Author: Anthropic

Translation: Deep Tide TechFlow

Deep Tide Guide: This is the first large-scale survey of users' real economic anxieties conducted by an AI company. The data reveals a cruel paradox: Those who are best at using AI, such as programmers and designers, are the most worried about being replaced by AI; the people who see the quickest efficiency gains are the least secure about the future. For investors, this means that the penetration of AI is happening faster than expected, and the impact on the job market has already begun at a psychological level.

Key Findings:

Our recent survey of 81,000 Claude users shows that people in jobs that are more easily replaced by AI are more worried about unemployment caused by AI. Respondents in the early stages of their careers are particularly concerned.

Occupations with the highest and lowest incomes report the largest productivity gains, primarily due to an expansion of the scope of work (taking on new tasks).

Those who experience the largest speed improvements due to AI express even greater worries about unemployment.

To help the public understand the AI economic changes we observe, our economic index shares what tasks Claude is being asked to perform, and the tasks in which Claude completes the largest proportion of the work. However, so far we lack information on how these usage patterns map to people's thoughts and perceptions of AI.

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

The survey asked people about their visions and fears regarding AI advancements. Many shared thoughts related to economic topics. We learned that many people are worried about unemployment—despite feeling more productive and capable. In some cases, AI has empowered them to start businesses or freed up time to do more important things; in other cases, AI feels oppressive or is imposed on them by employers.

The survey results provide preliminary evidence suggesting a relationship between the observed exposure (an indicator we use to measure AI replacement risk) and economic concerns surrounding AI. People in high exposure occupations—defined by the tasks tracked by Claude—are more anxious about economic displacement. This is consistent with the general awareness of the spread of AI and its potential impacts. We elaborate on our findings below.

Who is worried about unemployment?

"Like all white-collar workers right now, I am 100% worried, almost 24/7 worried that I will eventually be replaced by AI."—a software engineer.

One-fifth of respondents in our survey expressed concerns about economic displacement. Some worry about this issue in the abstract: a software developer warned, "AI is currently being used to replace entry-level positions." Others lament that their job, or certain aspects of it, is being automated. A market researcher stated, "Undoubtedly, my capabilities have increased. But in the future, AI may replace my job." In some roles, people feel that AI has made their jobs more difficult. A software developer observed, "When AI came along, project managers started giving me harder tickets and bugs to solve."

Throughout the report, we used a Claude-driven classifier to infer respondents' attributes and emotions from their answers. For example, many participants mentioned their field of work or provided details about their working life, allowing us to infer their occupations. Similarly, we quantified concerns about unemployment by prompting Claude to identify and interpret direct quotes where respondents indicated their roles are at risk of AI-driven replacement. Example prompts are provided in the appendix.

Respondents' perceived AI threat is related to our own observed exposure indicators, which reflect the percentage of tasks in a job performed using Claude. When the respondents' observed exposure indicators are higher, their worries about AI are greater. For example, primary school teachers are less concerned about being replaced than software engineers, which aligns with the fact that Claude usage leans towards coding tasks.

We illustrate this in Figure 1 below. The y-axis represents the percentage of respondents in a given occupation indicating that AI has already replaced or might soon replace their role. The x-axis is the observed exposure. The chart shows that, on average, individuals in higher exposure occupations tend to express more concerns about job automation. For every 10 percentage points increase in exposure, perceived job threat rises by 1.3 percentage points. Those in the highest 25% of exposure mention this concern three times more than those in the lowest 25%.

Figure 1: Perceived job threat from AI versus actual exposure level. The chart shows the percentage of respondents who believe AI constitutes some job threat, along with the actual exposure level metric proposed byMassenkoff and McCrory (2026). If a respondent states that their position has been replaced or drastically reduced, or that such changes might happen soon (using Claude coding), the respondent is coded as perceiving a job threat. The green line represents a simple linear fit.

