The paper maps four million Claude conversations to tasks in the US Department of Labor's O*NET database to measure where people are using AI at work.
2
Claude usage is concentrated in coding and content creation, while occupations involving physical work are less represented in the data.
3
The speakers question whether the paper's automation categories capture real job change, since most work consists of connected tasks that still require human judgment.
Summary
Sophia Skowronski introduces a framework from Anthropic's paper that maps four million Claude conversations to 20,000 O*NET tasks and about 1,000 occupations. The analysis estimates which tasks involve AI, associates them with skills and wages, and labels interactions as augmentation or automation. Usage is concentrated in software, coding, and content creation. Computer and mathematical occupations are overrepresented compared with their share of the workforce, while physical occupations are less visible. Adam Becker explains the privacy process behind Anthropic's Cleo system, which abstracts conversations into larger clusters and removes small clusters that may contain private information. The discussion questions whether those filters create bias and whether tasks can be cleanly separated from jobs. Valdimar Eggertsson argues that AI may change whole systems of work, as coding tools can alter a developer's workflow rather than automate one isolated task. The group also discusses adoption barriers, evaluation, responsibility, and the need for industry-specific tools.
The paper measures AI use from real Claude conversations
Sophia Skowronski presents the paper as a response to questions about whether AI will replace jobs. Instead of relying only on worker surveys, the researchers use four million Claude conversations. They map each conversation to a task in the US Department of Labor's O*NET database, then connect that task to occupations, skills, wages, and an automation or augmentation label. The result is a usage-based view of where people are already applying AI. Sophia says data from other generative AI companies could make it possible to compare usage and track changes in job demand, industries, and wages over time.
A hierarchical task taxonomy makes 20,000 O*NET tasks usable
The researchers could not send all 20,000 O*NET tasks to a model in one prompt. They first embedded the tasks, clustered them with K-means, asked Claude to name the clusters, and repeatedly refined the grouping. The result has a top level, a middle level, and a base level containing the original tasks. After cleaning duplicate tasks, the system classifies a conversation by checking whether it is occupationally relevant and then traversing the hierarchy from broad groups to a specific task. The researchers provide the classification prompts and validation details in the paper's appendix.
Claude usage is concentrated in software and content work
The analysis finds that coding and content creation make up about half of the conversations in the dataset. Software-related tasks appear at the top of the base-level and middle-level counts, including developing and maintaining software applications. Computer and mathematical occupations have more AI usage than their share of the economy, while construction, farming, fishing, forestry, and other occupations involving physical manipulation appear less often. Sophia frames this as a picture of where Claude is already being used, rather than a complete measure of where AI could eventually provide value.
The data compares AI conversations with the workforce distribution
The paper plots occupations by the percentage of tasks associated with Claude conversations and compares that with each occupation's share of the workforce. Sophia says roughly a third of occupations use AI for at least a quarter of their tasks. Adam Becker gives education and library work as an example of an overrepresented area, with about 9% of conversations compared with 5.8% of US workers. Healthcare is described as underrepresented in the conversations despite its larger workforce share. The speakers question whether concentrating data in already active occupations could reinforce those occupations while leaving others less visible.
Cleo abstracts conversations before analysts see them
Adam Becker describes Anthropic's Cleo system, which stands for Claude in Insights and Observations. A conversation is first turned into a facet containing a summary and metadata, such as the language and number of turns. Related facets are grouped into broader clusters, using examples such as tying shoes and tying bows becoming a cluster about tying knots. Analysts see only sufficiently abstracted clusters. Small clusters are removed because they may contain information from a single person. Adam says this process can reduce exposure to private information, but he also asks whether removing small clusters introduces bias into the analysis.
Automation and augmentation depend on disputed labels
The classification process first asks whether a conversation concerns an occupational task. It then identifies a collaboration pattern such as directive interaction, feedback loops, task iteration, learning, or validation, and groups those patterns into augmentation or automation. Adam is unconvinced that fixing a pasted software error counts as automating a job, because the developer still supplies business logic and interprets the result. Valdimar Eggertsson also says a single conversation can include several interaction patterns. In his view, the five-category scheme can be too narrow for the way people actually work with Claude.
Valdimar argues that the study measures tasks rather than jobs, and that a job is a connected system rather than a simple list of independent activities. A task can be automated while other parts of the same job remain dependent on human knowledge and responsibility. He says the current data reflects what people use Claude for now, so it may understate future use. Coding tools such as Cursor illustrate his point: the value can come from changing the whole development workflow, rather than saving a small amount of time on one task. He compares this with factories that became more efficient when electricity changed their layout instead of merely replacing a steam engine.
Adoption depends on trust, evaluation, and industry context
The discussion turns to farmers and other workers who have not adopted AI tools. A participant says smaller farmers may distrust new technology after previous investments in sensors and systems failed, while another describes the need for an AI tool that explains how to use the AI tool. The group suggests presenting a better way to solve a familiar problem rather than leading with the term AI. They also discuss responsibility when automated systems make mistakes. Software is easier to evaluate because code can be run, but other domains may need human validation, industry-specific benchmarks, and external evaluators who understand the work.
"They go and they classify them into distinct tasks that they found from O*NET and each of these tasks is then associated with one or multiple occupations."Adam Becker15:55
Who should watch
You are trying to understand where people are already using generative AI at work, rather than relying on predictions about future job loss.
You are building an AI product for a non-technical industry and need to think about trust, adoption, evaluation, and the language used with customers.
You want to examine the limits of task-level automation studies, especially their treatment of privacy, occupation labels, and human responsibility.
