What 81,000 People Want From AI

What 81,000 People Said They Want From AI

If someone asked you, “What do you want AI to do for you?”—what would you say?

My mind goes straight to work: sorting email, drafting reports. And in the interview study Anthropic published on March 18, 2026, “wanting to do more meaningful work” was indeed the largest category of responses. But reading through the individual answers, what caught my attention was what people wanted after the work was done. Some said they wanted time to cook with their mothers. Others just wanted the space to read a book.

The research team analyzed 80,508 interviews collected over one week in December 2025. Respondents answered in 70 languages from 159 countries. Keeping in mind that everyone involved was a Claude user who volunteered for the interview, let’s look at what this large-scale qualitative study¹ actually shows.

Response Rankings and Reasons from Follow-up Questions

The research team classified free-text answers to the question, “If AI could do anything for you, what would you want?” Each respondent’s answer was sorted into the single category that best captured their central wish. This isn’t a ranked vote respondents cast among the options below — it’s a categorization of open-ended responses.

• Doing more meaningful work better — 18.8%: reducing repetitive tasks to focus on what matters
• Personal transformation — 13.7%: changing oneself through advice and emotional support
• Managing daily life — 13.5%: organizing schedules and to-dos, easing the mental load
• Freeing up time — 11.1%: gaining time for family, hobbies, and rest
• Financial independence — 9.7%: securing income and financial stability
• Social change — 9.4%: tackling disease, poverty, and climate issues
• Starting a business — 8.7%: launching and running a venture with a small team
• Learning and growth — 8.4%: getting an education tailored to oneself and learning what one’s curious about
• Creative work — 5.6%: making the things one has always wanted to express

The Anthropic Interviewer² followed up on some responses to dig into the reasons behind them. One office worker in Colombia said that AI let them handle their work efficiently, and that the previous Tuesday, instead of catching up on leftover tasks, they were able to cook with their mother. A freelancer in Japan said they wanted to spend less mental energy on client issues so they could read more books. Behind the wish to “do work well” sat these very personal reasons.

I read this as a question of how to describe a product to customers. If a customer says they want to draft reports faster, you can also ask what gets better once that time is freed up. The same feature will carry different value for someone trying to focus on other important work versus someone trying to leave the office on time. You can’t assume every user’s underlying purpose is time with family — but the point of this kind of interview is precisely to hear the reason behind the desired feature, not just the feature itself.

The research team’s broader groupings tell a similar story. Roughly a third wanted more time, money, or mental bandwidth; roughly a quarter wanted more fulfilling work; and roughly a fifth wanted learning and personal growth. In other words, the desire to do less work and the desire to do work better show up side by side in the same data.

Hope and Worry Showing Up in the Same Person

What struck me most was that a single person’s answer often contained both the benefits and the worries about AI at once. The research team separately analyzed whether five paired benefit-and-harm categories appeared across the full interviews. The percentages below combine experiences people had already had or witnessed with expectations and concerns about the future. Interviews that didn’t answer the concern question were excluded from this analysis.

AI helps people learn, but it may reduce practice in thinking for oneself. Mentions of learning benefits came to 33%, while mentions of a possible weakening of one’s ability to think or study independently came to 17%. Only about 8% of the analyzed responses described actually experiencing or witnessing such weakening; the rest were forward-looking concerns. One respondent in Korea said they got a good grade by memorizing AI’s answers, then felt guilty because they hadn’t actually learned the material themselves. Among teachers and academics, the share reporting having directly witnessed this problem was about 2.5 to 3 times the overall average. This reflects respondents’ reported experiences, not a decline measured through cognitive testing.

People get help making decisions, yet also run into trouble from wrong answers. Mentions of better decision-making came to 22%, while mentions of reliability problems—such as misinformation—came to 37%. Of those who mentioned the former, 88% were describing actual experience; of those who mentioned the latter, 79% were. This issue stood out especially in fields where errors carry high stakes, like law, finance, and medicine. Among lawyers, many reported experiences of getting help with judgment, yet nearly half also said they had experienced reliability problems.

Work gets faster, but that doesn’t necessarily mean more time to rest. 50% mentioned time savings—37% from actual experience, 13% as future expectation. Meanwhile, roughly one in five people mentioned that verifying AI’s answers takes time, or that their workload had simply grown, so they didn’t actually feel more productive. One freelance developer in France said the ratio of work to rest stayed the same—they just had to work faster. It’s a reason to look separately at how fast a task gets done and how much breathing room someone actually has in a day.

Comfort and worry about dependency appeared together, too. 16% mentioned emotional support, and 12% mentioned dependency concerns. The two categories co-occurred about 3 times more often than would be expected if each were independent. That doesn’t mean using AI triples your risk of dependency. One graduate student in the U.S. said they confided in Claude about things they couldn’t tell their partner, and it felt like a kind of emotional affair.

Chances to earn more overlapped with fears of losing work. 28% mentioned economic opportunity, while 18% mentioned economic threats such as job displacement. Among creative-field freelancers, 23% reported actually experiencing economic benefit, while 17% reported experiencing instability. Even within the same occupation, people described both getting more done thanks to AI’s help and worrying that clients might just do the work themselves with AI, cutting them out.

