Anthropic Releases AI Economic Index Report: How AI is Reshaping the Labor Market

Artificial Intelligence (AI) systems will profoundly change the way people work in the coming years. To gain a deeper understanding of the long-term impact of AI on the labor market and economy, Anthropic has launched the "Anthropic Economic Index" (Anthropic Economic Index) research project.

The first report of the index, based on ClaudeData from millions of anonymized conversations on the .ai platform provide cutting-edge data insights and analysis. The report reveals the clearest picture yet of how AI fits into the tasks of the modern economy. Observers note that the volume of data and depth of analysis in the report is unprecedented, and Anthropic appears intent on establishing itself as a leader in AI economic impact research.

Anthropic has further publicized the dataset that underpins this analysis so that researchers can build on it. Commentators have suggested that this move demonstrates Anthropic's openness to encouraging wider academic participation in research on the economic impact of AI. The upcoming labor market transition and its potential impact on employment and productivity will require a multidimensional policy response. To this end, Anthropic invites economists, policy experts, and researchers from other fields to contribute their insights to the subsequent research of the Index.

Key findings of the first report of the Economic Index include:

• Industry concentration of AI applications: Currently, the use of AI is concentrated in software development and technical writing. More than one-third of occupations (about 36%) use AI for at least one-quarter of their relevant tasks, and about 4% of occupations apply AI in depth to three-quarters of their tasks.Commentators say these data provide initial confirmation of AI's penetration in specific knowledge-based jobs, while also signaling a broader landscape of AI applications that has yet to be explored.
• Enhanced applications dominate: Compared with the automation mode (43%) in which AI directly performs tasks, the application of AI focuses more on the augmentation mode (57%), that is, AI and humans work together to augment human capabilities in verification, learning and task iteration. Analysts interpreted that this shows that the current AI technology is better at assisting humans rather than completely replacing human work, and human-machine collaboration is likely to become the mainstream work mode in the coming period.
• AI Applications and Salary Levels: AI is more commonly used in tasks associated with middle- and high-wage occupations, such as computer programmers and data scientists. However, AI usage is relatively low in the lowest and highest paying jobs. Experts speculated that this may reflect the limitations of current AI capabilities and the practical barriers to adoption of AI technologies in different industries. It was further suggested that AI technologies may be more suitable for tasks that have a certain level of complexity but do not overemphasize creativity or interpersonal skills.

To take a deeper look at Anthropic's initial findings, the next sections detail how AI is being used in the labor market.

How AI is used and applied in various economic sectors, with data from Claude.ai's real user data. The numbers in the graph represent the percentage of Claude-related conversations that are related to specific tasks, occupations, and categories.

Insights into the AI adoption landscape in the labor market

Anthropic's latest research paper is based on a long-term study of the relationship between technological change and the labor market. Technological advances have consistently reshaped the labor landscape, from the jenny spinning machine of the Industrial Revolution to today's automotive manufacturing robots, and Anthropic focuses on the transformative impact AI is having. Unlike traditional questionnaires or future projections, Anthropic has direct data on how AI is actually being used. Critics say that this approach avoids subjectivity bias and draws directly from user behavioral data, making the findings more objective and compelling.

Occupational task analysis methodology

Anthropic's approach stems from an important insight in the economics literature: it is sometimes more analytically valuable to focus on "occupational tasks" than on the occupations themselves. Different jobs often contain certain tasks and skills in common. For example, the ability to recognize visual patterns is a task that designers, photographers, security screeners, and radiologists are all required to perform. Some scholars point out that this task-centered analytical perspective captures the structural impact of technological change on the labor market in a more granular way.

There are differences in the potential for different types of tasks to be automated or enhanced by new technologies. As a result, Anthropic anticipates that AI will be selectively applied to specific tasks in different occupations. By analyzing tasks, rather than just jobs as a whole, Anthropic is able to gain a more comprehensive understanding of how AI will be gradually integrated into the economy.

The Clio System: Connecting AI Apps to Career Missions

The study benefited from Anthropic's development of Clio (Claude Insight & Observation), an automated analytics tool that analyzes data on users' conversations with Claude while protecting their privacy.Anthropic utilized the Clio system to analyze approximately one million conversations (specifically, both Free and Pro versions of conversations on the Claude.ai platform) between users and Claude. Anthropic used the Clio system to analyze approximately one million user conversations with Claude (specifically, both the Free and Pro versions of conversations on the Claude.ai platform), which Clio was able to efficiently organize and categorize based on professional tasks. The reviewer praised the use of the Clio system as a highlight of the study, as it was able to extract valuable economic information from the vast amount of conversation data, while maintaining strict user privacy protections.

