Do AI Chatbots Use Your Conversations for Training? How to Check

Direct answer

The short answer

This guide explains how to inspect published controls. It is not a guarantee about any service beyond its current documentation, and policies can change after the review date shown above.

The useful question is not only “Does this company train on chats?” Ask: “Which product am I using, what is the default, can I change it, does feedback create an exception, which provider receives the prompt, and what storage remains for non-training purposes?”

Definitions

Training is not the same as storage or human review

A statement such as “we do not train on your content” addresses one purpose. It does not by itself establish zero retention, no logging, no safety review, or no provider processing. Conversely, a short operational retention period does not necessarily mean content is used for training.

Checklist

Five things to check before you share a sensitive prompt

1. Product tierConsumer chat, business workspace, enterprise service, developer API, and third-party app may have different defaults.
2. Account controlLook for settings labeled model improvement, activity, training, history, or data controls. Record the selected state and date.
3. Temporary modeCheck what it changes—training, visible history, retention, or all three—and note any safety-retention exception.
4. Feedback exceptionRead the notice attached to ratings, reports, shared links, or support tickets before including a sensitive conversation.
5. Provider chainA third-party app or router can have its own policy while sending the prompt to a separate model provider with additional controls.

After those five, check deletion, backups, legal or security exceptions, regional differences, and whether human review is described. Save the URL and update date, not only a paraphrase.

Examples

Why product, tier, and setting change the answer

As of the review date for this article, OpenAI documents consumer data controls that can exclude new chats from model improvement and describes Temporary Chat separately. Anthropic publishes different explanations for consumer use and commercial products. Google’s Gemini privacy hub describes how its activity controls affect storage and model improvement. These are provider-specific policies, not universal rules.

The details can change and may include exceptions. A consumer account setting should not be assumed to govern an API integration. A business default should not be assumed for a free personal account. A temporary mode may still retain content for a stated safety period. Feedback can also be treated separately.

SayAll

How to evaluate SayAll's provider chain

SayAll sends prompts through its backend and OpenRouter to a selected model provider for response generation. SayAll does not write chat content to its database as server-side conversation history; the reopenable history remains in the current browser. Those facts do not make inference local and do not remove the need to consider provider processing.

OpenRouter’s privacy policy states that OpenRouter does not train on user inputs or outputs, while also explaining that requests are routed to model providers. Provider behavior can differ. The current SayAll Privacy Policy, OpenRouter policy, and relevant provider controls should therefore be read together.

Whatever service you choose, minimize first. Remove names, credentials, exact locations, private third-party messages, and other details the answer does not need. “Not used for training” is useful, but sharing less remains the most portable privacy control.