benchr
The practical problem is fragmentation: the control, default, retention period, and exceptions differ by provider, product, account type, region, and whether you submit feedback. A “training off” label answers only one question; it does not describe storage, abuse review, administrator access, connected apps, or third-party recipients.
The table below records what each cited provider states for the named consumer product as of the review date. It is a starting point, not a promise that the same rule applies to another plan, country, workspace, or integration.
Current consumer training controls
Match the row to the exact service and account you use, then verify the linked policy. Retention entries are simplified and can have separate exceptions for feedback, safety investigations, legal obligations, deleted chats, or administrator settings.
| Provider | Trains by default? | How to opt out | Retention |
|---|---|---|---|
| ChatGPT (Free, Plus, Pro) | May use content unless you opt out | Settings → Data Controls → turn off "Improve the model for everyone," or use Temporary Chat | Opt-out applies to new chats; Temporary Chat may be retained up to 30 days for safety |
| Claude (Free, Pro, Max) | Used only if you choose to allow model improvement | Privacy Settings → do not allow model improvement (changeable later) | 30-day standard if not allowed; 5 years if allowed, with separate policy exceptions |
| Gemini Apps (personal account) | Keep Activity on: activity may improve services and models with human review | Turn off Keep Activity or use Temporary Chat; feedback is a separate exception | Future off/temporary chats: 72 hours; reviewed activity or feedback: up to 3 years |
| Copilot (signed-in consumer) | Yes, unless you opt out | Account settings while signed in | Opt-out stops future training |
| Named business products and APIs | OpenAI Business/Enterprise/API and Anthropic commercial/API: no by default; Microsoft lists specific excluded accounts | Verify the exact product, workspace controls, and any feedback or data-sharing option | Product documentation, administrator settings, contract, and DPA govern retention |
ChatGPT: individual-service and business defaults differ
OpenAI says content from services for individuals, including ChatGPT, may be used to train models. The switch in Settings under Data Controls is labeled “Improve the model for everyone”; turning it off excludes new conversations from training. It cannot undo training that has already occurred, and feedback can separately include the associated conversation.
Temporary Chat is not used for training and does not appear in history or use or create memories, but it may be retained for up to 30 days for safety. OpenAI states that ChatGPT Business, ChatGPT Enterprise, and API inputs and outputs are not used for training by default unless an organization explicitly opts in to data sharing. Retention and access controls still depend on the product and workspace configuration.
Claude: a consumer choice with different retention
Anthropic's consumer terms introduced a choice for Claude Free, Pro, and Max users, including Claude Code on those accounts. If a user allows model improvement, new or resumed chats and coding sessions may be used and retained for five years. If the user does not allow it, Anthropic states that the standard 30-day retention continues. Separate exceptions can apply to feedback, policy violations, legal obligations, and deleted data.
Users can change the choice in Privacy Settings. Anthropic says the consumer change does not apply to services under its Commercial Terms, including Claude for Work, Government, Education, and API use through Anthropic or named third-party platforms. That is a training-use statement, not a complete security or confidentiality assessment; commercial customers should review their applicable terms and retention controls.
A training setting, retention period, and confidentiality promise are three different controls.
Gemini: the human-review wrinkle
Google's Gemini Apps Privacy Hub says a subset of chats can be reviewed by trained human reviewers to improve services and protect users. Activity already reviewed and disconnected from the account can be kept for up to three years and is not removed when the user deletes Gemini Apps Activity.
With Keep Activity off, Google says future chats are not used to train its AI models unless the user submits feedback; the chats are still retained with the account for 72 hours to provide and protect the service. Temporary chats follow the same 72-hour period and are not used for model training. Safety review and voluntarily submitted feedback have separate rules, so “off” does not mean immediate deletion or zero human access.
Copilot and Meta AI: the rest of the field
Microsoft says it may use signed-in consumer Copilot conversation activity, voice, and uploaded files for AI training unless the user opts out. Its current FAQ lists exclusions for organizational Entra ID accounts, Copilot inside Microsoft 365 Personal or Family apps, signed-out users, users under 18, opted-out users, and certain regions. An exclusion from model training does not by itself establish that an account is appropriate for confidential data; apply the organization's access, retention, and data-handling rules.
Meta's Privacy Center says interactions with AI at Meta features can be used to develop and improve its generative AI models, with rights and controls varying by region. Meta also says that, from December 16, 2025 in most regions, AI interactions may be used as a signal to personalize content and ads across accounts connected in the same Accounts Center. Check the current scope for the specific Meta product, account connection, and region before sharing sensitive information.
