Anthropic has just published a nationally representative survey of nearly 52,000 Americans, and its most quotable finding is a verdict on Anthropic. Only 15% of respondents said they trust AI companies to make decisions about how the technology is developed and used, the lowest score for any institution the survey tested, below federal government, below state and local government, and less than half the level of trust placed in independent experts. The company reporting that number is one of the companies it indicts. That is not a contradiction the survey hides; it is the reason the survey exists.
What the Public Record actually is
Strip away the branding and the Anthropic Public Record is a recurring public-attitudes survey, fielded by YouGov across November and December 2025, weighted to US Census benchmarks, with a national margin of error of roughly plus or minus 0.6 percentage points. As market research it is serious and well-constructed. As a strategic act it is something more interesting: a frontier AI lab appointing itself the keeper of a document called “the public record” at exactly the moment the public says it does not trust that lab’s judgement.
The timing is not incidental. Anthropic released the survey alongside references to two of its own policy proposals: an Advanced AI Framework calling for mandatory independent safety testing, transparency requirements, and government authority to block or recall dangerous deployments, and an Economic Policy Framework on managing AI’s labour-market effects. The survey does not sit apart from that agenda. It supplies the evidentiary base for it.
Strategic Reality: When the organisation being regulated also runs the survey that shapes the regulation, the data can be impeccable and the framing still serve the sponsor. The methodology is the easy part to verify. The harder question is who chose the questions, and whose policy proposals the answers happen to support.
Consider what the public actually told Anthropic. Asked what would best ensure AI benefits humanity, respondents converged on holding AI companies legally liable for harm (47%) and prioritising safety over growth (44%). Over 70% want government involved in regulating AI, a bipartisan supermajority that barely moves across party, geography, or education. These are not fringe positions Anthropic is bravely amplifying. They are the settled preferences of a public that wants the industry constrained, surfaced by a member of that industry and routed towards frameworks that member has already written.
| What the survey found | Figure | Why it matters strategically |
|---|---|---|
| Trust AI companies to self-govern | 15% | Lowest of any institution tested; the legitimacy gap the survey addresses |
| Want government involved in AI regulation | 71% | Bipartisan supermajority; demand for external oversight is settled |
| Rank legal liability for harm as top action | 47% | Public wants accountability, not voluntary commitments |
| Rank safety over growth as top action | 44% | Direct endorsement of a “safety-first” lab’s positioning |
| Fear AI-driven job loss | 64% | The most common fear in every US state |
Manufacturing legitimacy is not the same as faking it
It would be lazy to read this as spin. The data looks robust, the disclosure of the trust figure is genuinely awkward for Anthropic, and there is real public value in a large, weighted attitudes survey that reaches non-users of AI rather than only the enthusiasts. The strategic move is subtler than deception. It is the conversion of a governance problem into a listening exercise.
A lab that publishes the public’s distrust, then says “the direction AI takes should not be set only by the companies building it,” positions itself as the reasonable party inside the industry, the one company willing to measure its own low standing and act on it. That is legitimacy earned through the performance of humility. It works precisely because the underlying research is credible. Bad data would be a liability; good data is an asset you can only spend once you own the instrument that produces it.
Critical Context: The finding that most flatters Anthropic is the one about “integrated users”, the roughly 6% who use AI daily for work and life. They are more trusting of every institution and less keen to slow AI down. The implied narrative is that trust arrives with familiarity, so the answer is more adoption, not more restraint. The survey is careful to note this probably reflects who early adopters already are. The framing invites the more convenient reading anyway.
There is a version of this that is straightforwardly good for the public: better evidence, transparently shared, is worth having. But evidence is never neutral about which decisions it makes easier. A survey built by a lab will tend to ask the questions a lab finds actionable, and to leave the more uncomfortable ones, about market concentration, compute ownership, or the terms on which these models are trained, off the instrument entirely.
The finding the UK should not ignore
For a UK audience the reflex is to note that this survey stops at the US border. Anthropic says it plans to expand outside the US “in the future”. Until then, the temptation is to treat the whole thing as an American conversation. That would be a mistake on two counts.
First, the substance travels. The single most durable finding, that fear of job loss falls as hands-on AI use rises (70% among non-users versus 54% among daily workplace users), maps directly onto the UK’s adoption problem. The Britons most anxious about AI are the ones least exposed to it, and that anxiety is highest among the more educated, whose work overlaps most with what the models now do. Any UK organisation rolling out AI is managing that gap between fear and fluency whether it names it or not.
Second, and more importantly, the UK has a choice about who produces its evidence base. British public-attitudes research on AI has tended to come from independent bodies such as the Ada Lovelace Institute and the Alan Turing Institute rather than from the model-builders themselves. That is a structural advantage worth defending. If the authoritative “public record” on how the UK feels about AI ends up being a vendor’s survey, expanded from the US as a marketing exercise, then British policymakers will be reading the public’s mind through a lens ground by the very companies the public says it distrusts.