Another important worker characteristic is career stage. In previous research, we reported preliminary signs of a slowdown in hiring among recent graduates and early-career workers in the U.S. For about half of the respondents in this survey, we were able to infer career stages from their answers. We found that early-career respondents are more likely to express concerns about unemployment compared to senior employees.

Figure 2: Concerns about economic unemployment by career stage. The percentage of respondents indicating AI poses some threat to their work, categorized by career stage. These two fields are inferred from open-ended responses using Claude classifier.

Who benefits from AI?

Using Claude to evaluate survey responses, we assessed the degree of self-reported productivity increase on a 1-7 scale, where 1 is "productivity decreased," 2 is "no change," and subsequent levels indicate greater improvements. Responses scoring 7 included testimonies like, "The website I created in 4-5 days used to take months"; Claude rated statements like "What used to take four hours is now done in half the time" with a score of 5, and "For me personally, I had AI help fix the code on my website. But I only got what I wanted after multiple attempts" received a score of 2.

Overall, respondents reported a significant average productivity gain. The average productivity score was 5.1, corresponding to "productivity greatly increased." Of course, our respondents are active users of Claude.ai willing to participate in the survey. This may make them more likely to report productivity gains compared to the average user. About 3% reported negative or neutral effects, and 42% did not provide a clear productivity indicator.

This varies to some degree by income. The left panel of Figure 3 shows that those in high-paying jobs, such as software developers, report the greatest productivity gains from AI. This finding holds not only for coding-driven work; it continues to apply even when we exclude computer and math professions. This echoes a previous finding from the economic index, which also favors high-wage workers: in tasks requiring higher education levels, Claude tends to reduce the time required to complete tasks (relative to not using AI) by a higher percentage.

Some of the lowest-paid workers also described significant productivity gains. This includes a customer service representative using "AI to create responses for me based on another reply, saving me a lot of time." In some cases, low-paid workers use AI for side gigs. For instance, a delivery driver is using Claude to start an e-commerce business and a gardener is building a music application.

Figure 3: Inferred productivity increases by occupation. The left chart shows the average inferred productivity gains from AI (using Claude classifier) divided by income quartiles based on median occupational wages provided by the Bureau of Labor Statistics (BLS). The right chart shows the same results, but categorized by major occupational groups. Error bars indicate 95% confidence intervals.

We take a closer look at this in the right panel of Figure 3, displaying inferred productivity gains by major occupational groups. At the top are management occupations. Most of these respondents are entrepreneurs using Claude to start businesses. The second highest category is computer and math, including software developers. The groups showing the most moderate productivity improvements are science and legal professions. Some lawyers are concerned about AI's ability to follow precise instructions. For example: "I've given very specific rules about what where, how to read legal documents, what I want it to do... but it deviates every time."

As AI spreads throughout the economy, a key question is where the gains will flow—to workers, their managers, consumers, or companies. About one quarter of respondents in interviews specified who the recipients of these gains are. Overall, most of these individuals mentioned benefits for themselves through faster tasks, expanded scope, and liberated time. But 10% of those mentioning recipients indicated that employers or clients demanded and received more work. A smaller proportion mentioned benefits for AI companies, and an even smaller fraction said AI will have a net negative impact. This varies by career stage: only 60% of early-career workers indicated they personally benefit from AI, while that figure rises to 80% for senior professionals.

Figure 4: Where do the surplus gains from AI productivity improvements go? Among respondents who identified beneficiaries of AI productivity gains, the percentage each type of beneficiary accounts for.

Scope and Speed

Respondents also shared where they experienced productivity increases. We categorized them into scope, speed, quality, and cost. For instance, many who use AI for coding tasks said, "I am not a technical person, but now I am a full-stack developer." This is an expansion of scope; AI has unlocked new capabilities for them. In contrast, some users accelerated tasks they were already doing, like the accountant who said, "I built a tool that helped me complete a financial task that used to take 2 hours in 15 minutes." Quality improvements typically come from more thorough checks of code, contracts, and other paperwork. A small number of respondents mentioned the low cost of using AI: "If I hire a social media manager, it exceeds my budget."

