Q&A | TaskNet: Understanding AI’s impact on work by understanding the work itself

Recently introduced in an NBER working paper, the Stanford Task Network (TaskNet) is an occupational taxonomy that draws its task inventories directly from workers. The project’s interactive platform is available to explore now.

The Lab spoke with PhD student Jennifer Wang and Digital Fellow Andreas Haupt about how TaskNet can help researchers and policymakers.

What is TaskNet?

Jenn Wang: TaskNet is a database of tasks performed in different occupations. It provides task lists for each occupation and describes the characteristics of each task.

Andy Haupt: It also is a suite of tools to “find” tasks. Our just-published paper creates tasks for six occupations—software engineers, lawyers, primary school teachers, emergency medicinephysicians, dishwashers, and personal care aides—based on AI-based interviews.

Why did we need this new taxonomy for work?

Jenn: To understand how AI might change work, we first need a clear picture of what people actually do. Much of the research on AI and jobs starts with existing task lists, so what those lists include—and leave out—changes our findings and interpretation. We wanted to develop a way to learn directly from workers and keep our descriptions of work current as their responsibilities change.

Andy: We have a powerful technology, AI, and work is changing. We might reshuffle who does what, redesign occupations, and what’s done by AI. The units to do such job design are tasks. We have decent information on tasks, but we could do better by using multiple methods: for example interviewers.

What is O*NET, and how does TaskNet complement it?

Andy: The Occupational Information Network (O*NET) is a database of occupational information maintained by the US Department of Labor. Among (many) other things, it provides information on tasks.

Jenn: TaskNet complements it by using interviews to collect detailed accounts of everyday work. Across the six occupations, we identified 79 tasks without an O*NET counterpart and 58 that provide more specific descriptions of existing tasks.

Andy: Missing tasks lead to wrong predictions about jobs, for example on how exposed (meaning how much of an impact AI could have) any given occupation is.

What do interviews with workers reveal that job descriptions or existing classifications might miss?

Andy: There’s not one and only way. Different methods have different strengths: postings are abundant, but they’re written by employers, so show their perspective, and probably not the less exciting parts of work. Experts bring an expert perspective.

Jenn: Workers have firsthand knowledge of what their jobs involve, including responsibilities that may not appear in formal job descriptions. Interviews give them space to describe that work in their own words and let us ask follow-up questions to understand the details. That helps us capture tasks that people might take for granted.

What findings stood out to you?

Jenn: One finding that stands out is how much of everyday work can be invisible in official descriptions of a job. These are tasks that keep work running, like coordinating handoffs, checking the quality of someone else’s work, or helping a colleague learn. Workers may take these responsibilities for granted, even though they require time, judgment, and experience. Making them visible gives us a fuller picture of what a job involves and how AI might affect it.

Andy: Our task aggregation method works, it produces task lists that represent individual worker accounts well. The task lists we find lead to higher exposure estimates, and tasks are often “invisible” in their work. To me, the invisible part was surprising. I had no prior on the impact on exposure

Did you use AI in your research?

Jenn: AI was central to the research method. We used it to conduct interviews, ask follow-up questions, and organize workers’ accounts into task lists. Workers then reviewed their individual lists and could correct them or add missing tasks.

Andy: So much AI assisted coding!

What can the focus on tasks tell us about AI’s impact on work?

Andy: Tasks are used to both compute [AI] exposure of occupations by aggregating task level scores and understand AI use. We primarily investigate the former, and find that exposure under our new task list is higher for the occupations that we already thought (based on O*NET, the existing task list) are pretty exposed.

Jenn: We find that changing the task list changes the exposure estimate, using the same scoring approach. TaskNet produces higher estimates for lawyers, software developers, and elementary school teachers, but lower estimates for personal care aides.

This also helps clarify what exposure means. A task may offer opportunities for AI assistance while still requiring human judgment and accountability. Higher exposure doesn’t, by itself, tell us that a job will disappear.

Why does task granularity matter when estimating exposure?

Jenn: Focusing on tasks gives us a more precise picture of where AI might affect someone’s work. For example, a broad task about coordinating medical staff can include both communicating patient information and supervising residents. Our analysis assigns very different AI exposure scores to those activities. Separating them helps us identify where AI might assist with a job and where other responsibilities remain, rather than treating everything within that job as equally affected.

What do articulability, verifiability, and differences in worker performance tell us about a given task?

Jenn: These characteristics help us understand what it takes to do a task well. Articulability asks whether you can explain the standards for success in advance. Verifiability asks whether someone can assess the result afterward. Those are different questions. You might recognize good work even when it’s difficult to spell out exactly how to produce it.

The gap between average and top-performing workers helps us see where expertise makes a particularly large difference. Together, these measures let us ask more specific questions about how AI might support a task, how its output could be checked, and where human expertise matters.

Andy: Articulability and verifiability are important for what data will allow AI models to be “trained”. A paper that I recently co-authored explores this. The worker performance heterogeneity highlights also whether there is a high ceiling in terms of performance for AI.

How can researchers, policymakers, and employers use TaskNet?

Andy: Researchers [can] compute exposure scores, model AI benchmarks after the tasks, [and use the task taxonomy to classify] AI usage. Policymakers, HR practitioners, and the general public can use TaskNet to consider what parts of jobs to focus training and workforce development on.