Anthropic asked 52,000 Americans about AI. Only 15% trust the companies building it.

Anthropic has published results from the first wave of what it's calling the Anthropic Public Record, a nationally representative survey of 51,993 Americans fielded with YouGov in November and December of 2025. It's a notable document, and not only because a frontier lab rarely commissions polling at this scale on its own industry. The findings are awkward for the people who paid for them.

Start with the trust number. Only 15% of Americans say they trust AI companies to make decisions about how AI is developed and used. That's the lowest figure for any institution Anthropic tested — below the federal government (20%), state and local government (19%), and international bodies (20%). Independent experts sit at 43%.

A company publishing a survey that ranks its own industry dead last on trust is, at minimum, an interesting choice.

What people want

The top hope for AI is curing disease (48% put it in their top three), followed by helping people with disabilities (36%). Hopes that AI might substitute for human contact — therapy, reducing loneliness — ranked lowest. This is roughly the inverse of how a lot of consumer AI is currently being marketed.

The top fear is job loss, held by 64% of Americans, and Anthropic reports it is the number-one fear in every state and across party lines (67% of Democrats, 62% of Republicans). Cognitive dependency comes second at 56%, misinformation third at 52%. The harms people worry about most are near-term and concrete; concerns about AI "going rogue" rank lower than criminal misuse, surveillance, and terrorism.

Two patterns inside the job-loss data are worth flagging. First, worry rises with education — postgraduates are nearly 10 points more worried than people with a high school education or less, which Anthropic attributes to their work overlapping more with what current models do. Second, worry falls sharply with hands-on use: 54% among daily AI users at work versus 70% among non-users.

Anthropic frames the second finding charitably (experience builds fluency, reveals limits). The less flattering reading is also available: people closest to the deployment decisions are the ones doing the deploying.

What people want done about it

71% of Americans say government should be involved in AI regulation — 79% of Democrats, 68% of Republicans, 69% of Independents. Anthropic calls this a bipartisan supermajority, and the state-by-state floor is 63% (Hawaii).

Asked what would best ensure AI benefits humanity, respondents converged on two answers: hold AI companies legally liable for harm (47%) and prioritize safety over growth (44%). Independent watchdogs with real power came in at 29%, slowing development at 27%.

The priority domains for government action are privacy (56%), child safety (52%), and liability for harm (49%).

The cognitive dependency wrinkle

One of the more honest findings: cognitive dependency, the second-biggest fear, is mostly anticipatory. Among the 56% worried about it, only about a fifth say they'd feel significant disruption if AI disappeared tomorrow. Among the 44% who aren't worried, about a third would. People are afraid of becoming dependent; the people who already are mostly aren't afraid.

Educators were 2.5 to 3 times more likely than average to report witnessing cognitive atrophy firsthand — presumably in students — which tracks with the qualitative work Anthropic has run through its Interviewer tool with 81,000 Claude users.

What we'd watch

Anthropic says it will repeat the survey and eventually expand outside the US. The interesting tension is structural: the company is producing high-quality public evidence that the policy interventions Americans most want — legal liability, mandatory safety prioritization, independent watchdogs — would constrain Anthropic and its competitors more than current voluntary frameworks do. The post ends by gesturing at Anthropic's Advanced AI Framework and mandatory independent safety testing for frontier models, which is the company's preferred frame for the liability conversation.

Whether regulators read the survey the same way the publisher does is the part worth watching.