What 81,000 People Want from AI

Key Findings

• — 112,846 interviews were collected in one week of December 2025; 80,508 met the quality threshold, from respondents in 159 countries writing in 70 languages.
• — Most frequent concerns (multi-label): unreliability 26.7%, jobs and economy 22.3%, autonomy and agency 21.9%, cognitive atrophy 16.3%; 'wellbeing and dependency' (social isolation, loneliness, negative psychological impacts, compulsive AI use, preferring AI companions to humans) 11.2%.
• — Emotional support was the primary thing 6.1% of respondents wanted from AI. In the benefit-harm analysis, 16% mentioned emotional support as a benefit (13% had experienced it, 3% expected it) and 12% mentioned emotional dependence as a harm (5% had seen it, 7% expected it).
• — A respondent who values emotional support from AI was three times more likely to also fear becoming dependent on it; this pair had the strongest within-person co-occurrence of the five benefit-harm tensions measured. People not currently working were twice as likely to raise it and twice as likely to describe some experience of dependence; healthcare professionals described using Claude for emotional support at twice the rate of other professionals.
• — 33% mentioned learning benefits and 17% worried about cognitive atrophy; 46% of those worried about atrophy said they had seen it firsthand. 37% said unreliability impedes good decisions, against 22% citing better decision-making.
• — 67% of interviewees expressed net positive sentiment toward AI (5 or above on a 1-7 scale); no country fell below 60%.

Methodology Notes

Opt-in conversational interviews of existing Claude.ai users (all account holders invited for one week in December 2025; exact dates not stated). Four core questions (last AI chatbot use; what AI would ideally do; whether AI has moved toward that; ways AI might be developed contrary to the respondent's values) with adaptive follow-ups. Transcripts were de-identified and coded by Claude-based classifiers, each validated at 90% or higher agreement with a human on 25 labels; the 'want' dimension is single-label and concerns are multi-label; respondents who did not reach the concerns question were excluded from that analysis. The authors checked representativeness against Claude.ai weekly active users by tier and region and against the Anthropic Economic Index usage mix. Stated limitations: the sample is opt-in Claude users and likely skews positive; question order (hopes before concerns) may inflate benefit-harm co-occurrence; label ambiguity; occupations are self-reported. Answers describe AI use generally, with other products' names redacted. Correction of 2026-03-19 reworded the 67% sentiment figure. Methods and limitations are in a separate appendix PDF (March 2026).

Sources