Strategic Insight: The question is not whether Anthropic’s data is accurate. It is whether the UK wants its governance debate anchored to evidence commissioned by a US frontier lab, or to independent research it can scrutinise, replicate, and hold to account. Those are different foundations for policy, even when the numbers agree.
Who is exposed, and how
The legitimacy contest plays out differently for each group watching it.
| Stakeholder | What the Public Record does for them | The catch |
|---|---|---|
| Frontier labs | Supplies public backing for preferred, self-authored frameworks | Reinforces the very self-governance the public rejects |
| UK policymakers | Free, high-quality attitudinal data | Risks importing an agenda framed by the surveyed industry |
| UK businesses adopting AI | Clear signal on workforce fears and where to build trust | Must separate genuine insight from vendor positioning |
| The public and workforce | Their views measured and amplified | Measured on questions someone else selected |
The workforce row is the one with operational teeth. Job loss was the top-ranked fear in every US state, held by 64% overall, and cognitive dependency followed at 56%. Yet the survey suggests dependency is mostly anticipatory: only about a fifth of those worried about it would feel significant disruption if AI vanished tomorrow. For UK leaders, that is a usable distinction. The fear is real and worth addressing, but it is often a fear of a future state rather than a description of present reliance, which means it responds to communication and involvement, not just to reassurance.
What UK organisations and policymakers should do
For policymakers and public bodies:
- Treat vendor-published attitude surveys as input, not authority. Read Anthropic’s data, then ask what an independently commissioned UK equivalent would need to measure that a lab would have no incentive to ask.
- Protect and fund the independent evidence base. The credibility advantage of Ada Lovelace or Turing research over a vendor’s survey is precisely that it can interrogate the industry rather than represent it.
For organisations early in AI adoption:
- Close the exposure gap deliberately. The people most fearful of AI are those who have never used it; structured, low-stakes hands-on experience does more to lower anxiety than any all-staff reassurance email.
- Name the job-loss question rather than talking around it. Silence reads as confirmation, and 64% are already primed to expect the worst.
For organisations already running AI in production:
- Distinguish anticipatory fear from real dependency in your own workforce. Ask who would actually be disrupted if a tool disappeared, and you will usually find the anxiety and the reliance sit with different people.
- Watch the accountability signal. A public that ranks legal liability for harm as its top governance priority is a public that will not be satisfied by voluntary corporate commitments, and neither, eventually, will regulators.
Take Action: Run a five-minute internal version of the trust question. Ask your teams how much they trust the AI vendors you depend on to act in your interest. If the answer rhymes with 15%, that is a procurement and governance issue, not a training one.
The challenges that do not show up in the headline number
Four risks in this picture are easy to miss until they matter.
The first is evidentiary capture. If the most-cited data on public attitudes to AI is produced by the industry being governed, the terms of the debate quietly narrow to what that industry finds convenient to measure. Nothing needs to be falsified for the framing to drift.
The second is the trust-through-adoption trap. The survey’s integrated-user finding is easily bent into an argument that distrust is just unfamiliarity, curable by more usage. That may partly be true. It also happens to be the conclusion most favourable to a company that sells AI, and it sidesteps whether some distrust is a rational response to real concentration of power.
Hidden Cost: Outsourcing the public’s voice to a vendor is cheap up front and expensive later. The price is paid when a governance decision leans on evidence no one outside the industry can fully audit, and the losing side asks who wrote the questions.
The third is UK complacency. Because this survey is American, it is tempting to file it under “not our problem” until Anthropic’s promised international expansion arrives pre-framed. The better move is to build the domestic evidence base now, on British terms, before the vendor version becomes the default reference.
The fourth is the accountability gap between what the public wants and what any lab can offer. The public wants legal liability and safety prioritised over growth. A voluntary framework, however thoughtful, is not liability. Reading the survey as endorsement of any single company’s proposals confuses measuring a demand with meeting it.
The takeaway for the UK
The Anthropic Public Record is good research doing strategic work. Its data on job-loss fear, the usage gap, and the appetite for regulation is genuinely useful, and UK leaders should read it. Its deeper lesson is about ownership: the same survey that records a 15% trust score is an attempt to convert that distrust into a mandate, held by the party being distrusted.
Three things matter most for a UK audience. Use the findings, because the workforce signal is real and actionable. Refuse the framing, because a lab’s survey will always foreground a lab’s preferred remedies. And invest in independent evidence, because the country that lets the model-builders write its public record has quietly let them set the terms of their own oversight.
Britain does not have to choose between ignoring this data and being governed by it. The productive path is to treat Anthropic’s survey as a well-made artefact with an interested author, mine it for what it genuinely reveals, and make sure the authoritative account of what the UK public thinks about AI is written by someone with no product to sell. For more on how AI policy lands on UK organisations, explore our insights and latest news, or get in touch with a tip or a question.
Analysis based on Anthropic, “Results from the first Anthropic Public Record” (12 June 2026). Strategic interpretation and UK context by Resultsense.