Inputs and Outputs train models unless you opt out, with two exceptions after opting out
Anthropic may use Inputs and Outputs to train and improve Anthropic AI models unless you opt out through account settings. Even after opting out, it will use Inputs and Outputs for model improvement when conversations are flagged for safety review to improve detection of harmful content, enforce policies or advance AI safety research, or when you have explicitly reported the materials, for example through feedback mechanisms.
Supporting quotations
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 to improve our ability to detect harmful content, enforce our policies, or advance AI safety research, or (ii) you've explicitly reported the materials to us (for example via our feedback mechanisms).
Qualifications & exceptions
Training use of Inputs and Outputs is the default unless you opt out through account settings.
The opt-out is not absolute: safety-flagged conversations and materials you explicitly report remain available for model improvement.
Not stated here
The policy does not say whether opting out affects models already trained, how opt-out status is verified, or which account settings control it.
Inference & Profiling · Company commitment
No decision-making based solely on automated processing with legal or similarly significant effects
Anthropic states that it does not engage in decision-making based solely on automated processing or profiling that produces a legal effect or significantly affects the person in a similar way, such as significantly affecting financial circumstances or access to essential goods or services.
Supporting quotations
Automated decision-making: Anthropic does not engage in decision making based solely on automated processing or profiling in a manner which produces a legal effect (i.e., impacts your legal rights) or significantly affects you in a similar way (e.g., significantly affects your financial circumstances or ability to access essential goods or services).
Qualifications & exceptions
The commitment is framed by the legal standard: solely automated decisions with legal or similarly significant effects.
Not stated here
The policy does not describe what automated review exists that falls short of that standard, or any human review process.
Inference & Profiling · Company practice
Feedback used for training has Inputs and Outputs disassociated from the user ID
When a user submits Feedback, Anthropic disassociates the Inputs and Outputs from the user's ID before using them to train and improve its models.
Supporting quotations
When you submit Feedback, we disassociate Inputs and Outputs from your user ID to use them for training and improving our models.
Qualifications & exceptions
Disassociation applies to Feedback-triggered training use.
Not stated here
The policy does not describe how disassociation is implemented or whether it is reversible.
Inference & Profiling · Company practice
Safety-flagged content is disassociated for classifier training but may be re-identified to enforce policies
Where systems flag Inputs or Outputs for harmful content or for potentially violating Anthropic's policies, the content is disassociated from the user ID and used to train trust and safety internal classification and generative models. Anthropic says it may re-identify that content to enforce the Terms of Service or Usage Policy with the responsible user where necessary.
Supporting quotations
If our systems flag Inputs or Outputs for harmful content or for potentially violating our policies, we disassociate the content from your user ID to train our trust and safety internal classification and generative models. However, we may re-identify the Inputs or Outputs to enforce our Terms of Service or Usage Policy with the responsible user if necessary.
Qualifications & exceptions
De-identification for safety training is qualified by a re-identification route for enforcement.
Re-identification is described as possible where necessary, not routine.
Not stated here
The Terms of Service and Usage Policy are separate documents and their text is not supplied here, so the enforcement grounds are not described.
The policy does not say who authorizes re-identification or how it is logged.