What 81,000 people told us about the economics of AI

economic indexlabor marketjob displacementproductivitysurveysanthropic interviewerclassifiers

Massenkoff and Huang use Claude-based classifiers to code economic themes in about 81,000 open-ended interviews with Claude users, finding that worry about job displacement rises with an occupation's observed AI exposure and is higher early in careers, that reported productivity gains are highest at the top and bottom of the wage distribution and come mostly from doing new tasks, and that the users reporting the largest speedups also worry most.

Maxim Massenkoff, Saffron Huang · 2026-04-22 · original

Anthropic's Economic Index measures what work Claude is asked to do and in which occupations it handles the largest share of tasks. What it could not show, the authors write, is how those usage patterns relate to "people's thoughts and impressions of AI". This report, published on 22 April 2026 by Maxim Massenkoff and Saffron Huang of Anthropic's Economic Research team, connects the two. It reanalyzes an earlier Anthropic survey in which about 81,000 Claude users described their hopes and fears about AI in open-ended interviews, and asks three questions: who worries about losing their job to AI, who reports benefiting, and how the kind of benefit relates to the worry.

The report was published as a web post and a 15-page PDF with an appendix; page numbers below refer to the PDF. It appeared alongside Announcing the Anthropic Economic Index Survey, which launches a monthly survey built on the same interview tool.

The data and how it was coded

The underlying data come from Anthropic's earlier study "What 81,000 people want from AI", run through Anthropic Interviewer; the acknowledgements thank "the 80,508 Claude users who shared their stories" (p. 12). The appendix (p. 14) refers readers to that study and its appendix for the sample, survey method and limitations, and says this report uses the same data with additional variables.

Those variables were not asked directly. They were inferred from the free-text answers by Claude-powered classifiers:

  • Occupation. Inferred mainly from the first survey question ("What's the last thing you used an AI chatbot for?") and incidental mentions (footnote 1). Claude was told to leave occupation blank when unclear, which happened for 61% of respondents. Of the 39% with a label, about 11% of all respondents stated their occupation explicitly and 28% had it inferred from clues (appendix, p. 14).
  • Career stage. Inferred for about half of respondents (p. 4), for example from mentions of homework (early career) or of running a business or hiring (senior) (footnote 2).
  • Job threat. A respondent was coded as indicating job threat if they said their role was already being replaced or substantially reduced, or that this was likely soon (Figure 1 caption).
  • Productivity. Rated by Claude on a 1 to 7 scale where 1 is "less productive", 2 is "no change" and higher values are larger gains up to 7, "transformatively more productive". Footnote 3 explains that the scale is uncentred because nearly all answers scored 6 or 7 on the original Likert scale.
  • Type of gain and speedup. Gains were classed as scope, speed, quality or cost; speedup was coded from 1 (much slower) through 4 (no change) to 7 (much faster) (pp. 8 to 9).

Observed exposure, the measure set against these answers, comes from Massenkoff and McCrory (2026), Labor market impacts of AI; this report describes it as reflecting "the percentage of a job's tasks for which Claude is used" (p. 3).

What it finds

Worry about displacement tracks exposure (pp. 3 to 4, Figure 1)

One fifth of respondents voiced concern about economic displacement. Across occupations, the share expressing job threat rose with observed exposure: each 10-percentage-point increase in exposure went with a 1.3-point increase in perceived threat, and people in the top quarter of exposure mentioned the worry three times as often as those in the bottom quarter. The report's example is that elementary school teachers worried less than software engineers, consistent with Claude usage skewing toward coding. The authors read this as "consistent with people being broadly aware of AI's diffusion and potential impacts" (p. 2).

Early-career respondents worry more (p. 4, Figure 2)

Early-career respondents were "much more likely" than senior workers to express concern about displacement. The report links this to its earlier tentative evidence of slower hiring of recent graduates and early-career workers in the United States. Figure 2 shows the shares by stage; the text gives no figures.

Productivity gains (pp. 5 to 7, Figure 3)

The mean productivity rating was 5.1, "substantially more productive". About 3% reported negative or neutral effects, and 42% gave no clear indication. By wage quartile (left panel of Figure 3), the highest-paid occupations reported the largest gains, and the report says this holds when computer and math occupations are excluded; some of the lowest-paid workers also described large gains, for instance using Claude for side businesses. By occupational group, management ranked highest, mostly entrepreneurs building a business, and remained tied with computer and math occupations when these "solopreneurs" were removed (footnote 4). Scientific and legal workers reported the mildest gains; some lawyers worried about AI's ability to follow precise instructions.

