Deep Tide Guide: This is the first large-scale survey by an AI company exploring users’ real economic anxieties. The data reveals a cruel paradox: programmers, designers—those who use AI most—are precisely the ones most worried about being replaced by AI; those experiencing the fastest efficiency gains feel the least secure about the future. For investors, this means AI is penetrating faster than expected, and its impact on the job market has already begun on a psychological level.

Key Findings:

Our recent survey of 81,000 Claude users shows that those engaged in jobs more susceptible to AI replacement are more concerned about AI-induced unemployment. Early-career respondents are especially so.

Jobs with the highest and lowest incomes report the greatest productivity improvements, mainly due to expanding work scope (taking on new tasks).

Respondents experiencing the fastest speed improvements from AI are actually more worried about unemployment.

To help the public understand the observed shifts in the AI economy, our economic index shares what kinds of work Claude was asked to do and in which tasks Claude completed the largest proportion. But so far, we lack information on how these usage patterns map to people’s perceptions and impressions of AI.

Our recent survey of 81,000 Claude users provides a way to connect people’s economic concerns with the quantified content in Claude traffic.

The survey asks people about their visions and fears regarding AI progress. Many responses involve economic topics. We learn that many people worry about unemployment—despite feeling more efficient and capable at work. In some cases, AI enables them to start businesses or gives them time to focus on more important things; in others, AI feels oppressive or is imposed by employers.

The results provide preliminary evidence that the observed exposure level (our measure of AI substitution risk) correlates with economic worries about AI. Those in high-exposure occupations—defined by tasks observed to be performed by Claude—are more anxious about economic displacement. This aligns with the general awareness of AI’s diffusion and potential impacts. We elaborate on our findings below.

Who worries about unemployment?

“Just like all white-collar workers now, I am 100% worried, almost 24/7, about being replaced by AI in the end.” — a software engineer.

One-fifth of respondents expressed concern about economic displacement. Some worry abstractly: a software developer warned about “the possibility of AI being used to replace entry-level positions in its current state.” Others lament that their work, or aspects of it, are being automated. A market researcher said, “It’s undoubtedly improving my capabilities. But in the future, AI might replace my job.” In some jobs, people feel AI makes their work harder. A software developer observed, “When AI arrived, project managers started giving me increasingly difficult tickets and bugs to fix.”

Throughout the report, we use Claude-driven classifiers to infer respondents’ attributes and emotions from their answers. For example, many participants casually mention their work field or provide details about their work-life, allowing us to infer their occupation. Similarly, we quantify concerns about unemployment by prompting Claude to identify and interpret direct references where respondents indicate their roles face AI-driven displacement risks. Sample prompts are provided in the appendix.

Respondents’ perceived AI threats are correlated with our own observed exposure metrics, which reflect the percentage of tasks in a job that involve Claude. When a respondent’s exposure level is higher, their concern about AI is greater. For instance, elementary school teachers worry less about being replaced than software engineers, consistent with Claude’s bias toward coding tasks.

We illustrate this in Figure 1 below. The y-axis shows the percentage of respondents in a given occupation who believe AI has or will soon replace their role. The x-axis shows the exposure level. The graph indicates that, on average, occupations with higher exposure levels tend to express more concern about automation. Each 10 percentage point increase in exposure correlates with a 1.3 percentage point increase in perceived job threat. The top 25% of respondents with the highest exposure report this concern at three times the rate of the bottom 25%.

Figure 1: Perceived job threat from AI versus actual exposure level. The figure shows the percentage of respondents who believe AI poses a certain threat to their job, along with the actual exposure metric proposed by Massenkoff and McCrory (2026). Respondents indicating their position has been replaced or significantly reduced, or that such changes may occur soon (as coded by Claude), are classified as perceiving a job threat. The green line indicates a simple linear fit.

