Workforce policy for the age of AI
B · Useful, limited Brookings synthesis by MIT FutureTech's Neil Thompson and Guy Ben-Ishai that turns cited empirical work into five policy principles; credible authors, but a narrative policy synthesis with no new data or systematic review method.
Summary based on the full text·Original: about 29 minutes to read
In short
Guy Ben-Ishai and Neil C. Thompson of MIT's FutureTech project review the economic evidence on AI and work for Brookings (15 September 2026). They argue that exposure scores, the share of an occupation's tasks AI could do, are the wrong basis for policy: what matters is whether AI is commercially viable for a task and whether it raises or lowers the value of human expertise. They propose five targeted measures, from support for displaced workers to apprenticeships and federal wage insurance.
Key numbers
- 65%AI success rate at end-2025 on tasks taking humans three to four hours
- 11%drop in model performance for each tenfold increase in task duration
- 2.5 yearsapproximate halving time of LLM failure rates (Mertens et al.)
- 13%lower earnings among low-wage workers six years after displacement
As reported by the source.
English·Web page·opens brookings.edu
What it reports
- At the end of 2025, AI models completed tasks taking humans three to four hours with a 65% success rate (Mertens et al.).
- For comparison, the HEART human-reliability framework puts human error at 2% for routine low-skill tasks, 9% for simple tasks with limited attention and 16% for complex tasks; the authors note these are not directly comparable.
- A tenfold increase in task duration lowers model performance by 11%; healthcare support tasks (just over an hour) score highest, legal and engineering tasks (over 20 hours) lowest.
- Model failure rates halve roughly every 2.5 years; if that holds, models could reach a 93% baseline success rate on most professional text tasks by 2029.
- At current integration costs, US businesses would automate only 23% of computer vision tasks AI can technically do (Svanberg et al.).
- Displacement in high-unemployment periods has cost workers earnings equal to 2.8 years of prior pay (Davis and von Wachter); low-wage workers earn 13% less six years after displacement.
- US mobility: roughly 90% of children born in the 1940s out-earned their parents, versus 50% for those born in the 1980s.
- Five recommendations: prioritise displaced workers and high-productivity sectors; domain-specific, employer-built training; programmes matched to whether AI raises or lowers the value of expertise; AI apprenticeships; federal wage insurance.
How they got it
A policy synthesis of recent economics and AI-capability research, including task-level reliability studies, the Autor-Thompson expertise framework, field experiments on AI use, and the displacement literature. No new data collection.
Limits
- The 2029 projection extrapolates a trend and the authors flag the uncertainty.
- Recommendations are US-specific (federal programmes, at-will employment) and are proposed as pilots, not tested designs.
- Human and AI error rates come from different frameworks and are not directly comparable.
Why it matters for companies
The paper gives managers a better test than "is this role exposed": how long are the tasks, what error rate can the process tolerate, and what does integration cost. Its expertise argument is useful for workforce planning: in some roles AI will raise pay and cut headcount, in others it will open the role to less-qualified people at lower pay. It also makes the case that generic AI literacy courses are not enough and that people need hands-on, domain-specific practice in judging when the tool is wrong. The argument for apprenticeships addresses a direct company problem: if AI lowers formal entry barriers, observed performance on the job becomes the main way to assess candidates.
Summary based on the full text. Read the original for the full argument and data.Read the original on brookings.edu