The authors caution that respondents were active Claude users willing to take a survey, which "could make them more likely to report productivity benefits than the average user" (p. 5).

Who gets the gains (p. 7, Figure 4)

About a quarter of interviews named who benefits from the respondent's productivity gains. Most of these named themselves. Ten percent of those who named a recipient said employers or clients were demanding and getting more work; fewer named AI companies, and fewer still called AI a net negative. Career stage mattered: 60% of early-career workers said they personally benefited, against 80% of senior professionals. Footnote 5 notes that enterprise users, absent from this sample, might be more likely to say value goes to their employers.

Scope more than speed (pp. 8 to 9, Figure 5)

Among respondents who mentioned productivity effects, 48% cited scope (doing things they could not do before) and 40% cited speed. Quality gains often came from more thorough checking of code, contracts and paperwork; a small share mentioned lower cost.

Speedup and worry form a U shape (pp. 9 to 10, Figure 6)

Respondents who said AI slowed them down were more likely to see it as a significant threat; the examples include fine artists and writers who found AI too rigid for their own work but feared its spread into their fields. Among the rest, perceived threat rose steadily with reported speedup. The authors suggest a reason: when the time a role's tasks require shrinks quickly, its future is less certain.

The prompts used

The appendix (p. 15) prints two of the classifier prompts. Each is applied to a full interview transcript and returns a single label. The productivity prompt asks "Along which dimension does this respondent say AI improves their work?", tells the model to pick the one the respondent leans on most, and defines speed, quality, scope, cost and not_discussed in a line each (scope, for example, is "doing things they COULDN'T do before"). The career-stage prompt asks the model to "Use any signal you can find", lists cues such as job title and whether the person manages others, and offers five labels from student_or_entry to executive, plus unclear. The report notes that results for Figures 1 and 3 were qualitatively the same when restricted to the 11% who stated their occupation explicitly (p. 14), a check on the occupation inference; it reports no comparable check of the other classifiers against human coding.

Limitations

The authors give the first three in the Discussion (p. 11); the fourth follows from how the report describes its main result.

  • Sample. Only users of personal Claude.ai accounts who chose to respond; they may be more inclined to see benefits as flowing to themselves.
  • Inferred variables. Occupation, career stage and other variables were inferred from contextual clues and "could be wrong".
  • Open-ended answers. Measures depend on what respondents happened to mention, so the findings "should be confirmed in structured surveys that ask about these topics directly".
  • Correlational design. The report presents the exposure result as "initial evidence" that observed exposure is correlated with economic concern (p. 2); it does not estimate a causal effect.

The authors also state that the large share of economy-related concerns in an open survey about AI "is a strong signal in itself" (p. 11).

Research questions

  1. Does the exposure gradient hold in a structured survey? The authors call for confirmation with direct questions. Asking a sample of respondents their occupation and perceived job risk explicitly, and comparing the slope with the 1.3 points per 10 points of exposure found here, would test it.
  2. How accurate are the Claude-inferred variables? Hand-coding occupation, career stage and productivity rating on a random sample of transcripts would give error rates for each classifier and show whether errors vary by group.
  3. Why do slowed-down users worry most? The U shape is shown for one survey. Following the users who reported slowdowns, by occupation, would show whether the pattern reflects creative fields specifically or a broader group.
  4. Is the scope effect stable as users gain experience? Repeating the scope versus speed classification in the monthly survey would show whether gains shift from new tasks toward faster existing tasks over time.
  5. Do enterprise users report different beneficiaries? Footnote 5 expects they might. Running the same beneficiary coding on interviews with enterprise users would answer it directly.

How it connects

References

  • Massenkoff, M., and Huang, S. (2026). What 81,000 people told us about the economics of AI. Anthropic, 22 April 2026. https://www.anthropic.com/research/81k-economics. PDF: https://cdn.sanity.io/files/4zrzovbb/website/3a8d990bc90098038eabd77b0d12ff636ed58d50.pdf (page numbers in this note refer to the PDF).