MIT Researchers Urge Public Policy Response to AI's Impact on Jobs

A paper built on existing research

The paper is titled Workforce Policy for the Age of AI: Recommendations from the Economic Literature. Its authors are Guy Ben-Ishai, Researcher and Senior Advisor at MIT FutureTech, and Neil Thompson, Principal Research Scientist at the MIT Computer Science and AI Lab and leader of the AI, Quantum and Beyond research group at the MIT Initiative on the Digital Economy. According to the IDE, the paper uses existing research, including work from IDE- and MIT Sloan-affiliated researchers, to offer data-backed strategies for dealing with AI job disruption.

The authors say the public debate is often framed as a choice between mass unemployment and widespread human augmentation, but that the evidence supports neither. Instead, they say the literature points to sharply divergent outcomes on very different timelines: some work will be displaced rapidly, other work augmented gradually, with variation across occupations, industries and employers. Because the impact is likely to be so profound, they write, "a public policy response is imperative." They also argue that such policies are central to deploying AI successfully and integrating it into the economy, making them a U.S. strategic imperative.

Four findings from the literature

The researchers distilled the prior work into four findings.

• Exposure is not automation. Prior research finds that when large language models complete tasks that would take humans three to four hours, their success rate is only 65%. Human error rates for routine and simple tasks are typically 2% to 9% for simpler tasks and 16% for longer tasks. A 10-fold increase in average task duration reduces AI model performance by 11%.
• Expertise matters as much as job counts. Some occupations may gain jobs at lower wages, as in the paper's physician example, where lower-paid nurse practitioners could do work that AI makes easier. Others may see higher wages but fewer jobs, as in its accountant example, where AI automates bookkeeping but remains inconsistent on tax advice.
• Knowing when to trust AI is a skill. AI systems are prone to errors, limited interpretability and non-deterministic outputs, so adoption involves a lot of "last-mile" customization. Workers need to understand when AI tools can reliably improve performance and when they are likely to fail.
• Better AI enables more autonomous work. Prior research suggests LLM error rates are halving about every 2.5 years. If that holds, LLMs could reach a 93% baseline success rate across most professional text-based tasks by 2029. Still, given integration costs, U.S. businesses would automate only about a quarter of the computer-vision tasks (23%) that AI could perform.

The five policy recommendations

The authors say their proposals do not require sweeping reforms of U.S. social safety nets or labor markets. They describe them as targeted, evidence-based programs designed to be fiscally sound, capable of attracting bipartisan support, and scalable if successful. The goal, they write, is not to predict every occupational change, but to build a framework "capable of responding before a temporary disruption becomes lasting economic harm."

The first recommendation is to prioritize both displaced workers and high-productivity opportunities. Prior research finds that workers displaced during periods of high unemployment have historically lost the equivalent of nearly three years of earnings. For them, retraining alone is unlikely to be enough, and the authors point to preemptive help such as job-search assistance and wage insurance. They also call for a focus on transitions into high-productivity, high-impact sectors, noting that U.S. employment is concentrated in slower-productivity sectors and that an aging population increases the burden on a shrinking workforce.

Second, training should be domain-specific and hands-on, and designed in partnership with employers. Workers need to assess AI reliability, exercise judgment and integrate AI into existing workflows. Employer involvement helps keep training demand-driven and tied to actual hiring.

Third, programs should be aligned with shifts in human expertise and distinguish between two risks: job displacement in some occupations and wage erosion in others. Where jobs are eliminated, policy should help workers move to other occupations and sectors through job-search assistance, wage insurance and occupational transitions. Where wages fall, it should help workers preserve or improve earnings, for example through apprenticeships and moves to higher-paid occupations.

Fourth, apprenticeships should be promoted and expanded. The authors say they are an effective mechanism for worker placement, lowering training costs, giving employers a source of screening and signaling, and aligning training with real business needs. The source text is cut off before the fifth recommendation is described in full, though the summary names a national wage insurance program among the proposals.

What it means

The paper's core message is that policymakers should not wait for market forces to sort out AI's effects, and should not base decisions on predictions of either mass job loss or universal augmentation. The authors favor targeted programs that can be tested and scaled, with different tools for workers who lose jobs and workers whose pay erodes. For employers, the recommendations place them at the center of training design and apprenticeship placement.