Anthropic's research team, using the Occupational Classification System (OCS) developed by the U.S. Department of Labor as a baseline and referring to the Occupational Information Network (OIN) maintained by the Department of Labor ONET (Occupational Information Network) database. oThe Clio system compares each conversation with the ONET task entries are matched (the process is shown below). Anthropic then follows the ONET's classification framework, which further categorizes tasks into occupations that represent those tasks, and groups occupations into broader categories such as "education and libraries," "business and finance," and so on.

Anthropic's Clio system converted conversations with Claude (data kept strictly confidential; top left) into occupational tasks (top center) and further into occupations/occupational categories derived from O*NET (top right). These data were then used for a variety of analyses (bottom row; discussed in more detail below).

Findings

Distribution of AI applications by job type. In Anthropic's analysis of the dataset, the tasks and occupations with the highest AI adoption rates were concentrated in the "Computers and Math" category, which primarily covers software engineering-related positions. 37.2% of user queries sent to Claude fell into this category, covering tasks such as software modification, code debugging and network troubleshooting. The review notes that it's no surprise that the software engineering field is at the forefront of AI adoption - after all, AI technology itself has its roots in computer science.

The second largest category was "Art, Design, Sports, Entertainment & Media" (10.3% of user queries), which largely reflects users utilizing Claude for all types of writing and editing. Analysts see this as a sign that AI is beginning to make its presence felt in the creative industries as well, especially in the content creation space. As expected, occupations that involve a lot of manual labor, such as the "Agriculture, Fishing, and Forestry" category (which accounted for only 0.11 TP3T of user queries), had the lowest share of AI adoption.

Anthropic also compared the percentage of AI adoption for each type of occupation in its own dataset with the share of those occupations in the overall labor market. The results are shown in the chart below.

For each job type, the orange bars represent the percentage of conversations related to Claude, and the gray bars represent the percentage of workers in the U.S. economy with that type of job (data from the U.S. Department of Labor's O*NET Occupational Classification).

Depth of AI adoption within careers. Anthropic's analysis reveals that only a very small number of occupations use AI in depth for most of the tasks to which they are relevant. the data shows that only about 4% of jobs have applied AI to at least 75% of their tasks. however, more modest applications of AI are much more prevalent: about 36% of jobs have begun to use AI for at least 25% of their tasks. commentators argue that this suggests that AI adoption is still in its early stages and that deep applications are yet to become widespread, but that shallow applications are beginning to penetrate. According to commentators, this suggests that AI adoption is still in its early stages, with deep applications yet to become widespread, but shallow applications beginning to penetrate a wide range of industries.

As Anthropic expected, the dataset does not show signs of jobs being completely automated by AI. Instead, AI is gradually penetrating many task segments of economic activity, only affecting some task groups more than others. According to the analysts, this is further evidence that current AI technologies favor task-level enhancements over job-level substitutions.

The relationship between AI adoption and salary levels. The O*NET database provides data on the median annual salary for each occupation in the U.S. Anthropic incorporates this information into its analytical framework, allowing it to compare the median salary level of each industry with the extent to which AI is being used in its related tasks.

One interesting finding is that AI usage is relatively low for both low- and very high-income jobs (which typically require a high degree of manual skills, such as hairdressers and obstetricians). In contrast, specific occupations that are in the middle to high income levels, such as computer programmers and copywriters, show a high reliance on AI in Anthropic's data. Observers have commented that this phenomenon is instructive, suggesting that AI may be more likely to find application in middle-skill, middle-income knowledge-based jobs, while being relatively less substitutable for jobs that require a high degree of manual skill or extreme expertise.

The graph shows the relationship between the median annual salary for an occupation (x-axis) and the percentage of Claude conversations involving that occupation (y-axis), and highlights some representative occupations.

Automation and Enhancement Segmentation. Anthropic also takes a deeper look at how tasks are "performed" - specifically, which tasks tend to be "automated" (the AI performs the task directly, such as document formatting), and which tasks are "augmented" (where the AI collaborates with the user to accomplish the task).

Overall, the results of the study slightly favored the "enhancement" model. Of all the tasks analyzed, 57% belong to augmentation applications and 43% to automation applications. This means that in slightly more than half of the application scenarios, the role of AI is not to replace humans in performing tasks, but rather to "synergize" with humans by engaging in validation (e.g., reviewing the user's work product), learning (e.g., assisting the user in acquiring new knowledge and skills), and task iteration (e.g., assisting the user in brainstorming or performing (e.g., assisting users in brainstorming or performing repetitive generative tasks). Expert analysis shows that augmented applications are slightly more common than automated applications, suggesting that AI tools like Claude are currently more focused on improving human productivity and creativity than on directly replacing the human workforce.

