Skip to content

Anthropic

Claude interviewed almost 81,000 people in 70 languages

2 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: anthropic.com

Anthropic on October 6 described how Anthropic Interviewer, the Claude-powered tool behind its almost 81,000-person global pilot, works. Claude drafts the interview guide, holds each interview and makes the first pass at the analysis, so the findings Anthropic publishes start as Claude’s synthesis, which researchers then validate and interpret.

Clio, the tool Anthropic’s earlier research relied on, has Claude sort conversations into topic clusters, and the clusters are all human analysts see. That showed a great deal about what people use Claude for and far less about how Claude’s outputs affect them or how they feel about AI, the gap the interviewer was built to fill.

Researchers set a study’s questions and goals, and Claude drafts a guide grounded in best practices, which the team refines and finalizes. In each interview Claude talks with the participant in real time, in their own language, adjusting its follow-ups to what they share. The analysis synthesizes the answers against the original research plan into emergent themes and illustrative quotes. Each participant is told how their answers will be used before agreeing to take part.

Several studies have run on the tool, from a focused test of how workers, scientists and creatives use AI at work to the global pilot, where people in 159 countries and 70 languages described in their own words how AI has affected them. For that pilot Anthropic invited everyone with a Claude.ai account over one week in December 2025, and 80,508 people took the interview, so its figures describe Claude.ai users who chose to answer. Claude-powered classifiers then tagged each conversation by what the person wants from AI, what they fear and their overall sentiment.

The results sit in separate survey write-ups. In the global pilot the largest group, 19%, sought “professional excellence” from AI. Each conversation was coded with a single primary want but could carry several concerns, so the shares for wants never overlap while those for concerns can.

Responses in that pilot were de-identified before a small team of researchers analyzed them, quotes picked for publication got a further manual review, and even the names of other AI products were redacted.

The insights feed Anthropic’s Societal Impacts research and the way it builds Claude, and go to policymakers, independent researchers and the public. From the first test, which came with the tool’s launch on December 4, 2025, Anthropic released all 1,250 interviews, with participants’ consent, for researchers to explore.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.