Anthropic published what it calls the largest multilingual qualitative study ever conducted on AI. 80,508 Claude users across 159 countries, interviewed in 70 languages over one week in December 2025. The study was led by Saffron Huang, a research scientist on Anthropic’s Societal Impacts team.
The top-line number is clean: 81% say AI has already taken concrete steps toward their vision. But that number hides something more important. The average respondent holds 2.3 distinct fears about the same technology they say is working.
Five tensions, one person
Anthropic’s most original contribution isn’t the data. It’s the framework. The study identified five “light and shade” tensions where a benefit and a risk co-exist inside the same user:
Learning vs. cognitive atrophy. 33% cited learning as a benefit. 17% worried about losing the ability to think independently. Educators reported atrophy at 2.5 to 3 times the average rate. Tradespeople who used AI to learn new skills rarely experienced it. The difference: AI threatens cognition when it replaces effort, not when it feeds curiosity.
Better decisions vs. unreliability. 22% valued AI decision support. 37% worried about hallucinations and errors. Both groups grounded their views in experience, not speculation. Nearly half of surveyed lawyers reported firsthand unreliability.
Emotional support vs. dependency. Only 22% of respondents raised either topic, but the overlap was striking. Users who valued AI emotional support were 3x more likely to also fear dependency. A Ukrainian soldier described AI friends that pulled him back to life during combat. A South Korean user talked with Claude instead of a friend and lost that friendship.
Time-saving vs. illusory productivity. 50% cited time savings. 19% said verification burden ate the gains. A software engineer in Mexico described leaving work on time to pick up his kids from school. A Brazilian user had to photograph evidence to convince the model it was wrong.
Economic empowerment vs. displacement. 28% experienced economic benefits. 18% feared displacement. Of the five tensions, this one was the most speculative, with the highest rate of hypothetical rather than experienced concern.
The geography of optimism
The global split was revealing. In sub-Saharan Africa, only 24.2% held negative AI sentiment. In Western Europe, 35.6%. The wealthier the country, the more anxious the population.
The reason isn’t abstract. In developing regions, AI functions as what respondents called a “capital bypass mechanism.” An entrepreneur in Cameroon reached professional level in cybersecurity, UX design, and marketing simultaneously. In Central and South Asia, learning was emphasized at nearly double the global rate, because AI offers access to education that doesn’t exist locally.
In North America and Europe, the dominant use case was simpler: surviving the workload.
What this means for the people managing adoption
The study has a clear selection bias. Every respondent was an active Claude user already finding value. This isn’t what the general population thinks about AI. It’s what experienced users think.
That makes the 2.3 fears-per-person finding more significant, not less. These aren’t people who fear AI because they don’t understand it. They fear specific things because they use it every day.
Only 6.7% worry about existential risk. The fears that actually dominate are practical: unreliability (26.7%), job displacement (22.3%), and losing the ability to think independently (21.9%).
For anyone leading AI adoption inside an organization, the challenge looks different through this lens. The resistance isn’t coming from the people who haven’t tried it. It’s coming from the ones who have. And they’re not asking you to slow down. They’re asking you to hold both things at once: the value and the risk, without pretending either one away.
Related: Your AI Tools Are Multiplying. Your People Aren’t Keeping Up. digs into the cognitive overload side of this equation: what happens when the tools multiply faster than people can absorb them.