Leadership1 publisher2 min readPublished
Starting pay in the most AI-exposed majors fell 13% after ChatGPT, Census researchers find
The Census Bureau study puts that hit on a par with graduating into a large recession. A separate study of 393 employees finds the productivity gains coming back to staff as a higher performance bar.
The Board Room · Leadership desk
What happened
- A Census Bureau study found graduates in the most AI-exposed majors, computer science and software, saw their job odds fall five percentage points and starting pay fall 13% after ChatGPT arrived in late 2022.
- The researchers put that earnings hit on a par in magnitude with what graduates lose when they finish school into a large recession.
- A separate study of 393 employees found regular AI users hit by a cyber whip effect, in which productivity gains prompt organisations to raise performance expectations and burnout rises.
- Gallup recorded 27% of US workers worrying that AI could make their own job obsolete, the highest share it has measured.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- cost A 13% lower starting salary is a saving on a hiring line and a loss carried by one cohort of graduates, and only the saving turns up in a budget review.
- decision A manager taking an AI gain is choosing between banking all of it as output and leaving part of it as slack, and the study points to supportive leadership as what changes the outcome.
- exposure The bar is being raised into a workforce where worry about AI taking one's own job is at a record in Gallup's series, so a higher target gets read as a headcount signal whether or not it is meant as one.
- constraint Because the AI bills before Congress remain unimplemented, a manager setting either the hiring ladder or the performance bar has no external standard to cite and owns the whole judgement.
The two findings sit in different populations. The Census Bureau work follows graduates in computer science and software into the labour market after late 2022 [1]. The other follows 393 people already employed who use AI regularly [3]. Read together they look like one productivity gain charged out twice, once on the hiring line and once to the people at the desks. Neither study puts a figure on the gain itself [2].
The employee study names something a manager controls. Researchers call it the "cyber whip effect": output rises with the tool, the organisation raises what it expects, and burnout climbs [3]. They label the wider pattern the "Generative AI Labor Paradox" [3]. The same study found that supportive leadership can soften the pressure [4].
The drop was measured in the most AI-exposed majors, and the researchers sized the earnings hit against "the earnings losses associated with graduating into a large recession" [2][1].
Pew and Gallup measure different distances. Pew's 37-country survey found most people now expect AI to cost jobs, with more than 70% in the US, Australia and South Korea saying it will lead to fewer jobs [5]. Gallup found 27% of US workers worried AI could make their own job obsolete [6]. The gap between the general belief and the personal one is at least 43 points [1].
Marcus Collins, a business school professor, told Charter that the definition everybody reaches for, "how we do things around here", from Deal and Kennedy in the 1980s, is only half the battle [8]. "But what actually grounds organizational culture are the beliefs reflected in those rituals and behaviors," he said [7]. A performance bar is one of those behaviours, and what it encodes is the organisation's working definition of enough. Collins's next book, Good Soil, is due in February [10].
The bar raised this quarter becomes next quarter's baseline whether or not the tool keeps delivering. In my view the starting-salary number is the one that outlasts the tool, because a cohort hired 13% below the last one sets the reference for the next offer [1].
What to watch
- Whether the graduate pay effect holds for the 2026 cohort in computer science and software, or reverses as technology hiring recovers.
- Whether the cyber whip finding replicates in a larger sample with named performance metrics rather than self-reported pressure.
- Whether any employer publishes the size of the AI productivity gain it is booking, which is the figure both studies work around.