Four years of AI job loss shows up as 38 people in a 1,250-worker panel

Three percent of 1,250 is about 38 people ^[1], and the arithmetic on 38 people deserves more attention than the percentage does: the standard error on a 3% share at that sample size is roughly 0.48 of a point, which puts a conventional 95% band at about 2.1% to 4.0% ^[2], or somewhere between 26 and 49 respondents ^[3]. The 3% to 4% who told YouGov they were not sure whether AI had changed their job at all ^[5] work out to 38 to 50 people ^[5], so the unsure group is at least as large as the displaced one. That is not a knock on the instrument, which says plainly that it is a snapshot of self-reports from a study in progress that has not been peer reviewed ^[2], and which is one of the few efforts to ask workers directly rather than model them from the outside ^[6].

The design detail that matters to anyone pricing AI against a payroll line is that every respondent had a job and all of them opted into an online panel ^[9], a limitation the second write-up concedes when it notes that workers displaced by AI may be underrepresented in exactly this kind of sample ^[13]. An employed-only survey cannot count the person who left in 2024 and never came back, and it certainly cannot see the requisition that was quietly never opened, which is where labor cost savings actually land. The researcher himself flags that the field period ran against a 4.1% official unemployment rate while the workforce was shrinking and AI was being reported as a leading reason for many layoffs ^[8]. Set that against the projections of hundreds of millions of jobs affected globally published by Goldman Sachs, McKinsey and the World Economic Forum ^[11], measured 3.75 years after ChatGPT shipped ^[12], and the gap is either evidence or an artifact.

The one channel the survey measures cleanly runs the wrong way for a margin story. About 113 respondents ^[6] reported a promotion tied to AI skills against 75 ^[4] who landed a role that did not exist before, a net three points of reported gain over loss ^[4]. A skills premium is money moving toward labor, not away from it, and cryptobriefing reads the same 6% as an investable segment of oversight and evaluation work ^[10] where the researcher reads it as limited effect ^[6]. Both readings fit the number, but only one of them is a P&L.

The counter-thesis comes in a few shapes. One is attribution failure, where the worker codes a displacement as a reorg. Another is censoring, the panel problem described above. The more interesting version is timing: if AI substitutes at the hiring margin, the effect appears in who never gets hired, which no survey of the employed can detect. This is probably wrong, but I read AI's near-term labor economics as running through wages before headcount, which is the expensive sequence for an employer. Falsification is cheap and specific: a second wave of this instrument printing 8% displacement, or exposed-occupation payrolls showing entry-level hiring fall away without a layoff count to match.