Another key worker characteristic is career stage. Previous research indicated early signs of hiring slowdown among recent graduates and early-career workers in the U.S. For about half of the respondents in this survey, we could infer their career stage from their answers. We find that early-career respondents are more likely than senior workers to express concerns about unemployment.

Figure 2: Concerns about economic unemployment by career stage. The percentage of respondents perceiving AI as a threat to their job, broken down by career stage. Both fields are inferred via Claude classifiers from free responses.

Who benefits from AI?

Using Claude to evaluate survey responses, we rated self-reported productivity improvements on a 1-7 scale, where 1 = “productivity decreased,” 2 = “no change,” and higher scores indicate greater gains. Responses scoring 7 include statements like, “A website I used to take 4-5 days to build now takes me a few months”; Claude assigns a 5 to “I completed what used to take four hours in half the time,” and a 2 to “Personally, I let AI help me fix code on my website, but it took multiple tries to get what I wanted.”

Overall, respondents report meaningful productivity gains. The average productivity score is 5.1, indicating “significant increase.” Of course, our respondents are active Claude.ai personal account users willing to participate, which may bias toward reporting higher productivity benefits. About 3% report negative or neutral effects, and 42% do not provide clear productivity indications.

These effects vary somewhat by income. The left panel of Figure 3 shows that high-income workers, such as software developers, report the greatest productivity gains. This result is not solely driven by coding; excluding computer and math occupations, the pattern persists. This aligns with a previous finding from our economic index, which also favors high-income workers: tasks requiring higher education levels tend to see larger reductions in time to complete when using Claude (relative to not using AI).

Some of the lowest-income workers also report high productivity gains. This includes a customer service representative who used “AI to save a lot of time creating responses based on another reply.” In some cases, low-wage workers are using AI for side tech projects. For example, a delivery driver is using Claude to start an e-commerce business, and a gardener is building a music app.

Figure 3: Inferred productivity gains by occupation. The left panel shows the average inferred productivity benefit (using Claude classifier inference) divided by median occupational wages from the U.S. Bureau of Labor Statistics (BLS), split into quartiles. The right panel shows the same results, grouped by major occupational groups. Error bars indicate 95% confidence intervals.

The right panel provides a more detailed view of productivity gains across major occupational groups. The top group is management, mostly entrepreneurs using Claude. The second-highest is computer and math, including software developers. The most modest gains are among science and legal workers. Some lawyers worry about AI’s ability to follow precise instructions: “I have given very specific rules about what, where, and how to read legal documents, and what I want it to do… but it always deviates.”

As AI spreads through the economy, a key question is where the benefits flow—workers, managers, consumers, or firms. About a quarter of respondents explicitly identified the beneficiaries during interviews. Most mentioned personal benefits—faster tasks, expanded scope, freed-up time. But 10% said employers or clients received more work, and a smaller proportion mentioned AI companies’ gains, with even fewer perceiving net negatives. This varies by career stage: only 60% of early-career workers say they personally benefit, compared to 80% of senior professionals.

Figure 4: Where do the productivity gains from AI flow? Among respondents who identified beneficiaries, the proportion attributed to each.

Scope and Speed

Respondents also shared where they experienced productivity improvements. We categorized these into scope, speed, quality, and cost. For example, many who use AI for coding say, “I’m not a technical person, but now I am a full-stack developer.” This reflects scope expansion; AI unlocks new capabilities. Conversely, some accelerate existing tasks, like an accountant who said, “I built a tool that helps me complete financial tasks that used to take 2 hours in 15 minutes.” Quality improvements often come from more thorough reviews of code, contracts, and documents. A small subset mentioned low-cost AI use: “If I hire a social media manager, it exceeds my budget.”

We find the most common productivity boost is in scope, with 48% of users explicitly mentioning this. Forty percent emphasize speed.

Figure 5: What types of productivity improvements do users report? The proportion of respondents describing each type.

User experiences with Claude may also influence their concerns about AI. To assess this, we measured reported speed improvements by extracting whether their work is now much slower (coded as 1), unchanged (4), or much faster (7).