Looking only at answers to the question “what worries you,” the breakdown was: reliability problems 26.7%, employment/economic concerns 22.3%, autonomy/agency 21.9%, and weakening of independent thinking 16.3%. The five-pair analysis above draws on full interviews, so the percentages differ. When one person raised multiple worries, all were counted. Reliability problems include hallucination³—AI generating answers that aren’t factual—incorrect citations, and the effort of verification.

Concerns about employment and the economy showed the strongest link to the overall AI attitudes the research team categorized. This suggests job-related concerns are closely tied to people’s overall evaluation of AI, but this analysis alone can’t confirm that employment concerns shift attitudes more than other worries do.

The Research Method and Its Limits: When AI Plays Interviewer

Honestly, I found myself paying as much attention to the research method as to the results themselves.

Claude asked the baseline questions, then generated follow-up questions tailored to each response—probing what people hoped for, what they’d actually experienced, and what worried them. Claude was also used to classify the answers: pulling themes out of long, freely written responses and tallying them up.

Over one week in December 2025, the study collected 112,846 interviews. After excluding joke answers and responses that were too short, 80,508 were left for analysis. What stands out is how many responses were gathered in such a short window—enough to rival established large-scale qualitative and oral-history studies.

Interviews are good at probing the “why” behind an answer, but asking questions and reading through responses takes enormous time. An AI interviewer shows that this work can be done at scale. Still, sheer volume and collection speed don’t by themselves prove the researchers achieved the same depth as a human researcher’s in-depth interview.

There are limits worth noting. Participants were volunteers drawn from existing Claude users—they can’t represent the views of people who don’t use AI at all, or who’ve stopped using it. And the order of the questions—asking about hopes before worries—may have nudged people to answer the two in a connected way.

Human checks were built into the classification process. The research team reported that for each classifier, humans reviewed 25 judgments and verified over 90% agreement. Even so, it’s hard to claim that every nuance and cultural difference across such a massive set of responses was captured accurately. How far to trust the automatically classified results is something that needs ongoing scrutiny.

Even human researchers can interpret the same answer differently. Going forward, I’d like to see the same interviews classified separately by humans and AI, compared to see where—and on what kinds of answers—the gaps widen. Only once that kind of verification accumulates can we judge which kinds of research this tool is actually suited for.

Oswarld’s Lens

From the perspective of GTM strategy—how you introduce a product to customers—I found myself looking at features alongside the reasons people want them. Automatic email sorting is a feature; cutting down processing time is a work outcome. Spending that freed-up time with family or on more important things is the life change the user actually wants. I think you have to go a step further than saying “it boosts productivity” and spell out, concretely, whose work changes and how.

Working in data analysis myself, I also paid attention to the researchers’ willingness to disclose the study’s limitations. Stating the sample and the order of questions lets readers know the boundaries within which they should interpret the results. Noting limitations doesn’t solve every problem, but not hiding them is what makes it possible to discuss, in the next study, what still needs to be addressed.

The comments from Korean respondents stuck with me too. One software engineer worried about how to deal with something smarter than a human, while another respondent said that even after getting good grades with AI’s help, they didn’t feel like they’d actually learned anything. A handful of comments can’t explain all of Korean society. Still, the voices grappling with how to protect their own judgment and learning while enjoying the convenience weren’t something I could brush aside lightly.

Closing

Reading this study, I realized that when evaluating how fast AI helps someone work, we also need to ask how the user’s whole day changed. Time saved writing a report can be eaten up by time spent verifying it, and finishing a task faster can simply mean more work gets piled on. You have to ask the same person both what helped and what got worse to understand what a product actually changed.

The original report includes an interactive visualization and a “Quote Wall.” You can read individual published quotes from respondents, organized by region and topic. It’s not a release of the full interview transcripts, but reading it alongside the aggregate numbers helps you understand the context in which respondents said what they said.

This fragment is correctly translated with no Hangul, matching numbers, links, headings, and footnotes. No corrections needed.

References & Further Reading

• Saffron Huang et al., What 81,000 People Want from AI, Anthropic, March 18, 2026. Contains the survey findings and quotes from individual responses.
• Anthropic, Research Appendix, March 2026. Explains the sample, classification methods, limitations, and additional analysis.
• Kunal Handa et al., Introducing Anthropic Interviewer: What 1,250 Professionals Told Us About Working with AI, Anthropic, December 4, 2025. An earlier study that piloted the interview tool on 1,250 working professionals.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. Qualitative research: research that reads through how people describe their own experiences and thinking in their own words, and explores the meaning and context behind them. Some findings are also tallied numerically. ↩

2. Anthropic Interviewer: an interview tool in which Claude asks questions according to a plan set by researchers, then poses follow-up questions tailored to the answers given. ↩

3. Hallucination: a phenomenon where AI plausibly generates information that isn’t factually true — including citing papers that don’t exist or presenting incorrect figures. ↩