We Asked 81,000 People What They Want From AI

Opening

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

Most of us would probably reach for “work efficiency” first — auto-sorting emails, drafting reports, organizing data. I’d say the same. But a study Anthropic released last week made me pause. When they asked 81,000 people that exact question, the initial answers looked similar at first. “Professional excellence” came out on top at 18.8%. But then the AI interviewer pushed one step further: “And once that happens, what do you want to do with it?” That’s where the answers changed.

159 countries, 70 languages — the largest multilingual qualitative study¹​ ever conducted, they say. Today I want to walk through three findings from this research, along with what the methodology itself means.

Between the Surface Answer and the Real One

Let’s start with the numbers. These are responses to the question: “If AI could grant any wish, like a magic wand, what would you want?”

• Professional excellence — 18.8% (AI handles routine work, I handle strategic thinking)
• Personal transformation — 13.7% (mentor, coach, emotional support)
• Life management — 13.5% (managing schedules, focus, cognitive load)
• Time freedom — 11.1% (freeing up time for family, hobbies, rest)
• Financial independence — 9.7% (generating income, economic security)
• Social change — 9.4% (solving disease, poverty, climate problems)
• Entrepreneurship — 8.7% (giving a one-person company team-level capability)
• Learning and growth — 8.4% (personalized education, satisfying intellectual curiosity)
• Creative expression — 5.6% (breaking down the barrier between imagination and realization)

Productivity ranks first — no surprise there. What’s interesting is what happened after the follow-up question. The Anthropic Interviewer²​ didn’t stop at “what do you want?” It asked again: “What’s the real hope behind that wish?” And the answers from people who’d ranked productivity first changed completely. One office worker in Colombia said that thanks to AI, they’d become more efficient at their job — and that last Tuesday, instead of finishing more work, they cooked with their mother. A freelancer in Japan said that by spending less mental energy on client problems, they wanted to read more books.

From a go-to-market strategy perspective, this is a textbook case of the gap between surface needs and deep needs. Ask a customer “what do you need?” and they’ll name a feature. But ask “why do you need that?” three times in a row, and an entirely different desire surfaces. AI companies are marketing “30% productivity gains” — but what users actually wanted to do with that 30% was have dinner with their family.

Structurally, roughly a third of all responses wanted AI to create more room in life — time, money, cognitive slack. About a quarter wanted more meaningful work (not escaping work, but elevating its quality). About a fifth wanted to become better people — learning, healing, growing. The rest wanted to create something (creative expression) or fix the world (social change).

Light and Shade Within One Person

The finding that struck me most in this study is this: hope and anxiety don’t split people into separate camps — they coexist within the same person. The research team called this “Light and Shade.” Five core tensions kept surfacing.

Learning ↔ cognitive decline. 33% of respondents mentioned AI’s learning benefits. At the same time, 17% worried that depending on AI was eroding their own ability to think. One respondent in South Korea said they got good grades using AI-generated answers but actually learned nothing — and that this was when they felt the deepest guilt. Educators stood out here: teachers and academics reported directly witnessing cognitive decline at 2.5 to 3 times the average rate — probably because they’re watching it happen in their students.

Better decisions ↔ reliability problems. This is the only one of the five tensions where the negative side outweighed the positive. 22% said AI helped them make better judgments, while 37% said AI’s unreliability actively undermined good judgment. Both sides were rooted in direct experience — 88% of those citing benefits and 79% of those citing harm had lived it firsthand. The tension nearly doubled in high-stakes professions like law, finance, and medicine. Nearly half of lawyers reported directly experiencing reliability problems — yet lawyers were also the group that reported the most decision-making benefits. People are leaning on AI’s judgment and getting burned by it at the same time.

Time saved ↔ illusory productivity. Half of all respondents (50%) mentioned time savings — the most cited benefit of all. But 18% said expectations rose right along with it, leaving them busier than before. One freelance developer in France put it memorably: “The ratio of rest time to work time hasn’t changed at all — I just have to run faster to stay in the same place.” It’s the AI-era version of the Red Queen Effect.³​

Emotional support ↔ emotional dependence. The percentages here are small (16% versus 12%), but this was the most tightly entangled tension of all. People who said AI gave them emotional support were three times more likely to also worry about becoming emotionally dependent on it. One graduate student in the US confessed to telling Claude things they couldn’t even tell their partner — and that it felt like a kind of emotional affair.

Economic empowerment ↔ economic displacement. 28% mentioned economic opportunities through AI, while 18% worried about job displacement. Freelance creators sat at the most extreme position here: the share reporting real economic benefit (23%) and the share experiencing real threat (17%) were nearly tied. AI is simultaneously their tool and their competitor.

The overall ranking of concerns matters too. Reliability issues (26.7%) topped the list — hallucination,⁴​ inaccurate citations, the burden of verification. Next came employment/economy (22.3%), autonomy/agency (21.9%), and cognitive decline (16.3%). Employment and economic concerns were the single strongest predictor of people’s overall attitude toward AI — more than any other issue, this is what set the emotional temperature of how people feel about it.

Another Finding: The Methodology Itself

Honestly, I found the research method just as striking as the findings themselves.

In this study, AI (Claude) played the role of interviewer. It asked a fixed set of core questions, then adaptively generated follow-ups based on the responses. The collected answers were then coded across multiple dimensions by a classifier, also built on Claude — things like “what people want,” “whether they’re already experiencing it,” “what they’re worried about,” “their occupation,” and “their overall attitude toward AI.”

The comparison to prior large-scale qualitative studies makes this clearer. The USC Shoah Foundation’s Holocaust testimony archive holds about 52,000 accounts, collected over five years from 1994 to 1999. The World Bank’s “Voices of the Poor” project gathered roughly 60,000 people across 60 countries, also a long-running effort spanning much of the 1990s. Anthropic’s study collected 80,508 responses across 159 countries and 70 languages — in a single week.

Qualitative research has traditionally faced a tradeoff between depth and scale. Deep interviews only worked with small samples; going large-scale meant falling back on multiple-choice surveys. The AI interviewer shows a real possibility for breaking that tradeoff.

Of course, the limitations are clear too. First, sampling bias — these are all active Claude users, people who already find enough value in AI to keep using it. Second, question-order bias — the interview asked about positive visions first and concerns second, which may have shaped the answers. Third, and more fundamentally, we need separate verification of what biases crept in as AI generated the questions and classified the responses — nuances the classifier might have missed, interpretive differences shaped by cultural context, and so on.

Even so, this methodology poses a serious question for social science. Human coders’ inter-rater reliability isn’t perfect in traditional research either. If a study eventually compares that against the consistency of AI classifiers, the results could be quite interesting.

Oz’s Lens

There are two things I find genuinely interesting about this study. First, there’s a fairly wide gap between the AI industry’s marketing message and what users actually want. AI companies lead with “productivity,” “efficiency,” “automation” — and sure enough, people say they want exactly that at first. But peel back one layer, and what’s underneath is “time to pick my kid up from school,” “the space to cook with my mother.” To me, this is a textbook case of the gap between a product’s feature and a customer’s outcome. “30% productivity gain” is a feature. “Cooking with mom on Tuesday evening” is an outcome. If AI companies want to reach the next real stage, they need to start promising outcomes, not features.

Second, I want to highlight the attitude of doing the research while openly acknowledging its bias. Anthropic spelled out the study’s limitations itself — that sampling only Claude users introduces bias, that question order may have swayed responses. And rather than treating that as a reason not to publish, they disclosed the limitations and pushed ahead anyway. The most dangerous thing in data analysis is claiming there’s no bias at all. Being upfront about the bias while still putting the findings out there is a far more honest research posture, in my view.

One more thing — several South Korean respondents were quoted in this study, and every one of them stuck with me. A software engineer who said “humanity has never had to deal with something smarter than itself.” A student who confessed to getting good grades off AI-generated answers while learning nothing at all. In a society with such high technology adoption as South Korea’s, I take this as a sign that the depth of reflection on technology is rising right alongside it.

Closing

One: what people want from AI isn’t ‘faster work’ — it’s ‘a better life.’ Productivity was the means, never the end. Two: hope and anxiety aren’t separate camps — they operate simultaneously within the same person. The people who used AI most skillfully were also worrying about it most deeply. Three: large-scale qualitative research using AI interviewers could become a genuinely new tool for social science. The limitations are real, but so is the potential.

The original study is published alongside an interactive data visualization. You can filter individual responses by region, concern type, and vision type. If you have the time, I’d recommend starting with the Quote Wall — it holds the texture that numbers alone can’t show.

References & Further Reading

• Saffron Huang et al., “What 81,000 People Want from AI”, Anthropic, 2026. : The full study, complete with interactive visualizations and a Quote Wall. I’d recommend reading individual responses rather than just the numbers.
• Anthropic, “What 81,000 People Want from AI — Appendix”, 2026. : An appendix covering the methodology, limitations, and additional analysis. A must-read if you’re interested in the research methodology.
• Kunal Handa et al., “Introducing Anthropic Interviewer: What 1,250 Professionals Told Us About Working with AI”, Anthropic Research, 2025. : The predecessor to this 81,000-person study — the first study to validate the AI interviewer methodology with 1,250 professionals.
• Deepa Narayan et al., Voices of the Poor: Can Anyone Hear Us?, World Bank, 2000. : A World Bank qualitative study gathering the voices of 60,000 people living in poverty across 60 countries — one of the “largest prior qualitative studies” this new research uses as its benchmark.

---

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. Qualitative Research: Unlike quantitative research, which measures things numerically, qualitative research digs deeply into people’s experiences, opinions, and feelings. If a survey counts “yes/no” answers, qualitative research asks “why do you think that?” ↩

2. Anthropic Interviewer: An AI-based interview tool developed by Anthropic. Claude asks a set of pre-designed questions, then adaptively generates follow-ups based on the respondent’s answers — a methodology aiming to combine the depth of traditional qualitative research with the scale of a survey. ↩

3. Red Queen Effect: A concept named after the Red Queen in Lewis Carroll’s Through the Looking-Glass, who says it takes all the running you can do to keep in the same place. It describes how, even as technology advances, expectations rise right along with it, leaving one’s actual sense of ease unchanged. ↩

4. Hallucination: When AI confidently generates information that sounds plausible but isn’t true — citing papers that don’t exist, or presenting incorrect figures as if they were accurate. ↩