What jobs is AI bad at?

In short

A newsletter roundup on the economics of AI and jobs. An OpenAI study of real-world tasks across 44 occupations finds AI already beats experts in several fields but struggles with roles like industrial engineers and nurses (and is surprisingly capable at private-detective work). It also covers a research agenda defining transformative AI as 3–5x historical productivity growth, and Pascual Restrepo's theory that wages could trend toward zero under either compute- or technology-constrained automation.

This is one of the Tech Futures Project newsletters, a roundup of the most useful recent research on AI and the economy, filtered through the publication’s political-economy lens.

The standout is an OpenAI study that actually measured model performance against experts across 44 occupations, with results that scramble the usual intuitions: AI already beats specialists in several white-collar fields, yet stumbles on industrial engineering and nursing, and, oddly, excels at the work of a private detective. The lesson is that “which jobs are safe?” has a far messier answer than the headlines suggest.

It also surfaces two pieces of serious labour economics: a research agenda for what would even count as “transformative” AI, and Pascual Restrepo’s unsettling argument about the paths by which AI could push wages toward zero. A good primer if you want the economics, not the vibes.

Key takeaways

An OpenAI study across 44 occupations finds AI already outperforms experts in some fields but is weak at industrial engineering and nursing.
AI proved surprisingly capable at private-detective work, automation potential is uneven and counterintuitive.
A leading research agenda defines transformative AI as sustaining total factor productivity growth 3–5x historical averages.
Pascual Restrepo argues wages could trend toward zero, gradually (compute-constrained) or unevenly (technology-constrained).

Read the full piece

This is a summary. Read the complete essay, with all the sources and argument, on Substack.