AI Chatbots That Don’t Use Chats for Training
Several AI chatbots say they do not use consumer conversations to train their models by default, including Claude, ChatGPT with its model-improvement setting turned off, and Gemini with its activity and training controls configured appropriately. The exact rule depends on the provider, account type, settings, and how long data is retained, so “not used for training” does not necessarily mean “not stored or reviewed.”
What “not used for training” means
Training is the process of using data to change a model’s parameters so it can produce better responses later. A company can exclude chats from that process while still processing and temporarily retaining them to provide the service, detect abuse, troubleshoot failures, or meet other operational needs. These are separate practices, and privacy policies often describe them separately.
Also distinguish model training from features such as chat history, memory, and product feedback. Saving a conversation so you can reopen it is not itself proof that it is being used to train a model. Conversely, deleting a chat from your visible history may not instantly remove every retained copy from service systems or backups. Check the provider’s current policy for its retention and deletion rules.
Chatbots and their default policies
Policies change, and some differ between free consumer accounts, paid plans, and business or API services. The following is a practical starting point, not a guarantee for every account or future policy revision.
- Claude: Anthropic says consumer conversations are not used to train its models by default. It may use conversations when a user chooses to provide feedback, and its policies describe exceptions related to safety and policy enforcement. Review the current privacy settings and terms before sharing sensitive material.
- ChatGPT: OpenAI may use content from individual consumer services to improve models unless the user turns off the relevant model-improvement control in Data Controls. Temporary Chat is also described as not being used to improve models, though it may be retained for a limited period for safety purposes. Business products and API use have different default terms.
- Google Gemini: Google’s consumer Gemini services may use activity to improve services, including model training, depending on the user’s activity setting and the applicable product terms. Turning off Gemini Apps Activity affects future activity use, but Google’s explanations note that some conversations may be reviewed and retained for a period for service safety. Do not treat the setting as an across-the-board promise that no human can review a conversation.
- Microsoft Copilot: Microsoft’s consumer Copilot privacy disclosures distinguish conversation data and optional personalization from model training. The treatment can vary by Copilot experience and account, so consult the privacy controls and current terms for the specific version. Organizational Microsoft 365 Copilot is governed by separate commercial commitments.
- Perplexity: Perplexity provides an AI data-retention or AI data-usage control for whether search and assistant interactions may be used to improve its AI systems. Check that control in your account rather than assuming a universal default across account types or product changes.
These examples illustrate why the useful question is not simply “Which chatbot?” but “Which product, under which account, with which controls enabled?” For a workplace account, the organization’s contract and administrator settings may override the consumer experience.
How to check and change the setting
Before entering private information, open the service’s privacy, data controls, or activity settings. Look for wording such as “improve the model,” “help improve AI,” “training,” or “conversation history.” Labels and menu locations change, so use the provider’s current help page if a setting is hard to find.
- Identify the exact service and account type: consumer, business, education, or API.
- Find the control for model improvement or AI training; do not confuse it with chat history or memory.
- Turn off optional training or data sharing, then confirm whether the setting applies to future chats only.
- Review temporary-chat options, deletion controls, and retention explanations separately.
- Recheck after major product changes or when switching accounts or devices.
If you submit a thumbs-up/down rating or report a response, read the accompanying notice. Feedback may be handled differently from ordinary conversations and can include the prompt or response for review.
Consumer, business, and local options
Business and API services often offer stronger contractual limits on training than consumer chatbots. For example, OpenAI states that business and API data is not used to train models by default; Anthropic makes a similar no-training commitment for commercial products, subject to its terms. Google’s paid workspace offerings also have distinct data protections. These statements apply to specified products and terms, not automatically to every service from the same company.
For confidential work, an organization should verify the actual agreement, retention period, access controls, and administrator options rather than relying on a marketing summary. A locally run model can keep prompts on a device or organization-controlled server, but only if the software, telemetry, hosting, and connected tools are configured accordingly. Local use is not automatically private, and it may require technical setup and capable hardware.
Choosing a safer option
Choose based on the sensitivity of the information and the protection you need. For ordinary questions, turning off optional training may be sufficient. For personal or business-confidential data, examine retention, human review, account access, and contractual protections as well as training.
- Do not paste passwords, authentication codes, financial account details, private keys, or another person’s sensitive information into a chatbot.
- Remove names, contact details, and identifying facts when they are not needed to answer the question.
- Use an approved business service for workplace information, and follow your organization’s rules.
- Check connected apps and tools: a chatbot may pass information to an external service even if its base model does not train on chats.
No setting can make a chatbot a confidential channel in every circumstance. Providers may retain data for service operation or safety, and legal or security obligations can also affect handling. Read the current policy before relying on a particular protection.
Frequently Asked Questions
Does turning off chat history stop training?
Not necessarily. Chat history controls whether conversations are saved or displayed; model-improvement controls govern training use. Check both settings and the provider’s explanation.
Are temporary chats completely private?
No. Temporary or incognito-style chats may be excluded from training and not appear in history, yet still be retained briefly for safety or operations. “Temporary” does not mean zero processing or immediate deletion.
Can employees or reviewers read my chats?
Some providers permit limited human review for safety, support, or policy enforcement, and feedback submissions may be reviewed. The scope depends on the service and its terms; do not assume a no-training policy means no review.
Does deleting a conversation remove it from training data?
Deletion controls may remove a chat from your account and trigger deletion under the provider’s retention policy, but they do not necessarily instantly erase backups or undo prior model training. Check the service’s deletion documentation.
Which option is best for confidential work?
Use a business or API service whose written terms meet your organization’s requirements, or an appropriately configured local system. Verify training, retention, review, and access provisions instead of relying on the product name alone.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.