We find a U-shaped relationship between speed improvements and perceived job threat (see Figure 6). The leftmost bars show respondents who report AI slowing them down. These respondents are more likely to see AI as a significant threat to their livelihood. For example, some creative workers, like artists and writers, find AI too oppressive and rigid to help in their work. They also worry that AI spreading into creative fields will make it harder to find jobs.

Figure 6: Work threat from AI and acceleration. The percentage of respondents who believe their job has already or may soon be replaced, based on inferred acceleration levels.

For the remaining respondents, perceived job threat increases with higher levels of inferred speed. Economically, this makes sense: if task completion times are rapidly shrinking, the future viability of that role may be more uncertain.

The economic index reveals what people do with AI. But another key input to understanding AI’s economic impact is listening directly to people’s experiences. The responses here show that people’s intuitions align with usage data: they are most worried about AI’s impact on the jobs where Claude performs the most work. We also observe higher levels of economic anxiety among early-career workers, consistent with prior research.

There are also signs that Claude empowers users. People are most likely to talk about benefits flowing to themselves rather than employers or AI companies. High-income workers are most enthusiastic about AI’s productivity effects, but low-wage and less-educated workers also report significant productivity gains. Most respondents say Claude enhances their abilities by expanding scope or speeding up tasks. However, users experiencing the greatest speed improvements are also the most anxious about AI’s impact on their jobs.

Due to the nature of the data, our analysis has important caveats. First, our survey only includes Claude.ai personal account users who chose to respond. These users may be more inclined to believe benefits flow to themselves. Second, respondents were not directly asked about many derived variables; thus, our inferences about occupation, career stage, and other variables based on context clues may be inaccurate. Relatedly, because the survey is open-ended, our measures are based on what respondents happened to mention; these findings should be confirmed with structured surveys explicitly asking about these topics.

Nonetheless, the interviews reveal genuine insights into how people perceive the AI economy, illustrating how qualitative data can surface quantitative hypotheses. Most economic concerns are themselves a strong signal.

Acknowledgments

We thank 80,508 Claude users who shared their stories.

Maxim Massenkoff led the analysis and authored the blog post. Saffron Huang led the interview program and provided guidance throughout.

Zoe Hitzig and Eva Lyubich provided key feedback and methodological guidance. Keir Bradwell and Rebecca Hiscott contributed editing support. Hanah Ho and Kim Withee contributed to design. Grace Yun, AJ Alt, and Thomas Millar implemented the Anthropic interview tool in Claude.ai. Chelsea Larsson, Jane Leibrock, and Matt Gallivan contributed to survey and experience design. Theodore Sumers contributed to data processing and clustering infrastructure. Peter McCrory, Deep Ganguli, and Jack Clark provided critical feedback, guidance, and organizational support.

Additionally, we thank Miriam Chaum, Ankur Rathi, Santi Ruiz, and David Saunders for discussions, feedback, and support.

This scale is not centered on the midpoint because most people rate productivity positively, with nearly all scores being 6 or 7 on the original Likert scale. The scale used here ranges from 1 = productivity decline, 2 = no change, 3 = slight increase, 4 = moderate increase, 5 = large increase, 6 = significant increase, to 7 = transformative increase—AI fundamentally changing what or how much they can produce.

Even excluding these “independent entrepreneurs,” management ranks alongside computer and math professions, showing the highest productivity gains.

A key limitation is that this survey targets users with Claude personal accounts. A more representative picture should include enterprise users, who may be more inclined to see value as belonging to their employer.

Anthropic surveys 80k Claude users: The people who improve efficiency the fastest with AI feel the least secure about the future.

Trending Topics

Gate13thAnniversaryLive

WCTCTradingChallengeShare8MUSDT

BitcoinBouncesBack

EthereumMemeSeasonReturns

USIranTalksProgress

Pin