Claude Science anthropic.com

Anthropic launched a formal public submission portal on Wednesday inviting anyone to submit the hardest questions they have about AI — and committing to publish not just its answers but the reasoning that produces them, including cases where the company acknowledges it is wrong or uncertain. The initiative, called "Hard Questions" and accessible at Anthropic's Hard Questions portal, is one of the more structurally unusual transparency moves by a frontier AI lab: it hands agenda-setting to the public rather than to the company's communications team.

The timing is not incidental. Anthropic's own nationally representative survey of nearly 52,000 Americans, published last month, found that only 15% of respondents said they trust AI companies to make decisions about how the technology is developed and used. That figure was the lowest of any institution the survey tested — below the federal government, state and local governments, international bodies, and far below independent experts at 43%.

AI Company Trust Ranks Last Among Every Surveyed Institution

The Anthropic Public Record survey, conducted by YouGov between November and December 2025 with a national margin of error of ±0.6 percentage points, documented a trust landscape that the company itself describes as the backdrop for the Hard Questions initiative. Job loss was the most common fear Americans cited, held by 64% of respondents and ranking as the top concern in every state — among Democrats at 67%, Republicans at 62%, and across every household type. Cognitive dependency, the worry that AI integration leaves people unable to think for themselves, ranked second at 56%. Misinformation ranked third at 52%.

On the hopeful side, nearly half of Americans — 48% — named curing diseases like cancer or Alzheimer's as one of their top three hopes for AI, placing it 12 percentage points above the next option. What the survey found missing was trust in the companies building the technology to navigate this terrain responsibly.

What Anthropic Is Actually Committing To: Mechanism, Limits, and What's Missing

The Hard Questions mechanism is structurally different from the transparency tools most AI labs already use — safety papers, whitepapers, voluntary commitments, and blog posts. Those are curated output: the company decides what it is ready to say, frames the questions it chooses to address, and controls every published finding. The Hard Questions portal inverts that: the public submits the agenda, and Anthropic has committed to publish not just positions but the reasoning that leads to them, including the company's acknowledgment of its own uncertainty and shortfalls.

That structural difference matters. Most of what passes for AI transparency is what researchers have called "self-referential opacity" — labs define what safety means, evaluate whether they meet their own definitions, and publish the results. An arXiv paper published in 2025 rated voluntary frontier AI commitments as failing on both democratic legitimacy (no external authorization of the standards) and accountability (self-reporting with no external verification). Brandie Nonnecke, director of the CITRIS Policy Lab at UC Berkeley, put it plainly in MIT Technology Review's 2024 assessment of voluntary commitments: "These are still companies that are essentially writing the exam by which they are evaluated."

The Hard Questions initiative does not fully resolve this. Anthropic still decides what constitutes an adequate answer to a submitted question, controls the publication timeline, and bears no legal consequence if its public reasoning trail diverges from its internal deliberations. There is no independent body with authority to audit the responses, no mechanism for the public to reject an answer as insufficient, and no enforcement architecture if the company falls short of its stated commitments. What exists is a reputational stake — stronger than a safety white paper because it invites external challenge, but weaker than any legally binding accountability mechanism.

Dario Amodei, Anthropic's CEO, has drawn that line himself. In a policy essay published June 10, Amodei wrote directly that "the rapid pace of acceleration means that transparency alone is no longer sufficient" and called on governments to take on binding authority over frontier AI deployments, including mandatory third-party safety testing and civil penalties tied to global revenue for companies whose models fail safety thresholds. The Hard Questions initiative is positioned as the voluntary public engagement layer of that larger picture — not the whole accountability structure.

What Americans Want That a Submission Portal Cannot Deliver

The Anthropic Public Record survey found that 71% of Americans — a bipartisan supermajority spanning 68% of Republicans and 79% of Democrats — believe the government should play a role in regulating AI. When asked what single action would best ensure AI is developed in humanity's interest, 47% of respondents named holding AI companies legally liable for harm. Another 44% named prioritizing safety over growth.

Neither of those outcomes is deliverable through a public submission process, however rigorously run. Legal liability requires legislation. Government oversight requires regulation. The survey's most politically unified finding — that the public wants something harder than voluntary transparency — is precisely the gap the Hard Questions initiative leaves open.

That gap is the honest context for the initiative: it is a genuine step, not a substitute. Anthropic has been building toward it for over a year, conducting in-depth interviews with 81,000 Claude users across 159 countries and 70 languages through a purpose-built tool called Anthropic Interviewer, running in-person focus groups, and publishing ongoing data from the Anthropic Economic Index. The Long-Term Benefit Trust, established early in the company's history, provides some independent oversight of how effectively Anthropic advances its public benefit mission — though it lacks the authority of an independent regulator. The Anthropic Institute, a research arm focused on AI's societal challenges, provides institutional grounding.

Anthropic is also structured as a Public Benefit Corporation, a legal form available in Delaware and approximately 40 other states that requires directors to weigh public benefit alongside profit and insulates them from shareholder suits for doing so. That structure gives the initiative legal cover — Anthropic can justify commercial costs to its mission without facing liability to investors. What it does not create is an enforcement mechanism for anyone outside the company.

How Does Anthropic's Hard Questions Initiative Actually Work?

The initiative operates through a dedicated portal at claude.com/hard-questions. Visitors can submit questions on topics including AI's effects on jobs and society, AI's potential in science and medicine, safety and governance, and the broader trajectory of AI development. Anthropic has committed to publicly tracking the specific actions it takes in response to submitted questions — and, notably, to disclosing where it falls short of its stated goals.

