In the future, will robots take all our jobs?
Daniel Rock thinks that’s the wrong question to ask.
Instead, Rock asks how, and how much, AI technology will impact the many tasks that comprise a given job — each one may be affected to different extents, changing a job in more nuanced ways than completely automating it.
“What I’m focused on is thinking about jobs in new ways,” says Rock.
Daniel Rock is a 2023 AI2050 Early Career Fellow and Assistant Professor of Operations, Information, and Decisions at the Wharton School of the University of Pennsylvania. His research has been cited by the Federal Reserve and Brookings Institution, and featured in outlets including The New York Times, Wall Street Journal, Bloomberg, Harvard Business Review, and Sloan Management Review.
Rock researches the economic effects of digital technologies, focusing on the economics of AI and machine learning in economic research. His AI2050 project quantifies the impact of AI on work. He measures how different jobs might change and be impacted by AI and aims to develop an open-source model to enable further research in the field. This work addresses Hard Problem #5, which concerns the future of work, as well as the economic challenges and opportunities resulting from AI technology.
The AI2050 initiative gratefully acknowledges Fayth Tan for assistance in producing this community perspective.
Your research examines the impact of AI on work. How would you define that impact?
When we think about work, you need a unit of analysis. If people default to jobs, and say, “All the jobs are going away,” that’s not necessarily helpful. You have to go within jobs, to tasks or skills. A task is ultimately the atomic unit of work, but you can have subtasks for whatever definition of task you pick. As robotics people will tell you, it’s not as easy as saying, “Pick up that apple.” It’s saying, “Rotate an arm along these 16 dimensions.” I’m not totally sure how many axes the latest robotic arms have.
I think you can also look at systems in a firm — for example, this module is a mixture of an engineer and a product manager and a designer, and they create this kind of output. There’s choreography that happens within companies to get their work done. The way that jobs are networked together, and work is networked together is interesting. Depending on what the question is, I try to find a reasonable unit of analysis to study how AI might change it, and affect people’s work processes.
This is a subject that's been in the popular discourse a lot—what motivated you to study it?
I started studying this question in some shape or form around 2013 when I started my PhD program — one of the AI2050 Senior Fellows, Erik Brynjolfsson, was my advisor. The two of us were listening to all these computer scientists working on really interesting problems, saying, “We’re going to have self-driving cars in five years”, but another researcher said it’d be 40 years before that happened.
I thought, well, if all these brilliant people disagree about when these technologies are going to start impacting society in a meaningful economic way, that seems like a good place for me to do some work.
We started thinking about what problems are solved by machine learning algorithms well, and developed an approach where you score a task and think about where it might change. That tells you the limits of what we could do — it’s technological feasibility, not what’s economically or socially viable. There’s a whole bunch of different reasons why you might not actually deploy AI into something, but this gives you a sense of what the technology can do and what might change.
Are there precedents for technological disruption before AI that might be a useful parallel to help people understand how AI might impact work?
You’ve got the Industrial Revolution, both steam power and electric power. There’s this great paper by Paul David about the dynamo and the computer. He talks about how factories were organized. Once electric power came out, they tried to organize factories in the same way as they did with steam — a big generator in the middle of the room and then offshoots into little machines. But the real gains came 30 years later, when many little engines were placed all over the room, as opposed to one big one.
That kind of reconfiguration needs to happen with any big technology. A lot of my work has been figuring out if AI is what we could consider a general purpose technology — in which case, we should expect this to take a long time. We can accelerate it with good choices, but it’s not the sort of thing that happens overnight. Somewhere along the line, we’ll figure out [what] will work well, and [what] won’t.
What does measuring impact entail? When do you think we’ll start to see some of these impacts?
From a forecasting perspective, I think this takes decades. When a lot of tasks are changing within an area, we can probably infer that something important is going on in terms of how people configure that work, and what kinds of machines and investments are necessary there.
One of my three projects is using large language models to build a new measurement system for how quickly certain types of roles are changing, and demonstrating its applications in macro statistics, but also applied micro studies of workers. We want to understand, for example, the effects of a minimum wage law change without controlling for aspects of the work, such as the specific type of job or industry.
Another project imports methods from finance, in the way we think about efficient portfolios. The optimal thing under certain models is to hold a little bit of every stock and diversify an index fund. Workers can’t do that with skills. You’re constrained — you can either invest more in one skill, or you can invest more of your budget into a wider range of skills.
The third project is creating encoder models, which are LLMs that capture the meaning and context of words without generating new text, for job postings, and then open-sourcing them so that other folks can work with them. I definitely couldn’t do all this work on my own, and I won’t come up with all the ideas either. It’d be great to expose some of these models to the rest of the community, and let brilliant people have an easy time working with language models.
Are there certain types of jobs that you think might be more affected or changed by AI?
It’s higher wage knowledge work that tends to be affected more or has more potential to change — lawyers, programmers, salespeople, marketers. These people tend to earn more than average workers do, and have more tasks that can be changed by AI.
We looked at automation in some work with OpenAI and GovAI colleagues that recently came out in Science. We looked through different types of roles that might be affected for overall exposure, and [an] experimental automation rubric, where generative AI [would be used] to automate stuff explicitly. The stuff that had higher exposure to automation tended to be clerical work and routine information processing. It looked a lot like the last wave of ordinary software-based automation, just using a new tool.
If you could rebut any popular narrative that irks you about AI, what would it be?
It’s the bouncing between extremes — either “robots are going to take away all your jobs”, or “AI technology is meaningless.” It’s going to help a lot in certain areas, and it’s not going to change too much elsewhere. I think the media tends to cherry-pick to build a story that is either scary or alarming. Don’t get me wrong, I think there’s things that we need to protect against. The boring story here is we have a lot of hard work to do as a society to integrate this stuff well.
More narrowly, it’s “AI will automate all the jobs”, or in particular, when they cite my work to say all the jobs are going to be automated. If I had hair, I’d be pulling it out! If people did precisely what they were told in their job, no more, no less, everything breaks. They can’t be in a situation at a customer service desk and realize, “The person I’m talking to looks like they’re going to blow a gasket, and they might be a bigger problem if I don’t just upgrade them on this flight.” But we have people doing work like that every single day. I think that’s our main long-term competitive advantage. We look at systems and we know how to adjust to make them function properly. I don’t think machines are taking that any time soon.