What to do next
Match the control to the exact product. In ChatGPT, review Data Controls or use Temporary Chat with its retention caveat. In Claude, choose the model-improvement setting in Privacy Settings. In Gemini, review Keep Activity and the feedback exception. In consumer Copilot, review the model-training control while signed in. Recheck the policy when the product, account type, region, or integration changes.
For sensitive data, do not rely on the training toggle alone. Use only an approved account and integration, minimize the data supplied, verify retention and human-access rules, check whether tools send content to third parties, and follow your organization's contract and security requirements. “Not trained on” is useful, but it is not a synonym for private, encrypted end to end, confidential, compliant, or immediately deleted.
Frequently asked
Does ChatGPT train on my conversations by default?
OpenAI says content from services for individuals such as ChatGPT may be used to improve its models unless you opt out. Review Settings → Data Controls → “Improve the model for everyone,” or use Temporary Chat, which is not used for training but may be retained up to 30 days for safety. OpenAI states that ChatGPT Business, Enterprise, and API data are not used for training by default unless an organization opts in. Check the exact product and current setting.
Did Claude change its privacy policy?
Yes. Anthropic introduced a consumer choice for Free, Pro, and Max users, including Claude Code on those accounts. If a user allows model improvement, new or resumed chats and coding sessions may be used and retained for five years; if not, Anthropic says the standard 30-day period continues. The choice can be changed in Privacy Settings. Anthropic states that services under its Commercial Terms are outside this consumer change, subject to their own terms and exceptions.
How do I opt out of training across providers?
Controls differ by product. In ChatGPT, review Settings → Data Controls and “Improve the model for everyone.” In Claude, use Privacy Settings to choose whether to allow model improvement. In Gemini Apps, turn off Keep Activity or use Temporary Chat, noting the feedback and 72-hour exceptions. In eligible consumer Copilot accounts, change the model-training control while signed in. These choices generally govern future use and cannot undo training already completed.
Does the API train on my data?
There is no provider-wide answer for every API. OpenAI says ChatGPT Business, Enterprise, and API inputs and outputs are not used for training by default unless an organization opts in. Anthropic says services under its Commercial Terms, including its API and named third-party platforms, are outside its consumer training change. Microsoft lists organizational Entra ID Copilot accounts among its training exclusions. Verify the exact service, retention, feedback, administrator, and contractual controls; no-training is not a complete privacy guarantee.
Is anything I type ever truly private?
No product tier is universally “truly private.” ChatGPT Temporary Chat is not used for training but may be retained up to 30 days for safety. Gemini Keep Activity off prevents future training absent feedback, but chats remain for 72 hours and safety review has separate rules. Business and API no-training defaults cover one use of data, not every retention, access, security, or third-party risk. Use only an approved product for sensitive information.
Can a human read my Gemini chats?
Google says a subset of Gemini Apps chats can be reviewed by trained human reviewers for improvement and protection. Reviewed activity disconnected from the account may be retained for up to three years. Turning off Keep Activity stops future chats from training unless you submit feedback, but chats remain for 72 hours to provide and protect the service; safety and feedback review have separate rules.
Changelog
- July 24, 2026 — Rechecked official policies; made every claim product-, setting-, and account-specific; removed motive claims and “privacy-safe” guarantees; separated training use from retention, human review, security, and compliance.
- May 30, 2026 — Originally published. Provider defaults, opt-out paths, and retention figures verified against Anthropic's consumer-terms post, Google's Gemini Apps Privacy Hub, Microsoft's Copilot privacy FAQ, and reporting on OpenAI's data controls.
References
- Anthropic, Updates to Consumer Terms and Privacy Policy, rechecked July 24, 2026.
- Anthropic Privacy Center, How long do you store my data?, rechecked July 24, 2026.
- Google, Gemini Apps Privacy Hub, last updated June 29, 2026 and rechecked July 24, 2026.
- Microsoft Support, Privacy FAQ for Microsoft Copilot, rechecked July 24, 2026.
- OpenAI Help Center, How your data is used to improve model performance, rechecked July 24, 2026.
- OpenAI Help Center, Temporary Chat FAQ, rechecked July 24, 2026.
- Meta Privacy Center, How Meta uses information for generative AI models, rechecked July 24, 2026.
- Meta, Improving recommendations with AI interactions, rechecked July 24, 2026.