Anthropic asks Claude users to share data for training

Anthropic has always pitched itself as the safety-first, privacy-minded alternative in the frontier AI race. But good training data is running dry, and now the company wants a slice of your daily chats.

Anthropic updated its Consumer Terms and Privacy Policy to give personal users a choice: allow your chats and coding sessions to train future Claude models, or keep them private.

If you say yes, Anthropic will extend its data retention window on your conversations from 30 days to five full years.

The fine print on your chats

The policy change hits users on Claude Free, Pro, and Max plans. That also covers anyone using the terminal tool Claude Code under those personal accounts.

Commercial and enterprise tiers are totally exempt. If you are on Claude for Work, Team, Enterprise, Gov, Education, or using Claude via the API, AWS Bedrock, or Google Cloud Vertex AI, your data will not be touched.

Existing consumer users have until October 8, 2025 to make a decision through an in-app prompt. After that deadline, you must pick an option to keep using Claude. New sign-ups will choose during onboarding.

If you opt in, Anthropic keeps your new and resumed chats, coding sessions, and feedback for up to five years. If you opt out, your conversations stick to the standard 30-day retention clock and never enter the training pipeline.

You can change your mind anytime in Privacy Settings. Turning it off stops Anthropic from pulling future sessions into training runs, though any data already baked into trained models cannot be deleted. Deleting an individual chat also pulls it from future training runs.

Anthropic claims AI development cycles take 18 to 24 months, so keeping data for five years ensures consistency across model generations and helps train abuse filters. The company also noted it uses automated tools to scrub sensitive data and never sells personal info to third parties.

Why it matters

We think this move shows just how desperate frontier AI labs are for high-grade human interaction data. Web scraping has hit a wall, and debugging real code alongside Claude Code is arguably the most valuable training material on the planet right now.

For everyday users and solo developers, the decision comes down to risk tolerance. If you feed proprietary client code, financial data, or sensitive documents into Claude, staying opted out is the obvious move. Giving any AI lab a five-year data leash is simply too long for sensitive work.

Competitively, Anthropic handled the rollout relatively well by making the opt-in explicit rather than burying a forced agreement in a quiet terms update. Keeping the enterprise side strictly partitioned also protects their high-paying corporate pipeline.

Still, it feels like a subtle shift in posture. When the company founded on safety and restraint starts asking paying Pro subscribers to surrender five years of interaction history, it is clear that model performance pressures are starting to outweigh brand purism.