Narayanan and Kapoor: the WARN data does not support the AI layoff story

Arvind Narayanan and Sayash Kapoor argue on their Normal Tech blog that the headline story of 2026, AI replacing software engineers, mostly is not showing up in the actual numbers. They build the case on hard data. New York added an AI checkbox to its WARN Act layoff filings in March 2025. In the first year, of roughly 25,000 workers covered by those filings, only about 46 were marked AI-driven. That is two-tenths of a percent. A Forrester analyst found 90% of companies blaming AI for cuts were not yet running applications that could plausibly do the work. An HBR survey found 21% of executives made large anticipatory headcount cuts against AI, while only 2% attributed cuts to real deployment results.

Their frame for software work is a "decide-execute-deliver sandwich". AI has compressed the execute layer, which is code. Deciding what to build and verifying that it ships correctly, the two slices around it, are still on humans. A GitHub study across 100,000 developers found agent use multiplied lines of code by eight while raising actual releases by only 30%. The human bottleneck lives in that gap. The authors also walk through the public layoff stories at Block, Snap, and Intuit, and find each one had its own non-AI cause, from activist investors to plain over-hiring during the pandemic.

The closer is Jevons' paradox. Cheaper software production drives more software, so aggregate engineer demand can grow even as the shape of the work shifts.

Why it matters

If you are deciding whether to hire engineers or cut them, the WARN number is a better signal than a CEO press release. The work that survives is deciding what to build and standing behind what ships, and that is also where engineer career value sits.