Anthropic Research · Anthropic · 2026-03-18
A large multilingual qualitative study of how Claude.ai users actually use AI — what they hope for, what they're afraid of, what they're getting from it. 81,000 people across many languages, processed through analysis pipelines that would have been impossible to do by hand. The methodology is itself the headline: at this scale, you stop having "user research" and start having something closer to a slice of the discourse.
The findings are heterogeneous, which is the point. Different demographics want different things; "what people want from AI" turns out to be a worse question than "what does this person, in this country, doing this job, want." That's a useful corrective to product roadmaps written by a small group of researchers in San Francisco who all have similar jobs.
- The most-mentioned hopes are not the ones in lab marketing materials — practical-life support beats "AGI" by a wide margin - Concerns cluster around honesty, manipulation, and labor displacement, not Skynet - Non-English-speaking users surface use cases (translation-for-bureaucracy, language-as-access) that are invisible from the US dashboard - "I want to be smarter" is a real, widespread motivation — and it's not the same thing as "I want to be more productive"
If you're teaching AI to a non-specialist audience, this is a reality check on what they actually showed up to learn. The product-team vocabulary ("agents," "RAG," "tool use") is invisible to most users. What's visible is "help me write this," "help me understand this," "help me decide this." Build curriculum around those.