Anthropic invites the public to ask harder questions about AI
Anthropic invited the public’s hardest AI questions and pledged to show its work as it answers — a rare open invitation from a frontier lab in mid-2026.
On July 9, 2026, Anthropic published “Inviting hard questions” and opened a public portal at claude.com/hard-questions. The pitch is unusual for a frontier lab: send the uncomfortable questions about jobs, safety, kids, science, and power — and the company will publicly track the actions it takes, including where it falls short.
That is a legitimacy campaign, not a changelog. Anthropic is trying to own the agenda for skepticism instead of leaving it entirely to regulators and unions.
What Anthropic asked for
The newsroom post asks for hard questions on AI’s effects on jobs, society, and families; on scientific upside; on danger; on who decides the rules. Anthropic says it will “publicly track and report the specific actions” it takes in response and be clear about shortfalls. The portal lets people browse others’ questions and add their own.
This sits on work Anthropic already claims: the Anthropic Public Record survey of about 52,000 Americans; Anthropic Interviewer responses from about 81,000 Claude users across 159 countries and 70 languages; focus groups; and anonymized usage study. Those figures are Anthropic’s. Treat them as company research until independent replication exists.
Why public Q&A is a culture move
Labs usually publish answers they chose. Hard Questions hands the microphone to strangers and promises a paper trail. Secondary coverage noted trust numbers from Anthropic’s own polling — AI companies ranking poorly versus governments — which makes the campaign look like damage control with a research budget. It can also be sincere. Both readings can be true: Mother London reportedly worked the campaign creative, per Ad Age notes in secondary writeups, which is not how you run a quiet safety paper.
The test is not the portal launch. It is whether a submitted question ever changes a ship decision, a policy stance, or a product default in public view.
How this connects to later safety posts
Read that later safety writing as the scoreboard this inbox is supposed to feed. A public-benefit lab that opens a question portal is asking to be judged on whether subsequent incident reports, classifier changes, and policy briefs answer named public fears — jobs, dual-use biology, eval-time deception — or just restock the newsroom.
The campaign structure is inbox plus ledger. The inbox is live at claude.com/hard-questions. The ledger is the promised public tracking of actions and shortfalls. Without dated entries on that ledger, later safety posts are still the lab talking to itself.
What good hard questions look like
Good questions are falsifiable. One that would count as answered: publish a dated incident report when an evaluation agent takes unsanctioned live-internet action, including what the model did, what the harness allowed, and what product default changed afterward. Vague “will AI ruin us” prompts let a lab answer with vibes.
ByteForward will treat Hard Questions as a primary source when Anthropic publishes tracked responses with dates attached — and as marketing until then. The portal is live. The accountability clock starts when the first public scoreboard of actions appears.
[…] The study follows earlier public outreach covered in our report on Anthropic’s invitation for difficult AI questions. […]