Economists Are Warning That AI Job Forecasts Are Built on the Wrong Data

The AI jobs debate may be running ahead of its evidence

Within the technology sector, predictions about AI-driven job loss have become increasingly dramatic. MIT Technology Review captures the mood clearly: executives and researchers are openly discussing recession risks, the breakdown of early-career ladders, and the possibility of AI acting as a broad labor substitute. But the publication also highlights a more sobering counterpoint from economist Alex Imas of the University of Chicago: the data tools commonly used to estimate labor disruption may be deeply inadequate.

The core criticism is that economists and policymakers are relying too heavily on task exposure. If a job contains tasks that AI could plausibly perform, that job is often treated as being at risk. Imas argues that this is not enough. Exposure, in his words as quoted in the source text, is not a meaningful predictor of displacement on its own.

Why task exposure is too blunt

The article explains the logic through a familiar example. Jobs are bundles of many tasks, some of which may be automatable and some of which may not. Researchers have used a government task catalog, first launched in 1998 and regularly updated, to estimate how exposed occupations are to AI. OpenAI used this type of data in December to judge occupational exposure, and Anthropic later compared those task lists with millions of Claude conversations to see which tasks users were actually performing with AI.

That sounds rigorous, but the problem is structural. A job is not simply the sum of its automatable tasks. Some tasks are central, some are peripheral, and some are tightly linked to trust, regulation, or in-person judgment. Replacing or augmenting one task does not automatically erase the role around it. Exposure data can therefore tell us where AI touches work without telling us how employment will change.

The missing data is worker-level reality

MIT Technology Review says Imas is calling for economists to gather a different kind of evidence: data that captures what is actually happening to workers as AI tools enter the labor market. That call matters because most of the public debate remains dominated by projections, anecdotes, or company-level rhetoric rather than longitudinal evidence about wages, hours, hiring, career progression, and substitution.

In practical terms, that means the debate is happening in the wrong order. Society is arguing about policy responses before it has built the measurement system needed to understand the scale and shape of the problem. If AI affects labor markets unevenly, with strong variation by age, sector, seniority, and geography, coarse occupation-level exposure scores may obscure more than they reveal.

Why this matters now

The urgency is not only academic. The article notes that workers are already panicking and lawmakers have not articulated a coherent plan for what comes next. That is a dangerous combination. When public fear is high and evidence is weak, policy can become reactive, symbolic, or captured by whichever narrative is loudest.

Even economists who previously warned against overreading AI’s labor impact are, according to the piece, moving closer to the view that this technology could have an unprecedented effect on work. That does not validate every apocalypse claim. It does suggest that waiting passively for better data to emerge on its own may be a mistake.

A measurement gap can become a policy failure

The most important idea in the piece is that bad measurement does not just create academic confusion. It can directly weaken policy capacity. If governments do not know which workers are being displaced, which roles are being transformed, or where early-career ladders are starting to break, they cannot design targeted responses. Training policy, safety-net planning, education reform, and even tax debates all depend on understanding what AI is actually doing inside firms and occupations.

That is why the call described in the article resembles an institutional challenge as much as an economic one. Building better labor-market data around AI may require coordination across researchers, employers, and public agencies. It may also require moving faster than the normal pace of labor statistics, which often lag behind real changes in how work is organized.

The future of work debate needs better instrumentation

One reason AI labor arguments have become so polarized is that the conversation is taking place with weak instrumentation. On one side are broad claims that AI will soon do nearly all jobs. On the other are reminders that mass job losses have not yet appeared clearly in aggregate data. Both can be true in limited ways while still failing to capture what is happening beneath the surface.

MIT Technology Review’s contribution here is to identify the gap between exposure and displacement as the key analytical fault line. That distinction deserves more attention. A job can be highly exposed to AI and yet remain durable for years. Another can be only partially exposed but vulnerable because junior tasks disappear first, cutting off the pipeline that produces future experts.

The next serious AI labor story will likely be statistical, not rhetorical

The strongest takeaway from the piece is that society needs less theater and better evidence. Grand forecasts about total labor substitution may dominate headlines, but they are not a substitute for disciplined measurement. If economists are right that current tools are abysmal, then the next crucial step is not another debate panel about whether AI will destroy work. It is a sustained effort to gather the worker-level data needed to see what is really changing.

Until then, both optimism and panic will remain underdetermined. The future of work debate is now large enough that the absence of better evidence is itself becoming one of the most important facts about AI.

This article is based on reporting by MIT Technology Review. Read the original article.