A study found that 90% of U.S. working hours involve tasks not well-covered by current AI benchmarks.

"Study finds 90% of US working hours are spent on tasks not covered well by AI benchmarks" — Hacker News (best) · AI & Technology

17 Oct 2026

View source

What it means

This study from Mercor, a labor research group, maps existing AI benchmarks (like those for coding, math, or image recognition) onto the tasks that make up actual U.S. jobs. It found that only 208 out of 923 occupations have even a single benchmark task with at least five associated prompts, and approximately 90% of all working hours in the U.S. are spent on tasks that have little or no coverage from any current AI benchmark. This suggests that the capabilities measured by today's AI evaluations are a narrow slice of the skills actually required in the workforce.

Accuracy & sourcesGrounded

The claim is directly from Mercor's own research page, which is the primary source. The page includes a detailed breakdown of the methodology (linking O*NET occupation data to benchmark tasks) and the figures (90% of hours, 208 of 923 occupations). The search results confirm that multiple outlets (Hacker News, daily.dev, OpenAI community) are reporting on the same study, making the factual claim solid.

How new

The study was published on 2026-10-16, making it genuinely new material — this is the first wave of reporting on this specific analysis.