Which AI chatbot does not use my conversations for training by default?

As of September 2026, none of the four big consumer chatbots is training-off by default on free plans. ChatGPT's 'Improve the model for everyone' is on for Free, Plus and Pro. Gemini's 'Keep Activity' is on by default and feeds human review and training. Claude asks consumer users (Free, Pro, Max) to choose at signup or in a prompt, with the toggle pre-set to allow training. Perplexity's 'AI data retention' is on by default. Business tiers and APIs are the exception: ChatGPT Business/Enterprise, the OpenAI API, Claude Team/Enterprise and the Anthropic API, and Gemini for Workspace do not train on customer data by default.

Do humans actually read ChatGPT, Claude and Gemini conversations?

Yes, at all three. 404 Media reported on September 14, 2026 that OpenAI's Project Lily employs hundreds of contractors to read and grade real ChatGPT chats. Anthropic confirmed it also uses human review to improve Claude. Google's Gemini Apps Privacy Hub states that human reviewers, including service providers, read a sample of conversations, which are disconnected from the account and kept up to three years. Opting out of training stops routine quality review; it does not stop access for abuse investigations, support or legal requests.

Which chatbot has the shortest data retention if I opt out?

Gemini: with Keep Activity off, chats are kept only 72 hours for service operation and feedback processing. Claude keeps opted-out consumer conversations 30 days (versus five years if you allow training). ChatGPT deletes cleared chats within 30 days unless already de-identified or held for legal reasons; Temporary Chats are not saved to history. Perplexity does not publish a fixed window for opted-out data beyond stating it is not used for training.

Which is the most private AI chatbot for sensitive questions in 2026?

For a one-off sensitive question, use a mode that is excluded from training on any of them: ChatGPT Temporary Chat, Claude Incognito, Gemini Temporary Chat. For ongoing use, Claude with training off (30-day retention, explicit choice, Team/Enterprise never trained) and Gemini with Keep Activity off (72-hour retention) offer the tightest published retention. If the requirement is that no vendor employee can ever read the text, no consumer chatbot qualifies; that needs an API with zero data retention or a self-hosted model.

Quick Answer

Who Reads Your AI Chats? ChatGPT vs Claude vs Gemini (2026)

Published: September 16, 2026

The short answer

All four read a sample of your chats to improve their models unless you turn it off, and none is training-off by default on a free consumer plan. The differences are in how loudly they ask, how long they keep the data and what the opt-out actually stops. Verified September 16, 2026, two days after 404 Media’s report on OpenAI’s Project Lily.

	ChatGPT (OpenAI)	Claude (Anthropic)	Gemini (Google)	Perplexity
Training on consumer chats by default	On (“Improve the model for everyone”) for Free, Plus, Pro	Choice screen; toggle pre-set to allow, applies to Free, Pro, Max	On (“Keep Activity”)	On (“AI data retention”)
Human review disclosed	Yes; Project Lily contractors rate replies 1-7 (404 Media, Sep 14, 2026)	Yes; confirmed to 404 Media	Yes; reviewers incl. service providers, Privacy Hub updated Aug 10, 2026	Policy permits processing “to improve the service”
Opt-out path	Settings → Data Controls → off	Settings → Privacy → “Help improve our AI models” off	myactivity.google.com/product/gemini → Keep Activity off	Settings → Preferences → AI data retention off
Retention when opted out	Cleared chats deleted within 30 days	30 days (5 years if opted in)	72 hours	Not specified
Reviewed-data retention	Not published	Not published	Up to 3 years, disconnected from account	Not published
Private mode	Temporary Chat	Incognito	Temporary Chat (72 h)	Incognito
Never-trained tiers	Business, Enterprise, Edu, API	Team, Enterprise, API, Bedrock/Vertex/Foundry	Workspace, Vertex AI	Enterprise Pro
Exception that survives opt-out	Abuse, support, troubleshooting, legal	Safety-classifier-flagged chats used for trust & safety	Abuse, legal, service operation	Legal obligations, service operation

ChatGPT: biggest scale, quietest default

OpenAI’s Data Usage FAQ says authorized personnel and “trusted service providers” may access content to investigate abuse, support you, handle legal matters or “improve model performance (unless you have opted out).” Project Lily is the improvement part: hundreds of contractors, recruited through Crossing Hurdles and paid via Mercor, reading real conversations with usernames removed and a “Privacy Filter” model applied, but with OpenAI admitting sensitive details can get through. With more than 900 million weekly users and the setting on by default, ChatGPT has the largest pool of reviewable conversations of any product. Full explainer: What is OpenAI’s Project Lily?

Claude: explicit choice, shortest opt-out retention among chat apps

Anthropic changed its consumer terms in September 2025 so that Free, Pro and Max users must decide whether conversations and Claude Code sessions can be used for training; the toggle appears at signup and can be changed at Settings → Privacy → Help improve our AI models. Opting out means new chats are not used, and Anthropic also stops using previously stored conversations in future training runs (data already inside a trained model cannot be pulled back out). Retention is 30 days when opted out and five years when opted in. Two caveats: conversations flagged by safety classifiers may still be used to improve trust-and-safety systems, and thumbs-up/down feedback can send that exchange to reviewers. Commercial products (Team, Enterprise, API, cloud marketplaces) are excluded from training by default.

Gemini: clearest disclosure, longest reviewed-data retention

Google’s Gemini Apps Privacy Hub is unusually direct: human reviewers, “including Google’s trained service providers,” read a sample of conversations to improve Gemini models, and data they review is disconnected from your account and saved for up to three years. Control is the Keep Activity setting; with it off, chats are kept 72 hours so Gemini can respond, process feedback and protect users. Audio, Gemini Live video and screen-shares are not used for improvement by default and have a separate checkbox. Temporary Chats disappear after 72 hours and are not used for training. Because Gemini can also reach Gmail, Drive and Calendar through Connected Apps, check Settings → Connected Apps as well.

Perplexity: one switch, thinner documentation

Perplexity’s opt-out is Settings → Preferences → AI data retention. Its help center says data collected after opting out is not used to train models, that earlier data is not removed, and that data may still be processed to operate and improve the service or meet legal obligations. It publishes no retention window for opted-out conversations and no description of human review, which makes it the least transparent of the four rather than necessarily the most or least private.

Which to use, by need

If you need…	Pick
A sensitive one-off question	Any private mode: ChatGPT Temporary Chat, Claude Incognito, Gemini Temporary Chat
Ongoing personal use with the least retained	Gemini with Keep Activity off (72 h) or Claude opted out (30 days)
A plain “did I consent?” record	Claude, the only one with an explicit consumer choice screen
Work data	Any business tier or API; none trains by default
Nobody at the vendor can ever read it	Not a consumer chatbot: API with zero data retention, or self-hosted — see ZDR vs customer-managed keys vs self-hosted LLM

What opting out never does

It does not stop safety, abuse, support or legal access at any vendor.
It does not remove your text from models already trained.
It does not cover data you send through third-party apps built on these models; their policies govern.