How Will AI Affect Jobs?

Summary

Allison Nathan hosts MIT economist Daron Acemoglu, MIT’s Neil Thompson, and Goldman Sachs Research economist Joseph Briggs to discuss AI’s labor-market effects. Briggs argues AI is already producing small sector-specific drags on hiring and that a full-adoption scenario could reallocate roughly 9% of U.S. workers over a decade, but with limited year-to-year unemployment pressure and strong long-run job creation. Thompson emphasizes that capability does not equal deployment: privacy, cost, data access, reliability, and adoption frictions will slow and shape real-world impact, while partial automation can sometimes raise wages or expand occupations. Acemoglu is more cautious about near-term complementarity, expecting some limited net job losses in the next five years, concentrated in cognitive routine work such as customer service and back-office tasks. The biggest long-term uncertainty, in his view, is whether AI investment goes toward complementary applications or toward worker replacement, with AI-plus-robotics a major wildcard.

Notes

• The episode frames the central question as whether rapid AI capability gains could cause a large-scale job shock before decade-end.
• Briggs says AI is already visible in a few sectors, but the current macro impact is still narrow.
• Briggs highlights tech, management consulting, and graphic design as the first sectors showing measurable AI-related hiring drag.
• His baseline assumes a 15% productivity uplift after full AI adoption.
• That baseline translates into about 9% of U.S. workers being reallocated during the transition.
• Briggs equates 9% displacement with roughly 15 million workers.
• He argues the shock should be spread over about 10 years, limiting annual unemployment pressure.
• Briggs compares the expected AI transition to earlier waves of technological reallocation.
• He says the long run should still absorb workers because technology historically created most job growth.
• Briggs cites an 80-year history in which about 85% of job growth came from technology-created positions.
• He also says the U.S. labor market is extremely dynamic, creating about 30 million jobs and destroying about 29 million each year.
• A modest 5% faster pace of new job creation would, in Briggs’ view, offset AI displacement.
• Thompson cautions that capability gains alone do not map directly into labor-market disruption.
• Deployment depends on access to the right information, including sensitive data.
• Privacy constraints are a real bottleneck for high-value use cases such as healthcare.
• Thompson says some candidate automations may fail on cost because exacting systems are expensive to run.
• He expects adoption to lag capability growth by a wide margin.
• Large businesses are likely to automate before small businesses, according to Thompson.
• He sees a long tail of tasks that may never be automated.
• Thompson says partial automation should be analyzed from the labor supply side, not just as lower demand for workers.
• Automating a low-expertise task can make a job more attractive and preserve or increase earnings.
• Automating a high-expertise task can lower wages but raise the number of entrants into an occupation.
• Thompson’s taxi example shows how automating expertise can expand the occupation while depressing pay.
• The same taxi example ends with a larger driver pool, even though wages fall.
• Proofreading is his counterexample: spellcheck removed a low-value task and left the higher-value work behind.
• Thompson contrasts AI with narrow tools like databases and Excel, which are highly reliable within a limited scope.
• AI is broader than those tools, but not as consistently correct.
• Hallucinations make AI better as an assistive tool than as a fully modular workflow component.
• Thompson says the dominant labor-market pattern is gradual adjustment, not a sudden wave.
• Acemoglu expects some layoffs or hiring slowdowns by 2027 in jobs where AI can substitute for tasks.
• He says current AI is not yet strong enough to complement workers at scale.
• Acemoglu expects limited net job losses in the next five years.
• He says widespread deployment is constrained by the lack of reliable, easy-to-use foundation-model applications for large employers.
• Acemoglu says agentic advances matter only if developers can package reliable, flexible tools for enterprises.
• He identifies customer service and back-office work as the most vulnerable near-term categories.
• Those vulnerable tasks amount to roughly 8 million to 9 million U.S. workers.
• He says the 10- to 15-year outlook depends heavily on where AI investment goes.
• Complementary AI requires different pretraining, application-level investment, and high-quality domain-specific data.
• The biggest wildcard is AI plus robotics, which could broaden labor exposure dramatically.
• Acemoglu notes that physical tasks are about half of U.S. work.
• He says judgment-heavy middle-manager tasks are another uncertain frontier for agentic AI.
• Acemoglu expects labor income inequality to rise as lower-paid groups face both wage pressure and weaker employment.