The company produced a two-minute film, directed by Myles McAuliffe and created by Mother agency as part of Anthropic's "Keep Thinking" brand platform, drawing on conversations with more than 12,000 people about their hopes and fears around AI. The film surfaced questions including "Who decides the rules for AI?", "Does AI make the world more dangerous?", and "Could AI help people stop feeling misunderstood?" — the range intentionally spanning existential and everyday.

What Anthropic has not specified is a timeline for publishing its first substantive responses, a format for what those responses will look like, or a mechanism for users who disagree with an answer to escalate their challenge. Those details will determine whether the initiative functions as genuine public accountability or as a more sophisticated version of the curated FAQ.

Why This Moment Is Harder Than It Looks for Anthropic

The Hard Questions launch arrives against a backdrop that makes the credibility test unusually concrete. Anthropic is currently in active federal litigation over a Trump administration directive issued in February 2026 that designated the company a "supply chain risk" and directed federal agencies to cease using its technology — the first time that designation, historically reserved for foreign adversaries like Huawei and ZTE, had been applied to an American company. A subsequent Commerce Department export control directive in June forced Anthropic to globally disable Fable 5 and Mythos 5 for all users, because the company's access architecture had no mechanism to selectively filter users by nationality at commercial scale.

The company disputed both actions and has filed lawsuits challenging them. But the episode demonstrated that Anthropic's most powerful deployed models operate in a legal environment that is neither stable nor transparent to users — a fact that sits in some tension with an initiative premised on accountability and showing its work.

Separately, the June 2026 launch of Claude Fable 5 — Anthropic's most capable public model — included safety classifiers that silently downgraded legitimate security and chemistry researchers' responses without notifying them. Anthropic apologized within days and made the downgrade visible, but critics noted the fix added transparency to the downgrade without removing it. A prominent AI safety researcher who departed Anthropic in February 2026, Mrinank Sharma, had spent his time there on exactly these problems: AI sycophancy and bioweapons defenses. His public departure letter described a world "in peril."

These are not disqualifying facts for the Hard Questions initiative. They are the honest context for what the initiative needs to demonstrate over time: that its public reasoning trail constrains actual company decisions, not just the ones Anthropic was already comfortable making.

What the Bipartisan Case for AI Accountability Actually Looks Like

The Anthropic Public Record survey found something unusual in an era of deep partisan division: AI governance is not a partisan issue. Job loss fears differed by just five percentage points between Democrats and Republicans. Support for government involvement in AI reached 79% among Democrats and 68% among Republicans. Integrated AI users — roughly 6% of Americans who use AI daily for both work and personal purposes — supported government involvement at essentially the same rate as the general public (74% vs. 71%). Even the Americans most embedded in AI could not be described as opposed to oversight.

That consensus has not translated into federal legislation, which remains stalled. In its absence, state-level transparency laws — including New York's RAISE Act, which Anthropic publicly supported and which takes effect January 1, 2027, and California's SB 53 — have created the first mandatory disclosure floor for frontier AI developers above $500 million in annual revenue. Those laws require annual risk frameworks and 72-hour incident reporting, with fines of up to $3 million for repeat violations. They are the nearest thing to enforceable accountability currently on the books.

The Hard Questions initiative sits in the space between those mandatory disclosure floors and the broader public accountability mechanism the survey data suggests the public actually wants. It is worth watching. Whether the public reasoning trail it promises turns out to be binding in practice — whether it constrains decisions the company would otherwise make differently — is the question that will determine its significance.

Submissions are open at Anthropic's Hard Questions portal.

Frequently Asked Questions

How does the Anthropic Hard Questions initiative actually work?

Anyone can submit a question at Anthropic's Hard Questions portal. Anthropic has committed to publicly tracking the specific actions it takes in response and to disclosing where it falls short of its stated goals. The structural distinction from prior AI lab transparency efforts is that the public sets the agenda — the company does not choose which questions to acknowledge. What Anthropic has not specified is a timeline for publishing responses, a format for those responses, or a process for users who find an answer inadequate.

Is Anthropic's transparency commitment legally binding?

No. The Hard Questions initiative is a voluntary commitment backed by reputational risk rather than legal consequence. Anthropic defines what constitutes an adequate answer, controls the publication process, and bears no legal penalty if its public reasoning trail diverges from its actual decisions. Dario Amodei, Anthropic's CEO, has publicly stated that "transparency alone is no longer sufficient" and called for binding government regulation — mandatory third-party safety testing, government authority to block deployments, and civil penalties. That regulatory framework, if enacted, would create enforceable accountability. The Hard Questions portal does not.

Why do Americans distrust AI companies more than their own government?

Anthropic's own Anthropic Public Record survey — 51,993 Americans, conducted by YouGov in November and December 2025, with a national margin of error of ±0.6 percentage points — found that only 15% of respondents trust AI companies to make decisions about how the technology is developed and used. That was lower than trust in the federal government (20%), state and local governments (19%), and international bodies (20%), and far below independent experts (43%). The survey found no significant partisan divide on the trust question. Researchers who study voluntary industry self-governance have documented a structural reason for the gap: AI companies currently define the standards they evaluate themselves against, report their own compliance, and face no external verification.

How does the Hard Questions initiative differ from other AI labs' transparency efforts?

Most major AI labs publish safety papers, whitepapers, and voluntary policy commitments — all forms of curated output where the company decides what it is ready to say. The Hard Questions initiative is different in that it hands agenda-setting to the public: submitted questions, not company-selected ones, drive what Anthropic commits to address publicly. The additional commitment to publish reasoning rather than just conclusions — including cases where Anthropic acknowledges uncertainty or shortfalls — is also structurally unusual. The test of whether this difference is meaningful is whether the public reasoning trail constrains decisions the company would otherwise make differently, which will only be visible over time.