Estimating time spent on work tasks
This paper introduces a principled method for estimating time shares for nearly 18,000 U.S. work tasks by combining ONET frequency data with language model-derived pairwise comparisons of task duration, revealing that weighting task exposure to AI by time rather than task count significantly alters the ranking of affected occupations and widens the gap between the most and least exposed jobs.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the economy as a giant, bustling kitchen where every job is a different recipe. For decades, economists have tried to understand how new tools—like a fancy new blender or a robot arm—change these recipes. They realized that jobs aren't just single, unbreakable blocks; they are actually bundles of smaller tasks, like chopping onions, stirring the sauce, or plating the dish. This "task-based" view is the standard way scientists study how technology reshapes work. If a robot can chop onions faster, that specific task changes, and that change ripples up to affect the whole job. But here's the tricky part: when we add up all these tiny tasks to see how a whole job is affected, how do we count them? Do we count every task as one equal vote, or do we weigh them by how much time a worker actually spends on them? It turns out, this question of "how much time" is the missing ingredient that could completely change our understanding of which jobs are most likely to be transformed by Artificial Intelligence.
This paper, written by researchers from Stanford and the University of São Paulo, is like a massive, meticulous time-tracker for the American workforce. The authors noticed that previous studies trying to figure out which jobs are most exposed to AI had been guessing at how much time workers spend on specific tasks. Some studies just counted how many tasks a job had, while others used rough categories like "core" or "supplemental" tasks, treating a quick 30-second check-in the same as a 4-hour deep-dive analysis. The researchers argue this is like judging a movie by counting the number of scenes rather than how long the audience spends watching the plot unfold.
To fix this, the team built a new method to estimate the actual time share for nearly 18,000 different tasks across almost 1,000 U.S. occupations. They didn't just ask a computer to guess; they broke the problem down into two parts. First, they looked at how often a task happens (frequency). Second, and this is the clever part, they used a language model (an AI) to rank tasks against each other to figure out which ones take longer to do just once. They then fed these rankings into a mathematical puzzle (a linear program) that had to fit everything into a realistic 7-hour workday. It's like solving a jigsaw puzzle where the pieces are tasks, and the picture is a full day of work.
The results of this time-tracking experiment are surprising and shift the landscape of the conversation. When the researchers applied their new time-based weights to see how exposed jobs are to AI, the list of "most at risk" jobs changed dramatically. Previous studies suggested that clerical and administrative jobs were the most vulnerable because they had the highest number of AI-exposed tasks. However, the new time-weighted view suggests that while these jobs have many small, exposed tasks, those tasks don't actually take up much of the worker's day. In contrast, jobs that involve deep analysis, research, and writing—which might have fewer total tasks—spend a huge chunk of the worker's time on those specific tasks.
When the authors re-ranked the top 25 most exposed occupations using their time data, 11 of them swapped places. The list moved away from clerical roles and toward analytical ones. For example, they found that for a Massage Therapist, while 44% of the listed tasks might be technically "exposed" to AI, those tasks only make up 16% of the actual time spent working. Conversely, for an Actuary, the exposed tasks might be fewer in number but consume nearly 90% of their working time. The paper suggests that by focusing on the fraction of working time exposed to AI, rather than just the fraction of tasks, we get a much clearer, and often more intense, picture of the impact on the most vulnerable roles. The authors are careful to note that this is a new way of measuring exposure, not a prediction that these jobs will definitely disappear, but it does suggest that the "headline" jobs most affected by AI might be different than we previously thought.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.