Does ChatGPT train on the notes I send it?

Whether OpenAI trains on your content depends entirely on which product you're using and which settings are enabled. Here's the 2026 reality.

ChatGPT free and Plus (consumer)

• Default: yes, your conversations are used for model training.
• To opt out: Settings → Data Controls → "Improve the model for everyone" → off.
• Effect of opt-out: your conversations are still stored for 30 days for abuse monitoring, then deleted. Not used for training.

ChatGPT Team and Enterprise

• Default: no, content is never used for training.
• Logging: conversations are stored for the user's history but excluded from training pipelines.

OpenAI API (the surface most apps use)

• Default since March 2023: API requests are NOT used for training.
• Retention: stored for 30 days for abuse monitoring, then deleted, UNLESS the developer has signed a "zero data retention" (ZDR) agreement with OpenAI, in which case requests are deleted immediately.
• Most apps using the API don't pay for ZDR — Notion, Mem, Reflect, etc. use standard API which has 30-day retention.

What this means for notes apps:

• Notion AI: powered by OpenAI API. Notion's privacy policy states they have a zero-retention agreement, so OpenAI doesn't keep your content. But Notion themselves can still read it because it passes through their servers unencrypted.
• Mem: similar — API-based, opted out of training, but Mem holds the data unencrypted.
• Apple Notes (with Apple Intelligence): Apple's on-device AI does not train on your content. Cloud-routed requests go through Apple's Private Cloud Compute, which Apple says doesn't store content beyond the request. No third-party AI provider involved.
• Némos: uses Apple's on-device Foundation Models exclusively. No content ever leaves your device for AI processing. No OpenAI involvement.

What about non-OpenAI providers?

• Anthropic (Claude): API not used for training by default. Consumer Claude.ai conversations not used for training by default since 2024.
• Google (Gemini): free Gemini consumer plan trains by default; can opt out. Workspace AI doesn't train.
• Microsoft Copilot: doesn't train on enterprise data; consumer terms vary.

The trust gap:

Even when policies say "we don't train," you're trusting:

• The provider's policy is accurate.
• The provider's policy doesn't change without notice.
• The provider's *employees* don't access the data (the technical capability often exists).
• The provider isn't compelled by law to hand it over.

For genuinely sensitive content (medical, legal, financial, journalistic sources), the only safe path in 2026 is on-device AI — Apple Foundation Models, local LLMs via Ollama or LM Studio, or apps explicitly designed around on-device processing.

Quick reference card:

If privacy is a deciding factor, the on-device options are the only ones that don't require trusting a third party.