Anthropic Research: The Economics of AI Based on 81,000 User Surveys

Anthropic has released findings from a survey of 81,000 Claude users, revealing a strong correlation between a worker's actual exposure to AI tasks and their fear of job displacement. While users report significant productivity gains—particularly in expanding the scope of their capabilities—those experiencing the fastest speedups in their work are paradoxically the most concerned about the future viability of their roles.

Job Displacement Concerns and AI Exposure

Perceived threat of job displacement is directly correlated with "observed exposure," which Anthropic defines as the percentage of a job's tasks for which Claude is used. For every 10-percentage-point increase in observed exposure, the perceived threat of job loss increases by 1.3 percentage points. Workers in the top 25% of exposure mention displacement concerns three times more often than those in the bottom 25%.

Career stage also significantly influences economic anxiety. Early-career respondents are much more likely to express concern about job displacement than senior professionals. This aligns with previous research indicating a slowdown in the hiring of recent graduates and early-career workers in the United States.

Productivity Gains Across Income and Occupations

Users report meaningful productivity improvements, with a mean productivity rating of 5.1 on a 1–7 scale (where 5.1 corresponds to "substantially more productive"). These gains are distributed unevenly across income levels and occupations:

• High-Wage Workers: Software developers and other high-paying roles report the largest productivity gains. This trend persists even when computer and math occupations are excluded.
• Low-Wage Workers: Some of the lowest-paid workers also report high gains, often using AI to automate repetitive responses or to launch technical side projects (e.g., a delivery driver starting an e-commerce business).
• Management and Technical Roles: Management occupations and computer/math roles show the highest inferred productivity gains. Conversely, scientific and legal professions show the mildest improvements, with some lawyers citing the model's inability to follow precise, complex instructions.

Regarding the distribution of these gains, the majority of respondents who named a beneficiary cited themselves, noting faster task completion and expanded scope. However, 10% stated that employers or clients are receiving the benefits by demanding more work.

The Impact of Scope and Speed on Worker Sentiment

Productivity gains generally manifest in two primary ways: expansion of scope and increase in speed.

• Expansion of Scope: Cited by 48% of users who mentioned productivity, this refers to AI unlocking new abilities (e.g., a non-technical user becoming a "full stack developer").
• Increase in Speed: Cited by 40% of users, this refers to completing existing tasks faster.

Anthropic observed a U-shaped relationship between speedup and job threat. Users who reported that AI slowed them down (such as some fine artists and writers who find AI too rigid) expressed high levels of concern about displacement. Simultaneously, users who experienced the most significant speedups also reported the highest levels of job threat, likely because the rapid shrinkage of time required for their tasks creates uncertainty about the role's future necessity.

Methodology and Limitations

Anthropic used Claude-powered classifiers to infer attributes such as occupation, career stage, and sentiment from open-ended survey responses. The researchers acknowledge several caveats:

1. Selection Bias: The survey was limited to users of personal accounts on Claude.ai who volunteered to respond, meaning they may be more likely to perceive benefits as flowing to themselves.
2. Inference Errors: Because occupation and career stage were inferred from contextual clues in free-form text rather than asked directly, some data points may be inaccurate.
3. Mention Bias: The findings are based on what respondents happened to mention in open-ended interviews rather than structured survey questions.