Anthropic Wants Your Hard Questions on AI. Who Grades the Answers?

The AI lab is asking the public to submit tough questions about AI's impact and promising to show its work. The commitment sounds impressive until you look for the part where anyone outside Anthropic gets to verify the results.

In early July, Anthropic published a post titled "Inviting hard questions", calling on the public to submit their most challenging concerns about artificial intelligence. The company pledged to track and publicly report its progress on the issues raised, covering everything from job displacement to the erosion of creative work to the existential question of whether AI undermines human agency. It's the kind of initiative that looks great in a corporate newsroom. The harder question is what it actually commits Anthropic to doing, and what happens when the answers are uncomfortable.

As we noted in our earlier coverage, a public inbox is not the same thing as independent oversight. Six weeks later, that gap between process transparency and substantive accountability remains the central tension in Anthropic's approach, one sharpened by the competitive context in which the initiative landed.

What the Initiative Actually Promises

Anthropic's "hard questions" initiative has two concrete components. First, an open call: anyone can submit a question about AI's societal impact. The company highlights the spectrum of public sentiment it's trying to capture, from daily AI users who see productivity gains to people worried about job loss, devalued creative work, and diminished human connection. Second, Anthropic says it will "transparently chart our progress toward our public benefit goals," as described in its announcement, essentially promising a public paper trail.

The company frames this within its status as a Public Benefit Corporation, pointing to existing efforts like AI safeguards research, free model access for scientists, and a fellowship program pairing early-career Claude users with nonprofits. These are real programs, and the Public Benefit Corporation structure is meant to signal that Anthropic answers to more than shareholders. But the "hard questions" initiative itself is, structurally, a feedback mechanism. Anthropic decides which questions to prioritize, how to frame its responses, and when to publish them. There is no external review board, no third-party auditor, no binding timeline.

The promise is to "show our work." The missing piece is who checks it.

The Data Anthropic Is Building Around Public Sentiment

To its credit, Anthropic isn't operating blind. The company has been investing in understanding how the public actually feels about AI, and the results are more nuanced than the usual tech-industry assumption that everyone either loves or fears the technology.

In June, Anthropic published results from its first Public Record survey, a nationally representative study of nearly 52,000 Americans conducted in late 2025 via YouGov. The findings were notable for what they didn't show: AI attitudes did not heavily divide Americans along partisan, geographic, or educational lines. Instead, the survey found broad consensus. Americans want AI's benefits but fear disruption, and they want accountability from the companies building it. Where disagreement existed, it was mostly about intensity, not direction.

That's a useful finding, and it's the kind of data that could inform genuinely responsive policy. The Public Record survey joins other Anthropic research efforts, including a global qualitative study of 81,000 Claude users conducted through the company's own interviewing tool, and the Anthropic Economic Index, which draws on anonymized usage data.

The breadth of this research is impressive. But notice the pattern: every data-collection tool here is designed, administered, and interpreted by Anthropic. The company is building a sophisticated picture of public opinion, but it's a picture it controls from end to end.

Process Transparency Is Not Accountability

Anthropic's Transparency Hub offers a broader view of the company's responsible development practices. It covers the Responsible Scaling Policy, first published in September 2023 and updated in February 2026, which establishes capability thresholds that trigger additional safeguards. It also describes the Frontier Compliance Framework, published in December 2025, which lays out how Anthropic evaluates risks across categories including cyber offense, CBRN threats, AI sabotage, and harmful manipulation.

These are substantive documents. The Responsible Scaling Policy's tiered approach, where stronger model capabilities require proportionally stronger protections, is a more structured framework than most competitors have published. And the Transparency Hub itself, organized by key areas of focus and inviting feedback from policymakers and the AI community, represents a level of disclosure that is unusual in the industry.

But process transparency and accountability are different things. Publishing your safety framework is a form of showing your work. Having someone outside your organization verify that you're following it is accountability. Anthropic acknowledges this distinction implicitly by noting that its Responsible Scaling Policy "is designed to evolve as our understanding of AI risks improves." Evolution is good. But who decides when the evolution is sufficient? Right now, the answer is Anthropic.

Our earlier reporting on this initiative raised the same concern: without independent oversight, the "hard questions" program is a speech commitment, not an accountability mechanism. Nothing in the weeks since has changed that structural reality.

The Competitive Backdrop Matters

Timing offers a clue to why Anthropic is making these moves now: the company's original announcement landed in the same week OpenAI was fielding pressure about AI's economic effects. That timing read as deliberate positioning: Anthropic as the company that invites scrutiny rather than waiting to be cornered.

The Ramp Data

The competitive dynamics have only intensified since. Data from corporate expense platform Ramp, reported by TechCrunch on August 20, shows OpenAI gaining ground on Anthropic among U.S. business users:

• Anthropic held roughly 44% market share among Ramp's paying business customers as of July, compared to OpenAI's approximately 40%.
• OpenAI has been growing faster in Q3 to date.
• Ramp's data covers more than 70,000 American businesses, though it skews toward the tech industry and doesn't capture large enterprises using other spend-management tools.

The volatility is the point: businesses are willing to switch providers as new models drop, which should give both companies' investors pause about how sticky enterprise AI revenue really is. In a market where customers flip based on the latest benchmark, trust and brand differentiation become strategic assets. Anthropic's transparency push isn't just altruism. It's also competitive positioning in a market where the product itself is increasingly commoditized.

None of that makes the initiative cynical. Companies can pursue genuine public benefit and competitive advantage simultaneously. But it means the "hard questions" program should be evaluated on its structural merits, not its marketing value.

What Would Real Accountability Look Like?

If Anthropic wanted to move from transparency to accountability, the path is reasonably clear. Independent audits of its safety evaluations. External review of how it selects and responds to public questions. Binding commitments with consequences for non-compliance. A governance structure where at least some oversight sits outside the company's control.

Some of these steps would be unprecedented in the AI industry. That's partly the point. Anthropic has positioned itself as the safety-first lab. The "hard questions" initiative, the Public Record survey, and the Transparency Hub all reinforce that brand. But brand and governance are different categories.

The Public Record survey found that Americans want accountability from AI companies. Anthropic is betting that asking for hard questions is a form of delivering it. Whether that bet pays off depends on what happens next: whether the responses are substantive, whether the company changes course when the feedback is uncomfortable, and whether anyone outside Anthropic's walls gets to verify the answers.

The questions are invited. The accountability is still optional.