The figure illustrates the percentage of conversations with Claude that are augmented versus automated applications, and the composition of the breakdown of task subtypes within each category. The specific definitions of the subtypes are explained in Anthropic's research paper. "Instruction": task delegation with minimal human interaction; "Feedback loops": task completion guided by environmental feedback; "Task iteration": optimizing the task execution through a collaborative process; "learning": knowledge acquisition and understanding; "validation": verification and improvement of work results.

Limitations of the study

This study by Anthropic provides a unique perspective on understanding how AI is reshaping the labor market. However, as with all research, the study has some important limitations. The main ones include:

• Definition of work scenarios: Anthropic is unable to accurately determine whether a user's use of Claude to perform a task is in the realm of work. Users who turn to Claude for writing or editing advice may indeed be working on a task, but they may also be doing so simply to fulfill a personal hobby (e.g., writing a novel). Commentators have noted that this is an inherent methodological challenge, and that it is difficult for any study based on user dialog data to fully distinguish work scenarios from non-work scenarios.
• Interpretation of user behavior: Linked to the above issue, Anthropic does not understand how users actually use the results of Claude's responses. For example, do users copy and paste code snippets directly? Do they fact-check the content of replies, or do they uncritically accept them wholesale? In Anthropic's data, certain application scenarios that appear to be "automated" may actually still be in "augmented" mode. For example, a user may ask Claude to write a complete memo on his behalf (which is ostensibly automated), but then make edits and refinements on his own (which would be an enhancement). Analysts believe that this "automated shell, human core" phenomenon may lead to an overestimation of the degree of automation.
• Limitations of data sources: Anthropic currently only analyzes user data from the Claude.ai Free and Pro plans, and does not cover data from API, Team, or Enterprise users. While Claude.ai's data may be mixed with conversations from non-work scenarios, Anthropic has filtered the data using a language model that seeks to retain only conversations related to occupational tasks, which mitigates the data bias issue to some extent. While the data filtering attempts have mitigated the bias to some extent, relying only on Claude.ai's free and Pro version user data may make it difficult to fully reflect the true picture of enterprise-level AI applications, experts say.
• Errors in task categorization: Due to the large absolute number of task types, the Clio system may have incorrectly labeled some of the conversations during the categorization process (for more information on how Anthropic validated the results of the analysis, see the full paper, especially Appendix B).
• Limitations of modeling capabilities: Claude does not currently have the ability to generate images (except indirectly through code), so some creative image-related applications were not addressed in the study data.
• Over-emphasis of coding use cases: Given that Claude has been touted as one of the most advanced coding models, Anthropic expects that coding-related use cases may be overrepresented in the data. As such, Anthropic does not believe that the distribution of applications in the current dataset is fully representative of the overall AI adoption landscape. The reviewer emphasizes that the research team also candidly acknowledges the limitations of the dataset, particularly given Claude's market position as a programming assistant, which may have resulted in a bias in the data results towards areas such as software development.

Conclusions and directions for future research

The applications of AI technology are expanding rapidly and the capabilities of AI models continue to grow. The face of the labor market may change significantly in the near future. As a result, Anthropic plans to repeat many of these analyses over time in order to continue to track potential social and economic changes, and Anthropic will periodically publish the results and related datasets as part of the Anthropic Economic Index, a research project that Anthropic seems intent on perpetuating to continue to track and analyze the far-reaching impact of AI on the economy and society. Anthropic seems intent on perpetuating and normalizing this research project to continue to track and analyze the far-reaching economic and social impacts of AI.

These types of longitudinal analyses will help Anthropic gain deeper insight into the complex relationship between AI and the job market. For example, Anthropic will be able to dynamically monitor trends in the depth of AI adoption within occupations. If AI continues to be applied primarily to specific tasks in the future, and only a few jobs apply AI in depth to the vast majority of tasks, then the future labor market may favor the iterative evolution of existing occupations over mass extinction.Anthropic can also continually monitor changes in the ratio of automation to augmentation applications, thus capturing timely signals of areas where automation applications are becoming more prevalent. Analysts believe that tracking changes in these key metrics over time will help determine more accurately whether AI's impact on the labor market will be a "disruptive impact" or an "incremental evolution."

Anthropic's research provides valuable data on how AI is actually being used, but does not directly offer policy recommendations. The answer to the question of how to adequately prepare for the impact of AI on the labor market cannot come from isolated research findings alone. Rather, it requires a combination of evidence, social values, and practical experience from a variety of fields, and Anthropic looks forward to utilizing its new research methodology to contribute to a clearer articulation of these critical issues. The ultimate goal of Anthropic is not just academic research, but also to provide insights for future policy making and social response strategies.

For more detailed information on Anthropic's analysis and findings, please read Full PaperThe