already seen by reviewers can stay for three years. ChatGPT and Grok both hold private mode chats for up to 30 days. The safest one is the one whose settings you have checked. For the research side of AI risk, see our AI safety explainer; for Meta AI in WhatsApp, see our WhatsApp chatbot guide.
FAQ
Is it safe to tell ChatGPT personal information?
Only what you would be comfortable with OpenAI storing. Chats can be used for training unless you opt out, and even Temporary Chats may be kept for up to 30 days. Leave out passwords, ID numbers and medical or financial records.
How do I stop ChatGPT from training on my data?
Open Settings, then Data controls, and turn off "Improve the model for everyone." The choice applies to your whole account on web and mobile. Avoid rating sensitive answers with thumbs up or down, because feedback can send that conversation to training anyway.
What is the difference between temporary chat and incognito chat?
They are different names for similar features. ChatGPT and Gemini call it a temporary chat, Claude calls it an incognito chat, and Grok calls it Private Chat. All four keep the chat out of your history and out of training, but the company still stores it for a while: 72 hours on Gemini and up to 30 days on the others.
Can AI chatbot companies read my conversations?
Yes, in limited cases. Google says human reviewers read a subset of Gemini chats, and OpenAI says chats showing plans to harm others go to a trained review team. Turning off training reduces how your chats are used but does not switch off safety review.
Are AI chatbots safe for kids?
None is risk free. If your teen uses ChatGPT, link accounts through parental controls for extra content protections and settings such as quiet hours, and pair that with family rules, since OpenAI says the guardrails can be bypassed.
