Automation will transform work over the next few decades, but it won't be a binary "before and after," at least not at the job level. Most occupations won't simply disappear. Instead, the tasks within them will change.
Economic context is a key consideration for these changes. Is automation cheap enough that replacing human work makes sense? In AI, direct costs are falling dramatically for equal capability. Other costs, such as maintenance, understanding, collaboration, management, and safety, are often hard to measure. Saving money can also create more human work: When spinning machines made yarn cheap in late-1700s Britain, demand for hand-loom weavers soared to turn that yarn into cloth.
Automation and augmentation don't always push wages in the same direction, either. In our paper "Expertise," David Autor and I ask whether automation makes experts more valuable or automates away their expertise. Drawing on U.S. wage data, we argue it can do both, depending on the tasks within a job that are automated. Computers took over the routine bookkeeping of accounting clerks, which was the simpler part of the job. The work that remained required more expertise, so wages rose and fewer people were qualified. By contrast, inventory clerks lost the harder parts of their job, such as computing prices and compiling inventories. What remained, such as counting and stocking shelves, was easier to learn, so wages fell and more people could do the work.
My lab's work has convinced me that we can't predict AI's effect on employment from how many tasks it can automate. We need to know which tasks it will automate and how that changes the expertise a job requires.
AI progress appears to lift performance across many tasks at once rather than advancing in concentrated, successive leaps. For a paralegal, that means AI gets somewhat better at research, drafting, and summarizing all together, rather than mastering one task while the rest stay untouched. Because progress is broad, most jobs will change gradually before they disappear. That gives us time, as a society, to decide where we want automation and where we don't.
