Anthropic repeats its global AI survey, but this time it promises to publish the answers unedited

By Anthropic · September 29, 2026. Anthropic has opened a new round of its opinion study on artificial intelligence, this time with an unprecedented condition: participants can decide to have their full interview published as is, unedited, so that anyone -researchers, journalists, regulators or the merely curious- can read it. The survey, which runs from September 29 to October 6, 2026, is open to Claude and Claude Code users (free, Pro and Max plans) with accounts at least two weeks old, and each session lasts about 15 minutes.

The interviewer is not a person but "Anthropic Interviewer", an AI system that asks about the most significant moments -good and bad- the user has experienced with the technology, and about what they would want AI to change in work, education, healthcare or public administration. It is the second edition of a similar exercise carried out in December 2025, in which 81,000 people took part and whose results, according to Anthropic itself, fed the Anthropic Institute's agenda and were presented to international leaders at the World Economic Forum. This time, instead of publishing only aggregate findings and a few scattered quotes, the company wants to release the full transcripts of those who give their consent, along with the country of origin of each conversation, though without account data such as name or email address.

The company itself admits the frictions of that approach. In its FAQ it acknowledges that, although it will not publish account data, it cannot guarantee that no one will be identified from seemingly innocuous details -age, profession, city, education- combined by another AI; it cites previous studies that have already managed to re-identify participants in Anthropic Interviewer interviews this way. It also makes clear that the decision to make an interview public is, in practice, irreversible: the company can delete its copy if asked, but not what third parties have already saved. And it bluntly acknowledges the underlying limitation: those who respond are Claude users, not a representative sample of the population, and those who additionally agree to publish their interview are an even more biased subset.

In general, an AI lab opening up its raw user-opinion data -beyond the usual executive summary- remains rare in the sector. Most generative AI companies communicate the social impact of their technology through press releases and studies with closed methodology; here anyone will be able to check whether the findings Anthropic publishes later match what people actually said, and researchers will be able to trace patterns the company itself never even thought to look for. The very framing of the exercise, moreover, rests on an idea we share: "how to weigh these benefits and risks should not be left to AI companies alone", Anthropic writes, in a sentence that verges on a kind of institutional self-criticism rarely seen in the sector.

Our reading is that the real value of this exercise will be measured later on, when the interviews are published and we see whether they actually change anything or remain a public relations gesture with a scientific veneer. The December precedent suggests it is not pure cosmetics: that study did leave a documented mark on Anthropic's own research agenda and reached international decision-making forums. But the sample bias the company itself acknowledges -only Claude users, and among them only those who agree to expose themselves publicly- limits what this kind of consultation can tell us about society as a whole; at best it is a high-resolution thermometer of a population already convinced to use AI, not a citizens' referendum.

What does strike us as revealing is the content of the questions: what people would like AI to change about work, school, healthcare or government. These are exactly the questions that matter in the transition we have long argued for at Zendoric -harsh in the short term, with job losses and real disruption, but with a horizon of abundance if managed well- and it is a good sign that one of the companies with the most influence over where that transition is heading is asking people directly, instead of assuming it already knows the answer. That it does so with an AI interviewing people about the impact of AI itself adds an almost recursive layer to the exercise, but it does not make it any less serious: on the contrary, it is a legitimate and relatively underexploited use of these models as a tool for large-scale social research.