How do I stop ChatGPT from using my chats for training?

Open ChatGPT, click your profile icon, go to Settings → Data Controls and turn off 'Improve the model for everyone'. It is on by default for Free, Plus and Pro. Once off, new conversations are not used for model improvement or routine human quality review (the Project Lily program). Authorized staff can still access chats for abuse, support, troubleshooting or legal reasons. For a single sensitive conversation, use Temporary Chat instead.

Does opting out delete what the AI already learned from me?

No. Opting out stops future use. Text already incorporated into a trained model cannot be removed from it. Anthropic goes furthest: when you turn off 'Help improve our AI models' in Claude, it also stops using your previously stored conversations in future training runs. OpenAI and Google apply the setting going forward and delete stored chats on their normal schedules (30 days for cleared ChatGPT chats; 72 hours for Gemini chats with Keep Activity off).

Is the API or a business plan safer than the consumer app?

Yes, on training. The OpenAI API and ChatGPT Business/Enterprise, the Anthropic API and Claude Team/Enterprise, Gemini for Workspace and Vertex AI, and Perplexity Enterprise Pro do not use customer content for training by default. For the strictest guarantee, ask for zero data retention (ZDR) on the API, where prompts are not stored at all after the response is generated.

What is the fastest way to protect one sensitive conversation?

Use the private mode every major chatbot now has: ChatGPT Temporary Chat, Claude Incognito, Gemini Temporary Chat, Perplexity Incognito. These conversations are not saved to your history, do not feed memory and are excluded from training. They are still processed by the vendor to generate the answer and are usually kept briefly (Gemini says 72 hours) for safety and reliability.

Quick Answer

How to Stop AI Chatbots Training on Your Chats (2026 Guide)

Published: September 16, 2026

The short answer

Every major chatbot lets you opt out of training in two or three clicks, but the setting is on by default for consumers at OpenAI, Google and Perplexity, and Anthropic pre-selects “allow” on its choice screen. This guide gives the exact path for each product as of September 16, 2026, then explains what the toggle does and does not stop, because the 404 Media report on OpenAI’s Project Lily showed that “training” in practice means human contractors reading your conversations.

Step 1: Turn off training in each product

ChatGPT (OpenAI)

Profile icon (bottom-left on web; sidebar → profile on mobile).
Settings → Data Controls.
Turn off Improve the model for everyone.

Default: on for Free, Plus and Pro. Off by default for Business, Enterprise, Edu and the API.

Claude (Anthropic)

Click your name → Settings → Privacy.
Turn off Help improve our AI models.

Applies to Free, Pro and Max (consumer) plans and includes Claude Code sessions on those plans. Anthropic will also stop using your earlier stored conversations in future training runs. Retention drops from five years to 30 days. Team, Enterprise, API and cloud-marketplace traffic is never trained on by default.

Gemini (Google)

Go to myactivity.google.com/product/gemini (or profile icon → Gemini Apps Activity in the app).
Under Keep Activity, choose Turn off.
Optional: while Keep Activity is on, uncheck Improve Google services with your audio and Gemini Live videos & screenshares (off by default).
Review Settings → Connected Apps to limit what Gemini can read from Gmail, Drive, Calendar and Maps.

With Keep Activity off, new chats are kept 72 hours for operation and feedback, then deleted, and are not reviewed or used for training.

Perplexity

Settings → Preferences.
Turn off AI data retention.

Stops future data from being used for training; previously collected data is not removed.

Microsoft Copilot (consumer)

Sign in → profile → Privacy (or account.microsoft.com → Privacy).
Turn off Model training on text and Model training on voice.

Copilot in Microsoft 365 (work and school accounts) is not used for training.

Grok (xAI)

Grok app or grok.com → Settings → Data Controls.
Turn off Improve the model.
On X, also check Settings → Privacy and safety → Grok for the separate X-data setting.

Step 2: Understand what the toggle does and does not stop

Stopped by opting out	Not stopped by opting out
Use of new chats to train future models	Access to investigate abuse or a security incident
Routine quality review by contractors (e.g., OpenAI’s Project Lily)	Support and troubleshooting when you contact them
Feeding your chats into reviewer sampling pools (Gemini)	Legal holds and law-enforcement requests
Long retention tied to training (Claude: 5 years → 30 days)	Safety-classifier-flagged chats used for trust and safety (Claude)
	Anything already inside a trained model

Step 3: Use a private mode for sensitive one-offs

ChatGPT Temporary Chat — not saved, not used for training, does not use or update memory.
Claude Incognito — never used for training, not saved to history.
Gemini Temporary Chat — not saved to history or Activity, deleted after 72 hours.
Perplexity Incognito — not saved to your library.

Use these for medical, legal, financial and relationship questions even after you have opted out, because opting out does not cover the safety and support exceptions above.

Step 4: Clean up what is already stored

ChatGPT: Settings → Data Controls → Delete all chats. Deleted from OpenAI systems within 30 days unless already de-identified or legally held. Also review Personalization → Memory and delete stored memories; reviewers under Project Lily reportedly saw a “user memories summary.”
Claude: delete individual conversations or the whole history from Settings; opted-out data falls under the 30-day window.
Gemini: myactivity.google.com/product/gemini → delete by date range or Delete all time.
Perplexity: Settings → Delete all threads.

Step 5: For work or client data, move to a tier that never trains

Vendor	Never-trained option	Strongest option
OpenAI	ChatGPT Business, Enterprise, Edu; API	API with Zero Data Retention (eligibility required)
Anthropic	Claude Team, Enterprise; API; Bedrock/Vertex/Foundry	API with ZDR
Google	Gemini for Workspace; Vertex AI	Vertex AI with customer-managed encryption keys
Perplexity	Enterprise Pro	Enterprise Pro with retention disabled

See Zero data retention vs customer-managed keys vs self-hosted LLM (2026) for when each is worth it. If the requirement is that no vendor employee can ever read the text, only self-hosting satisfies it.

Step 6: Re-check quarterly

Defaults and menu paths change. Anthropic rewrote its consumer policy in September 2025; Google updated the Gemini Privacy Hub on August 10, 2026; OpenAI’s ads program and the DSA designation of ChatGPT as a very large online search engine (September 2026) may bring further changes. Put a 90-day reminder on it.