The AI jobs shock may begin as a quiet productivity J-curve

Hard Fork's latest episode opens with the Apple–OpenAI fight over alleged hardware trade secrets, then turns to a more consequential question: what would it look like if AI changed the labor market faster than economists could measure it? The interview with Erik Brynjolfsson is useful because it does not treat today's employment numbers as a verdict. Its argument is that the visible shock may be delayed by the time it takes firms to redesign work.

Brynjolfsson is a Stanford economist, a senior fellow at the Stanford Institute for Human-Centered AI, and director of the Stanford Digital Economy Lab. His work focuses on the economic effects of digital technology, including AI. 1 The episode's central claim is not that mass unemployment has already arrived. It is that the absence of a headline collapse is weak evidence that the transition is harmless.

Brynjolfsson helped organize "We Must Act Now", a statement signed by nearly 200 economists, researchers, and technology leaders. The New York Times reported that the signatories included 15 Nobel laureates and chief economists from OpenAI and Anthropic. The statement says AI could become radically more powerful over the next decade, producing an economic transformation larger than the Industrial Revolution but compressed into a much shorter period. 2

The wording is notably cautious. The statement names both large-scale job displacement and major gains in living standards. Its request is for economists, policymakers, and technology leaders to understand the economics of transformative AI and build incentives, guardrails, and institutions that steer it toward complementing people. 3 That is not a forecast of a specific unemployment rate. It is a warning about a preparation gap.

That distinction matters. A prediction can be tested against a near-term number and dismissed if the number does not move. A preparation argument asks a different question: are firms, schools, labor institutions, and governments building the capacity to respond before the effects become obvious?

A skeptical listener in the episode makes a fair point: job losses rarely have one cause, so it is difficult to isolate AI in the data. Brynjolfsson does not deny that problem. He says the research examined other explanations, including interest rates, remote work, technology overhiring, and changes in education. His position is that analytical difficulty is a reason to improve the measurement, not to stop looking. 4

He uses the idea of a productivity J-curve to explain why the labor impact can lag behind the capability curve. Electricity did not transform factory productivity the day motors were installed. Firms first reproduced old processes with the new technology. The larger gains came later, after factories and workflows were reorganized. Brynjolfsson's analogy is that AI may be in that earlier stage now: companies are adding a powerful tool to existing jobs before changing the jobs themselves.

The most concrete signal he cites is not economy-wide unemployment. It is a divergence within the workforce. In the conversation, he points to a Stanford dashboard showing early-career employment down 2.7 percent year over year while mid-career employment rose 1.6 percent. Those figures do not prove that AI caused the entire difference, and the episode does not present them as proof of a completed labor-market transformation. They do suggest where to look: at entry points, task composition, and the bargaining power of workers who have not yet accumulated experience. 4

That is a more useful frame than asking whether AI has "taken all the jobs". A system can preserve the number of jobs while changing who gets hired, which tasks are assigned to junior workers, how quickly people are expected to produce, and what counts as an acceptable first draft.

The interview becomes more practical when it asks whether companies can intentionally build AI that complements workers instead of merely replacing them. Brynjolfsson's answer is that technology is shaped by incentives and management choices. The same capability can be used to cut headcount, or to develop a new service, handle more customers, improve quality, and reduce employee churn.

He describes a chief financial officer who wanted AI to improve return on investment but worked inside a company that measured the program mainly through headcount reduction. That metric makes substitution look like the natural definition of success. Change the metric, and other forms of value become visible. The point is not that every AI project should preserve every job. It is that an organization cannot claim to have discovered AI's productivity impact if it measures only the labor it removes.

Brynjolfsson also points to tax incentives that favor capital over labor. If replacing a worker reduces a company's tax burden relative to hiring or training people, the economy is quietly subsidizing one direction of technical change. A policy that wants complementarity would have to address those incentives rather than rely on corporate goodwill.

For practitioners, this translates into a concrete evaluation question: when an AI system is deployed, what new output, service level, or capability does it make possible, beyond the hours or roles it eliminates? If the answer is only "fewer people", the organization is measuring automation, not the full productivity opportunity.

The conversation also contains a quieter institutional warning. Brynjolfsson says many economists and students are moving into frontier AI labs, where they can access valuable data and do important work. He has declined those offers because he values an independent perch from which to study the industry. He is careful not to portray academia as perfectly neutral. His point is that a field needs researchers who are not perceived as speaking for a particular company.

That independence is especially important when companies are both deploying the technology and defining the metrics used to judge it. The AI industry can supply experiments, data, and technical expertise. It cannot be the only institution deciding whether the results are good for workers or society.

Hard Fork's episode therefore lands on a modest but demanding conclusion. The AI jobs debate does not need a more confident prediction. It needs earlier measurement of where work is being reorganized, better incentives for augmentation, and institutions capable of saying when a claimed productivity gain is simply a transfer of risk onto workers. The quiet phase of a transformation is still a phase of the transformation.