Glonce

St. Louis Fed economists find AI adoption at work is broad but shallow

researchThe Glonce editorsThe Register🔥48Glonce rating48Impact40Novelty25Relevance85Surprise20Players60Credibility65
St. Louis Fed economists find AI adoption at work is broad but shallow

A paper titled "What Work Does Generative AI Do?" by Alexander Bick of the Federal Reserve Bank of St. Louis, Adam Blandin and Tyler R. Schumacher of Vanderbilt, and David J. Deming of Harvard measures how workers actually use generative AI on the job by analyzing Real-Time Population Survey (RPS) data, rather than relying on chat logs collected by AI vendors. Their headline finding: generative AI reaches 80 percent of occupations and more than 40 percent of tasks, but in most of those occupations fewer than half of workers have adopted it. Use is highest in management and professional occupations, particularly those tied to finance, business and computers, and lowest in personal service occupations and jobs that require manual activity or interpersonal interaction. The authors write that four out of five detailed occupations have adoption rates above 20 percent, but only one out of six occupations exceed 70 percent adoption. About 15 percent of occupations, largely computer-oriented ones, had adoption rates above 70 percent. At the task level, adoption is even thinner: only 2.8 percent of tasks had adoption rates above 50 percent, and none surpassed 70 percent. The paper contrasts these figures with occupational exposure estimates that Anthropic, Microsoft and OpenAI have each produced from their own chat logs, which the authors say come up with different results because their classifiers associate chats with a small number of generic, activity-based task descriptions such as "Edit written material or documents." As an example, OpenAI data concludes that 15 percent of chats involve that sort of editing, but the US Department of Labor's O*NET database shows only 2.4 percent of workers are in jobs that actually include that task. The researchers say this mismatch means vendor exposure figures likely overstate how relevant generative AI is to people's jobs. Separately, the paper reports that as of May 2026, 55 percent of US adults aged 18 to 64 had used generative AI for non-work reasons and 45 percent had used it for work, for an overall adoption rate of 62 percent. The authors also identify experience as a driver of adoption: people who start using generative AI in one domain tend to adopt it in other domains too. On that basis, they argue it is at least as important to understand why some workers adopt AI and others don't as it is to understand which tasks the technology can handle.

Key facts

  • A new paper using Real-Time Population Survey data finds generative AI reaches 80 percent of occupations and more than 40 percent of tasks, but adoption within most occupations stays below 50 percent of workers.
  • Four out of five detailed occupations have adoption rates above 20 percent, but only one in six exceed 70 percent; about 15 percent of occupations, mostly computer-oriented, top 70 percent.
  • At the task level, only 2.8 percent of tasks show adoption above 50 percent and none exceed 70 percent.
  • The authors say chat-log-based exposure estimates from Anthropic, Microsoft and OpenAI overstate relevance: OpenAI classifies 15 percent of chats as document editing, versus only 2.4 percent of US workers whose jobs include that task per the O*NET database.
  • As of May 2026, 62 percent of US adults aged 18 to 64 had used generative AI overall: 45 percent for work and 55 percent for non-work purposes.

Why it matters

Most public claims about how much generative AI has penetrated the workforce come from vendors reading their own chat logs. This paper uses independent labor-survey data instead, and the resulting picture is more modest and more uneven than vendor-supplied exposure estimates suggest, with real consequences for how seriously policymakers and employers should treat those vendor numbers.

Who it affects

Workers in management, professional, finance, business and computer-related occupations show the highest generative AI adoption, while personal service occupations and jobs built around manual activity or interpersonal interaction show the lowest. The findings also directly concern Anthropic, Microsoft and OpenAI, whose own chat-log-based occupational exposure estimates are the ones the paper says run too high.

How to use it

Anyone sizing generative AI's real footprint in a workforce, for planning, investment or policy purposes, should treat vendor chat-log exposure percentages with caution and weigh them against survey-based adoption data like this paper's, which tracks actual task-level use rather than the topic of a conversation with a chatbot.

How solid is it

The paper comes from named economists at the Federal Reserve Bank of St. Louis, Vanderbilt and Harvard, using the Real-Time Population Survey, a data source independent of any AI vendor. The Register article does not state the paper's publication venue, working-paper number, or the survey's sample size, methodology detail or margin of error.

Risks and caveats

The article gives no sample size, methodology detail or margin of error for the underlying survey, and no publication venue or date for the paper itself. It includes no response from Anthropic, Microsoft or OpenAI, and does not break adoption rates down by specific named occupation beyond the broad categories mentioned, nor does it explain exactly how the authors define or measure 'adoption.'

“GenAI is used in many occupations and tasks, but few of these exhibit very high adoption rates. For example, four out of five detailed occupations have adoption rates above 20 percent, but only one out of six occupations exceed 70 percent adoption.”

— the paper's authors