What 81,000 People Actually Want From AI

Anthropic interviewed 81,000 Claude users across 159 countries. Hope and fear live in the same person. Here is what that means for anyone rolling AI out to a workforce.

EXECUTIVE SUMMARY

Every leader rolling AI out to a workforce in 2026 is implicitly answering three questions — what do people want from this, what are they afraid of, and what happens when those two things live inside the same person? Until now, every answer to those questions came from a checkbox survey. Anthropic did something different earlier this year.

// Anthropic published the largest qualitative AI study ever conducted — 81,000 open-ended interviews across 159 countries and 70 languages, with Claude itself doing the interviewing. The methodology is the story almost as much as the findings.

// The defining finding is Light and Shade: hope and fear are not separate populations. They live inside the same person, as opposite poles of the same handful of concerns — learning vs. cognitive atrophy, empowerment vs. displacement, connection vs. dependence. Someone who valued emotional support from AI was 3x more likely to also worry about becoming dependent on it.

// The leaders winning the next two years won't be the ones with the strongest pro- or anti-AI position. They'll be the ones whose AI rollouts acknowledge that the average user is both hopeful and afraid at the same time — and whose communications, training, and tooling are built for ambivalence rather than either enthusiasm or resistance.

This deep dive walks the full study — hopes, fears, regional splits, productivity claims, the economic-anxiety data — and ends with a four-phase playbook for AI rollouts that take the Light and Shade finding seriously.

I'll admit something before we get into the data.

When I read the title — "What 81,000 people want from AI" — my first instinct was that this would be another vendor study where Anthropic asks Claude users if they like Claude, and the answer is yes. That's the genre. Most "AI sentiment" research published in the last twenty-four months has been somewhere on the spectrum between marketing and motivated reasoning.

This isn't that.

What caught me was the methodology. Anthropic's Interviewer didn't ask people to pick from a list. It conducted open-ended interviews — the kind a good qualitative researcher would conduct — at a scale no qualitative researcher has ever operated at. 81,000 people. 159 countries. 70 languages. White-collar workers in Colombia. Entrepreneurs in Nigeria. Students in Indonesia. Tradespeople in the United States. People who don't normally show up in AI research because no one has ever covered the labor costs of interviewing them at scale.

The whole study reads like the first piece of AI research designed for an actual planet of users instead of a Silicon Valley conference. That alone made it worth a Friday afternoon.

The second thing that caught me was the Light and Shade framing. The mainstream AI discourse is binary: enthusiasts on one side, doomers on the other, with a comms layer trying to convince each that the other is wrong. The 81k data says that binary is a fiction. The same individual who is excited about AI is also afraid of it, and the framing they bring to your rollout will depend on which side you speak to.

That's the entire problem with most AI strategy decks I've reviewed in 2026. They are written for a workforce that does not exist — a workforce that has already made up its mind. The actual workforce is ambivalent, and our communications are built for everyone except them.

So here is what 81,000 people actually said, what it means, and what to do about it.

The 81k study has three layers that matter for any leader making AI decisions in 2026: what people hope AI will do for them, what they're afraid of, and what they say has actually happened since they started using it. Each layer has a top-line stat. Each layer has a counter-stat that complicates the top-line. The art of reading this study — and building a strategy on top of it — is holding both at once.

One note before going further. This is a study of Claude users who agreed to be interviewed by an AI. That sample skews toward people who are already AI-curious and AI-comfortable. The findings below are best read as the upper bound of where the rest of the population is headed, not as a representative snapshot of "all employees today." The directional signals are strong. The exact percentages are treated as indicative.

The number that should be on every AI rollout deck: 19% of respondents named professional excellence as their top hope. Not productivity. Not efficiency. Professional excellence — the ability to clear away the mundane so they can focus on the work they were actually hired to do well. I like this; I think it’s a qualitative, not a quantitative, measure, though, which makes it harder to figure out.

The number two hope was entrepreneurship at 9% — AI as a partner for building or scaling a business. That number jumps dramatically in lower-income regions, where AI is described as a capital bypass mechanism — the thing that lets a Nigerian solo founder or a Kenyan freelancer access skills, tooling, and markets that were previously gated by money, credentials, or geography.

Beneath the headline hopes, three patterns deserve attention.

Pattern 1: Time savings is the dominant concrete benefit, named by roughly half of the respondents. This tracks with every other AI productivity study. What the 81k data adds is what people want to spend the time on. The answer is rarely "more work." The most-cited downstream benefit was time with family, time for rest, time to think. The quote that should be on every AI strategy deck:

The product isn't "productivity." The product is time with mom. Any AI rollout whose only promise is "do more with the same people" is selling the input, not the outcome users actually want.

Pattern 2: The regional split is sharper than expected. In wealthier regions, people described AI primarily as a life-management tool — a way to reduce the cognitive overload of email, scheduling, decisions, and the mental tax of modern white-collar life. In developing regions, people described AI primarily as an entrepreneurship and learning tool — a way to build skills and businesses that previously required capital they didn't have. Same product. Two completely different jobs-to-be-done.

For any multinational rolling AI out globally, a single "AI will give you more time for what matters" message will land cleanly in Copenhagen and miss entirely in Lagos. The reverse is also true.

Pattern 3: 81% said AI has already taken a real step toward their vision. This is the most important deployment statistic in the study and the one most likely to be misread. It does not mean 81% of users are satisfied. It means 81% of users are already past the " Should I try this question and are now in the " How do I get more of what I want from it question. The decision is made. The rollout you're planning isn't introducing a new tool to a workforce — it's catching up to a workforce that has already deployed AI on their personal accounts in their own browsers, without IT.

