You can stop most personal AI agents from training on your data, but the setting is rarely on the agent's own screen. Meta Muse trains by default and offers an opt-out. OpenAI dots follow your ChatGPT "Improve the model for everyone" setting. Claude Duet falls under Anthropic's Model Improvement setting, and Grok Bot follows your Cursor account settings. Here is where each control sits as of November 2026, based on the official docs and launch coverage.
Training vs retention: two different questions
Training means your conversations or connector data help improve the vendor's future models. Retention means the vendor stores your data. Opting out of training does not delete anything, and deleting a chat does not pull it out of a model that was already trained on it. For storage and deletion timelines, see how long AI agents keep your data.
- Your agent activity
- Vendor storage
- Training dataset (if allowed)
- Future model
The opt-out switch sits on the third arrow. It only works going forward.
Opt-out settings by agent
| Agent | Default for personal users | Where to change it | Notes |
|---|---|---|---|
| Meta Muse | Training on | Muse privacy settings (opt-out) | Meta says it strips "critical personally identifying information" first |
| OpenAI dots | Follows ChatGPT setting | ChatGPT Settings, Data controls, "Improve the model for everyone" | No training on proactive research or private notes; Business, Enterprise, Edu excluded by default |
| Claude Duet | Not stated for Duet | Claude Settings, Privacy, Model Improvement | Consumer policy applied by inference; Incognito chats never used |
| Grok Bot | Follows Cursor settings | Cursor account data settings | Legacy Privacy Mode not supported |
| OpenClaw | No vendor training | Your model provider's API terms | Depends on the model you plug in |
As of November 2026. Menu names can change; check the vendor's help pages linked below.
Meta Muse: on by default
CNBC reported that Muse uses your data for training by default and offers an opt-out. Meta says it removes "critical personally identifying information" before training. That is narrower than removing all personal information, so details such as preferences, habits and the content of your requests may still be used.
Meta also says Muse conversations and VM data are not shared with its ad systems. However, what Muse browses on your behalf can influence your ads indirectly. If that bothers you, the training opt-out does not address it; limiting what you ask Muse to browse does.
OpenAI dots: your ChatGPT setting decides
According to OpenAI's dots privacy FAQ, dots on personal plans follow the ChatGPT setting "Improve the model for everyone". Turn it off in ChatGPT's data controls and the same choice applies to your dots. Business, Enterprise and Edu workspaces are not used for training by default. VentureBeat reported that OpenAI does not train on proactive research or private notes regardless of the setting.
OpenAI notes that conversations can still be reviewed by humans for safety purposes. The training toggle does not switch that off.
Claude Duet: Model Improvement, by inference
Anthropic's Duet safety article does not include a Duet-specific training statement. Anthropic's general consumer policy says it trains on your chats only if Model Improvement is on, if a conversation is flagged for safety review, or if you submit feedback. Incognito chats are never used, and feedback is kept for up to five years.
Applying that policy to Duet is our inference, not an Anthropic statement. The safest course is to turn Model Improvement off and avoid sending feedback on conversations that contain sensitive data.
Grok Bot: Cursor settings apply
Cursor's documentation says Grok Bot uses your Cursor account data settings, so the training opt-out is the Cursor setting. The older Cursor Privacy Mode is not supported for Grok Bot, so do not assume that switching on Privacy Mode in an editor covers your Bots.
How to opt out, step by step
- Find the parent accountMuse uses your Meta account, dots your ChatGPT account, Duet your Claude account, Grok Bot your Cursor account.
- Open data or privacy settingsLook for model improvement, training or data controls.
- Turn the training toggle offConfirm it saved, and check again after app updates.
- Skip feedback on sensitive chatsThumbs-up or thumbs-down can attach the conversation for review.
- Consider a business plandots Business, Enterprise and Edu are excluded from training by default.
- Recheck each quarterVendors change defaults, especially when they add new features.
Pros
- Stops your future agent activity entering training datasets
- Takes a minute and usually applies across the vendor's products
- Business plans give stronger defaults
Cons
- Does not delete stored data or memory
- Does not stop safety-flagged human review
- Does not cover indirect effects such as Muse browsing shaping ads
- Data already used in training is not removed
What opting out doesn't fix
Training is one leak path. Your agent still reads every email and message it is connected to, including messages from people who never agreed to share anything with an AI, as our guide to bystander privacy explains. Sensitive categories such as medical details need extra care regardless of training settings; see AI agents and health data. If you use an agent for work, your employer's contract may override personal settings, so read personal AI agents at work.
UK users have a legal right to object to some processing; see UK data rights with AI agents. For a fuller comparison of memory controls, see AI agent memory privacy compared.
Bottom line
Turn off training on the parent account, not just inside the agent: Meta for Muse, ChatGPT for dots, Claude for Duet and Cursor for Grok Bot. Meta Muse is the one that trains by default. Then treat the opt-out as one layer: also limit connectors, reset memory and keep the most sensitive information out of the agent altogether, as covered in what not to share with an AI agent.



