Which AI tools train on your chats by default (2026)
- ChatGPT (consumer): trains on chats by default; opt out via Settings → Data Controls.
- Claude (consumer): Anthropic's 2025 policy trains on chats unless you opt out, with multi-year retention if you don't.
- Gemini: conversation data feeds improvement and sampled chats may be human-reviewed; controls live in Activity settings.
- Business/API tiers are generally excluded by default. Local models never see your data leave the machine at all.
The major AI assistants flipped a quiet default over the last two years: your conversations help train the next model unless you say otherwise. The toggles exist. They're just buried, named blandly, and switched on. Here's the state of play in 2026, where each opt-out lives, and the one architecture where the question never comes up.
The 2026 training-default scoreboard
| Tool (consumer tier) | Trains on chats by default? | Where the control hides |
|---|---|---|
| ChatGPT | Yes, unless opted out | Settings → Data Controls → model improvement toggle |
| Claude | Yes since the 2025 policy change, unless opted out | Privacy settings; presented at login as a choice most people click through |
| Gemini | Conversation data used for improvement; sampled human review | Gemini Apps Activity controls |
| Copilot / Cursor and other wrappers | Varies by underlying provider and plan | Each vendor's privacy page, separately |
The pattern: defaults favor collection, controls exist but require you to go looking, and the policies move. Anthropic's 2025 change is the instructive one: Claude spent years marketing a no-training-by-default stance, then reversed it for consumer accounts with an opt-out window. Whatever today's setting says, you're one policy update from a different deal.
Why opting out isn't quite enough
The opt-out governs training. It doesn't stop your conversations from being transmitted, stored, retained under the provider's schedule, or producible under legal process (the 2025 preservation order against OpenAI swept in deleted chats, opt-outs notwithstanding). Opting out is worth doing. It shrinks one use of your data while leaving the architecture (your words on their servers) fully intact.
Business tiers: better deal, same architecture
Enterprise and API agreements typically exclude customer data from training by default, and that's a real improvement. The data still leaves your machine, though, and the stronger promises are contractual rather than structural. Contracts are good. Architectures that make the contract unnecessary are better.
The architecture with no toggle to find
A model running on your own Mac can't train on your chats, because the company never receives them. There is no default to audit, no setting to dig for, no policy update to track. Outlier's models run entirely on-device; your conversations are files on your disk and nothing else. The trade, as always with local: slower than cloud flagships, and the biggest cloud models keep the edge on the hardest problems. What you get is a privacy position that doesn't depend on anyone's settings page, including ours.
Frequently asked questions
How do I stop ChatGPT from training on my data?
Open Settings, go to Data Controls, and turn off the model-improvement option. This stops training use on consumer plans but doesn't stop storage or retention. Enterprise and API usage is excluded from training by default.
Does Claude train on my conversations?
Under Anthropic's 2025 consumer policy, yes by default unless you opt out, with extended retention if you allow it. Commercial and API tiers carry different terms. The policy reversed Anthropic's earlier no-training default, so it's worth rechecking periodically.
Is there an AI where opting out is unnecessary?
Local AI. When the model runs on your own machine, your conversations never reach a provider, so there's nothing to train on and no setting to manage. That's the structural difference between local tools like Outlier and any cloud assistant.
Try Outlier free
Free Nano + Lite — local, private, no account. Pro $20/mo or $149/yr adds everything (all 7 model tiers incl. Plus 397B). Lifetime Pro from $99 (Founding 200, first 200 seats) or $200 (Founders 500). Apple Silicon only.
Download for Mac