Does Claude train on your data?
Claude.ai trains on your conversations and other inputs by default on Free and Pro, with an opt-out in account settings. Safety-flagged content is used for training even after you opt out.
A self-serve opt-out exists
Graded on the plan most people are on. The other plans are in the matrix below.
"We may use your Inputs and Outputs to train and improve Anthropic AI models, unless you opt out through your account settings. Even if you opt-out, we will use Inputs and Outputs for model improvement when: (i) your conversations are flagged for safety review"
- 14 Sept 2026
- 5 Oct 2026
- Free & Pro (Claude.ai consumer)
- global
- Fetched from the company
- 102CA4EA…F922
How we read it: The quote runs past the opt-out on purpose. It previously stopped at "...through your account settings", which left the headline claim — that safety-flagged content is trained on even after you opt out — resting on nothing a reader could check, though it is stated in the same sentence of the same capture. Capped at 360 characters, so one exception is not shown: the policy also carves out material you explicitly report. The consumer terms' global text has listed "and to develop other products and services" among the permitted purposes since our 1 Aug 2026 capture; the UK/Ireland variant served at the same URL omits it, which was our 24 Aug "change". The privacy policy has neither phrase.
opt out, opt-out, account settings, object, objection, withdraw consent, by default, default
What the document does offer
- the training control, in the same sentence as the training use: "unless you opt out through your account settings" — the location is named, the setting is not
- two uses that survive the opt-out, stated in the same paragraph: conversations "flagged for safety review", and materials "you've explicitly reported ... for example via our feedback mechanisms"
- "Objection: the right to object to processing of your personal data, including profiling conducted on grounds of public or legitimate interest" — gated by "In places where such a right applies", and subject to Anthropic demonstrating "compelling legitimate grounds ... which override your interests"
- the marketing opt-out, the targeted-advertising opt-out and the Cookie Policy controls — none reaches the training use
- "Withdrawal of consent" applies "where Anthropic's processing of your personal data is based on consent"; the training sentence does not state consent as its basis
- "default" appears zero times in the capture, as do "turned on", "turned off" and "enabled"
Why none of it applies: The control is real, free and in account settings, and the policy states it in the same sentence as the training use rather than leaving a reader to find it. Three things it does not do. It does not name the setting, so this is not in the "switch is named" pattern — it gives a location, as Airbnb does, and that group is limited to documents naming an actual switch. It states no default: "default" appears zero times. And it does not say whether opting out reaches Inputs and Outputs already used — the question Perplexity answered plainly and OpenAI answered for new conversations only.
"Anthropic may not train models on Customer Content from Services."
- 24 Aug 2026
- 5 Oct 2026
- Team & Enterprise (business offerings)
- global
- Fetched from the company
- BDF24834…182B
How we read it: Resolved from UNCLEAR on 24 Aug 2026. The consumer privacy policy sends business offerings to "customer agreements" without saying what they contain, which is why this tier sat unclear; the commercial terms say it in one sentence, and in the opposite direction from the consumer tier. B rather than A because it is an undertaking in terms Anthropic can revise, and the same grade OpenAI's business tier carries on an equivalent statement.
What each plan can refuse
The same questions, asked of every plan we could identify. Every cell comes from a quoted clause or a documented search - see the evidence above.
- Self-serve toggle
- Not stated
- 50 / 100
- Qualified
- Unknown
- Nothing to switch off
- Not stated
- 19 / 100
- Not stated
- Unknown
How to opt out of Claude AI training
Plan by plan, from what Claude's own policy says about refusing. Where it publishes the steps, they're listed here with the limits they carry. Where there's nothing to switch off, or nothing graded, this says so instead of inventing a menu path.
Free & Pro (Claude.ai consumer)
Yes. Claude has a self-serve setting.
- Click your name in the settings menu, then Settings.
- Open Privacy.
- Under "Help Improve our AI models", switch the toggle off.
Applies to new chats from the moment you switch it off. Anthropic says chats already in a training run stay in it, and chats its safety classifiers flag can still be used.
- Tap your name in the settings menu, then Settings.
- Tap Privacy.
- Under "Help Improve our AI models", switch the toggle off.
Team & Enterprise (business offerings)
Nothing to switch off: Claude's policy says training is off here.
What this policy has done since we started watching
17 changes detected since 25 Feb 2024. Each one is a diff between two archived copies, and each links to the passage that moved.
- Touched retention and sub-processors.
- Touched retention.
- Touched human review, opt-outs, retention, what's collected, sub-processors, plan tiers and training.
- MajorTouched jurisdiction, opt-outs, retention, what's collected, sub-processors, plan tiers and training.
- Touched jurisdiction, opt-outs, retention, what's collected, sub-processors, plan tiers and training.
- Touched jurisdiction, opt-outs, retention, what's collected, sub-processors, plan tiers and training.
- Touched jurisdiction, sub-processors and plan tiers.
- Touched jurisdiction and retention. Not yet reviewed by a human.
Everything we read
Vendor response
We wrote to Anthropic on 20 Aug 2026, quoting the clauses above and inviting a correction. Nothing has come back. That's published because a company declining to comment on its own policy is something a reader should know, and because the invitation should be on the record whether or not it's taken up. Any reply goes here, word for word.
Report a wrong clause