Does AI train on your chats, and how do you opt out?

The short version is that on consumer plans, the default is participation. Your conversations are used to improve future models unless you go into settings and say otherwise, and buying a subscription does not change the default. The details differ enough per product that the general answer is not much use on its own.

Where each one stands

Claude. Anthropic used to be the exception here, with short retention and a clear line against training on consumer chats. That changed in August 2025, when the company updated its consumer terms. Free, Pro and Max users are asked to choose whether their chats and coding sessions train future models, and the setting defaults to on. Existing users had until October 8, 2025 to make a decision, after which making the selection became a condition of continuing to use the product. The retention consequence is the part that gets missed: allowing training extends how long that data is kept to as much as five years, while opting out keeps the prior thirty-day retention. The toggle lives in the privacy section of settings.

ChatGPT. OpenAI uses consumer conversations to improve its models by default. The control sits under Settings, then Data Controls, where the model-improvement option can be turned off. The important boundary is between products rather than tiers: content submitted through the API is not used for training by default, and business and enterprise agreements carry their own terms, while the consumer app defaults the other way on both free and paid plans.

Gemini. Google’s consumer default is also on, governed by the Gemini Apps Activity setting rather than a switch inside the chat interface. Turning off that activity stops future conversations from feeding training. Google has been explicit that a subset of conversations is reviewed by humans and that reviewed conversations are handled separately from the activity setting, so people should avoid entering data they would not want a reviewer to read.

What opting out does not do

Four limits apply to all three, and they matter more than the toggle.

It is not retroactive. The setting governs new and resumed conversations. Once data has gone into a completed training run, there is no mechanism to pull it back out of the resulting model weights. This makes the timing of the decision the whole decision.

It does not cover safety review. Anthropic’s privacy policy reserves the right to use inputs and outputs for model improvement where a conversation has been flagged for safety review, even for users who opted out. Comparable carve-outs exist across the industry. The category is narrow, but it is not empty.

It does not cover what you paste. The setting governs the conversation, not the source of the material in it. A contract, a customer list or someone else’s medical information pasted into a chat is covered by whatever the setting was at the time, and by whatever retention period applies.

It does not cover anything you published. Training corpora are built largely from the public web, and opting out of chat training says nothing about that. In August 2026, the entire contents of Cara, a platform built specifically as a refuge for artists opposed to AI training, were scraped for under ten dollars. The economics of collection are the reason platform-level promises are weak protection.

The pattern behind the defaults

The direction of travel across the industry has been one way: from opt-in to opt-out, and from short retention to long retention, usually announced as an update to consumer terms rather than as a change of position.

The reason is straightforward. Real conversations are the highest-value training data available, because they contain the tasks people actually bring, the follow-ups when the first answer misses, and the corrections. Synthetic data does not reproduce that. As frontier labs run into the limits of what the public web offers, the chat logs become the differentiated corpus, and the default setting is the cheapest way to acquire them.

The same logic is now visible in pricing. Meta launched a cheap Muse Code tier in August 2026 that trains on user prompts and code, making the trade explicit: a lower price in exchange for the data. Samsung Health asked users to allow AI training or lose sync. Those are honest versions of the arrangement, in that the exchange is stated. The consumer chat defaults are the quiet version of the same exchange.

For anyone handling work that belongs to someone else, the practical line is not the toggle but the product. Consumer assistants default to training and are the wrong place for client material, regardless of how the switch is set. API and enterprise agreements are where the contractual protection actually lives.

Related coverage

• Meta’s cheap Muse Code tier trains on your prompts and code, the trade stated as a price.
• Samsung Health asks you to allow AI training or lose sync, consent tied to functionality.
• Scraping all of anti-AI platform Cara cost under ten dollars, why platform-level promises are weak.
• Meta’s Muse Image can use your photos by default, the same default in an image product.
• Shared Claude conversations turned up in Google search results, a reminder that training is not the only exposure.
• Claude vs ChatGPT vs Gemini: how to choose in 2026, the rest of the comparison.