If there is a single statistic from this study that AI vendors don't want to talk about, it's this one: roughly 22% of respondents named loss of autonomy and agency as a top concern, with forced adoption called out by name. One student in the study described it like this: "The line isn't something I'm managing — it feels like Claude is drawing the line."

Three fear patterns matter.

Pattern 1: Job displacement fear is calibrated, not generic. Anthropic's economic-impact follow-up shows respondents in the top 25% of observed AI exposure mentioned displacement concerns roughly 3x more often than the bottom 25%. For every 10-point increase in exposure (the percentage of a job's tasks Claude is observed performing), perceived job threat increased by 1.3 percentage points. Software engineers worry significantly more than elementary school teachers, because the data shows Claude is doing more of a software engineer's work than a teacher's.

The operator implication: employees most worried about AI displacement are not irrational. They are doing approximately accurate math about their own exposure. "Don't worry, AI won't replace you" is the wrong message because in the most exposed roles, workers already know it's replacing parts of them. The better message is honest: AI is doing more of the tasks in this role; here is what the role becomes; here is the path we're investing in for you.

Pattern 2: Cognitive atrophy is the educator-flagged fear. Educators in the study reported witnessing skill loss due to AI overreliance among their students at two to three times the average rate. Tradespeople using AI on their own terms reported it at roughly 4%. The variable isn't the technology. The variable is choice. People who adopt AI because they want to retain the underlying skills. People who adopt AI because they were told to seem to lose them.

This has a direct rollout consequence. How you introduce AI changes what AI does to the people you introduce it to. "Here, you must use this" is a different intervention than "here, you can use this for the parts you want." The 81k data suggests the second framing actually protects skills the first one erodes.

Pattern 3: "Illusory productivity" is the most under-discussed finding. 18% of respondents described what the study calls illusory productivity — the experience of saving time without ever experiencing the time saved. One freelancer put it bluntly: "You just have to run faster and faster to stay in place." I observe this myself, I have become the bottleneck, I have myriad unfinished projects because I am not able to spend the time required for the human to bring these across the line.

If your AI rollout's promise to employees is "you'll be more productive," but the lived experience is "you'll do the same hours and produce more output that someone else captures the value of," the trust collapse is not far behind. Communications that promise a benefit users don't experience are worse than communications that promise nothing.

Anthropic's economics follow-up adds a productivity-and-power dimension to the picture.

The average self-reported productivity rating across respondents was 5.1 on a 1–7 scale, corresponding to "substantially more productive." 3% reported negative or neutral impacts. 42% didn't give a clear productivity answer.

More interesting is how people got more productive. Anthropic separates gains into four categories: scope, speed, quality, and cost.

The scope-vs-speed split has direct strategy implications. If your AI ROI model measures only speed — minutes saved per task — you are measuring the smaller of the two effects. The larger effect is people doing work they couldn't previously do at all, and most ROI models don't capture it because the baseline didn't exist.

There is also a power dimension. When asked who benefited from the productivity gains:

The split by career stage is sharp. Only 60% of early-career workers said they personally benefited from AI, versus 80% of senior professionals. Translation: the people most worried about displacement are also the ones least likely to be capturing the upside. That is not a coincidence. It is the Light and Shade finding in a single data point.

// Key Takeaways

Hope and fear are not separate populations — they live inside the same person. Build communications, training, and tooling for the high-hope/high-fear majority. Materials designed for enthusiasts will land for a small slice of the workforce and fail for the rest.

Time savings only count if the time gets returned somewhere employees can feel. Eighteen percent of users in the largest AI study ever conducted describe "illusory productivity" — same hours, more output, no benefit captured. If your rollout's promise is time savings, you need an explicit time-return plan.

Choice protects the skills that mandates erode. Educators in the study saw 2–3x the average cognitive-atrophy rate in students using AI; self-directed tradespeople saw 4%. Default to opt-in tooling; reserve mandates for narrow, well-justified workflows; measure quality of use, not just adoption.

Scope expansion is roughly equal in size to speed gains — and most ROI models miss it entirely. Add "new work enabled" as a measured line item. The non-developer who now ships full-stack code, the marketer who now designs product, the analyst who now builds models — that is the bigger half of the productivity story and the one with the longest tail.

The upside is concentrating where the power already is. Distribute it deliberately, or pay for it later. Sixty percent of early-career workers say they personally benefit from AI versus 80% of senior professionals. Same productivity gains, asymmetric capture. Fix it explicitly — through assigned scope, training, and compensation — or watch retention degrade in the population you most need.

The temptation in 2026 is to build AI strategy around the parts of the workforce that are easy to plan for — the enthusiasts who self-onboard, the senior professionals who already capture the gains, the use cases that fit a clean productivity narrative. That strategy will succeed for those segments and quietly fail for everyone else, which is most of the company.

The 81k study is the first time we have at-scale data that, in users' own words, show what the rest of the workforce is actually experiencing. It is more hopeful than the doomer narrative and more honest than the vendor narrative. It says: people want this, people are afraid of it, the same people hold both feelings, and the rollout that ignores either side will lose them.

The most likely outcome of an AI rollout in 2026 isn't "workforce embraces it" or "workforce rejects it." It's something messier in the middle: most people use it, many people benefit, a nontrivial share feel the treadmill speed up, an even smaller share quietly disengage, and the gap between the adoption dashboard and the lived experience widens quarter over quarter until somebody finally measures it.