Anthropic to train models on chats by default, unless you opt out

Anthropic has long positioned itself as the high-road AI lab, winning over privacy-conscious coders and researchers who balked at OpenAI’s data hunger. That halo just slipped a bit.

On Thursday, Anthropic announced an update to its consumer privacy terms: unless you explicitly opt out, the company will begin using your chat logs and coding sessions to train future AI models. It is also extending its data retention window from 30 days to five years for anyone who agrees.

The default-on switch and five-year retention

Previously, Anthropic automatically purged consumer prompts and outputs from its backend after 30 days, only keeping flagged conversations for up to two years. More importantly, it did not touch consumer chats to train its foundation models.

Under the new rules, all consumer tiers—Claude Free, Pro, and Max—are fair game. That includes developer sessions logged through the newly released Claude Code tool when tied to personal plans. Commercial customers using Claude Gov, Claude for Work, Claude for Education, or developers hitting Claude via APIs, Amazon Bedrock, and Google Cloud’s Vertex AI remain exempt.

Existing users have until September 28 to make a call, but the user experience is already raising eyebrows. When prompted, Claude presents a large "Accept" button. Tucked underneath in smaller print is the toggle to allow model training, and it is switched on by default. Skip reading the fine print, and your sessions are opted in instantly.

You can turn it off later in Settings under the Privacy tab, but Anthropic notes the change will only apply moving forward. Any transcripts already pulled into the training pipeline cannot be retracted.

Why it matters

For everyday users and developers, this changes how freely you should dump proprietary code or sensitive workflows into Claude. If you rely on Claude Pro or Claude Code for private projects, go into your settings right now and toggle it off. While Anthropic says it runs automated filters to scrub sensitive personal data, no redaction pipeline is foolproof against code or business logic leaks.

For the AI landscape, it shows the data bottleneck sparing no one. Synthesized data only goes so far; human feedback on multi-step reasoning, nuances, and edge-case bug fixes is the gold standard of frontier training. Anthropic is matching OpenAI’s and Google’s data-gathering playbook because rivaling their pace requires real-world prompts.

We think the move is a miscalculated blow to Anthropic's core branding. Relying on default-on dark patterns while framing it as a way to "improve model safety" feels distinctly un-Anthropic. It will harvest data from casual users who smash "Accept," but it risks pushing privacy-first developers toward self-hosted models or pure API endpoints where zero-retention rules still hold.