We found that the most common productivity improvement is in scope, with 48% of users who explicitly mentioned productivity impacts reporting this. 40% of those mentioning productivity emphasized speed.

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

People's experiences using Claude may also impact their worries about AI. To assess this, we measured the speed improvements reported by respondents by extracting whether their work is much slower (coded as 1), has no change (4), or is much faster (7).

We found that the relationship between speed improvements and perceived job threat is U-shaped (see Figure 6). The leftmost bars indicate respondents who report AI slowing down their speed. These respondents are more likely to indicate that AI poses a significant threat to their livelihoods. For instance, some creative workers, such as artists and writers, found AI too oppressive and rigid to assist them in their work. At the same time, they worry that AI's spread into creative fields will make it harder for them to find jobs.

Figure 6: Job threats posed by AI and acceleration. The percentage of respondents indicating their position has already been or might soon be replaced, based on inferred acceleration.

For the rest of the respondents, the perceived threat to their job increases continuously with the level of speed improvements suggested by their answers. This makes economic sense: if the time needed to complete tasks is shortening rapidly, there may be more uncertainty regarding the future viability of that role.

The economic index reveals what people are doing with AI. But another key input to understanding AI's economic impact is directly listening to people's experiences. The responses explored here show that people's intuitions are consistent with usage data: they are most worried about AI's impact on the work we observe Claude doing most. We also found that early-career workers exhibit higher levels of economic anxiety, aligning with previous research.

There are also signs that Claude is empowering users. People are most likely to talk about benefits flowing to themselves rather than to employers or AI companies. High-wage workers are most enthusiastic about the productivity impact of AI, but individuals in low-wage jobs and those with lower education levels are also reporting substantial productivity gains. Most respondents report that Claude has enhanced their capabilities in the form of expanded scope or accelerated speed. However, those experiencing the greatest speed improvements also show the most anxiety about AI's impact on their jobs.

Given the nature of the data, there are important considerations for our analysis. First, our survey was limited to voluntary responses from Claude.ai personal account users. Among other potential biases, these users may be more inclined to believe that the benefits flow to themselves. Secondly, users were not directly asked about many of the derived variables here, so our inferences about occupations, career stages, and other variables from contextual clues may be incorrect. Relevantly, since the survey was open-ended, our measures rely on topics respondents happened to mention; these findings should be corroborated in structured surveys directly asking about these subjects.

Nonetheless, the interviews reveal real insights about people's feelings regarding the AI economy, showing how qualitative data can emerge to support quantitative hypotheses. Much of the economic-related anxiety itself is a strong signal.

Acknowledgments

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

Maxim Massenkoff led the analysis and wrote the blog post. Saffron Huang led the interview project and provided guidance throughout.

Zoe Hitzig and Eva Lyubich provided key feedback and methodological guidance. Keir Bradwell and Rebecca Hiscott provided editorial 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 key 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 a midpoint as most people's assessments of productivity are positive, with almost all being scores of 6 and 7 on the original Likert scale. The scale we use here ranges from 1 = productivity decreases, 2 = no change, 3 = slight improvement, 4 = moderate improvement, 5 = significant improvement, 6 = remarkable improvement, to 7 = transformational improvement—AI fundamentally changes what or how much they can produce.

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

But an important limitation is that this survey targets users with Claude personal accounts. A more representative picture should also include corporate users, who may be more inclined to view value as belonging to employers.

Related Content

Launch of the Anthropic Economic Index Survey

We are launching the Anthropic Economic Index Survey, a monthly survey conducted through the Anthropic Interviewer.

Automation Alignment Researcher: Expanding Scalable Supervision with Large Language Models

Can Claude autonomously develop, test, and analyze alignment ideas? We conducted an experiment to find out.

Trustworthy Agents in Practice

AI "Agents" represent a significant shift in the ways people and organizations use AI. Here, we explain how they work and how we ensure their trustworthiness.