one year on
Anthropic updates consumer terms, will train on user chats unless they opt out
Claude Free, Pro, and Max users must decide by October 8 whether to allow their conversations to be used for model training, with a five-year data retention period for those who do not opt out.
Anthropic is overhauling its consumer privacy policy, now requiring Claude Free, Pro, and Max users to actively opt out if they do not want their chats and coding sessions used for model training. The company announced today that starting immediately, new or resumed conversations—including those in Claude Code on those accounts—will be retained for five years and fed into training unless the user changes the setting. Users who opt out keep the prior 30-day deletion policy.
The change, which effectively reverses Anthropic’s earlier practice of not training on consumer data, applies only to consumer tiers. Commercial customers using Claude for Work, Claude for Government, Claude for Education, and API access (including via Amazon Bedrock and Google Cloud Vertex AI) are excluded. Users must make a choice by September 28, 2025, to continue using the service, after an in-app pop-up that has drawn criticism for its design: a large Accept button sits above a much smaller toggle that defaults to allowing training.
The company frames the move as a user-empowerment step, arguing that sharing data helps improve model safety and capabilities. Critics note that the opt-out structure, combined with a quietly extended retention period, creates a high hurdle for privacy-conscious users. The policy shift arrives amid broader industry tension over data practices; OpenAI is currently fighting a court order to retain all ChatGPT conversations indefinitely in a copyright lawsuit. For Anthropic, the bet appears to be that more data outweighs any brand damage from abandoning its earlier privacy stance.
The record
raised concerns that the pop-up design—a prominent Accept button with a small, pre-set-on toggle for training—might lead users to accidentally agree to data sharing.
long warned that the complexity of AI makes meaningful consent nearly unattainable.
One year later — open only if you can handle spoilers
Anthropic pushed the decision deadline for existing users back from September 28 to October 8, 2025, and the policy took effect: consumer chats and coding sessions now default to training use with five-year retention unless the user opts out. The consent pop-up — a bold "Accept" button above a pre-toggled "On" switch — drew dark-patterns criticism from privacy researchers who argued it sat uneasily with the GDPR standard for freely given consent, though no formal regulatory action had followed by mid-2026. In June 2026 an updated privacy policy spelled out a carve-out critics said undercut the promise: chats flagged for safety review can still be used to train Claude even for users who opted out.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy