Anthropic Is Building a Public Ledger for Its Own AI Promises

Anthropic launched “Inviting hard questions” on July 9 — a standing site at claude.com/hard-questions and anthropic.com/hard-questions where anyone can submit their toughest question about AI’s effect on jobs, families, science, or where increasingly powerful systems are headed, and browse what others have already asked. The commitment that makes it more than a suggestion box: Anthropic says it will publicly track the concrete actions it takes in response and be transparent about where it falls short of its own stated goals — a self-imposed scorecard, published on a timeline the company doesn’t control alone.

The initiative isn’t a standalone gesture. It follows the company surveying roughly 52,000 Americans through the Anthropic Public Record and interviewing about 81,000 Claude users across 159 countries and 70 languages via the Anthropic Interviewer, and it leans on two structures built earlier this year: the newly formed Anthropic Institute, tasked with researching AI’s biggest societal risks, and the Long-Term Benefit Trust, which just added former Federal Reserve Chair Ben Bernanke as a trustee with no equity and no profit share — compensated only for his time. The same week, Anthropic also published GRAM research, a technique for surgically deleting dangerous knowledge from a trained model rather than just filtering it out during training.

Three moves in nine days — a public accountability tracker, an outside-the-industry trustee, and a technical off-switch for dual-use capability — read less like separate announcements than a single bet: that the AI lab willing to publish its own report card, not just its safety commitments, is the one enterprise buyers and regulators will trust first when the next capability jump lands. Whether “hard questions” produces real answers or just a well-documented list of unresolved ones is the thing worth watching over the next two quarters.