Science1 publisher3 min readPublished
Top-1% industry AI researchers out-earn matched academics by $1.5 million a year
The NBER analysis matches specialisms before it compares pay, which is the right control, but the headline gap describes the tail of the industry distribution, and the academics Nature asked all stayed.
The Scientist · Science desk
What happened
- An NBER analysis found that the top 1% of industry AI authors earn US$1.5 million more per person each year than academic researchers working on comparable specialisms.
- The comparison was built by identifying researchers with similar specialisms on each side of the academic and private-sector line, rather than comparing sector averages.
- Nature's accompanying article collects 14 university academics in AI and computational science explaining why they stay and what might push them into industry.
- Those who stayed cited the freedom to pick their own questions and the satisfaction of training students, while complaining about the difficulty of finding stable funding.
- Nature also reports that hybrid roles, splitting time between a technology company and a university, are being adopted more often by researchers who want both.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint A 99th-percentile gap cannot be turned into a retention line item, because the figure a university needs is the premium facing an ordinary senior hire, and that one is absent from the published account.
- decision Institutions now weighing hybrid appointments as the retention instrument are choosing on a reported trend with no prevalence figure attached, which makes the choice a bet rather than a costed policy.
- exposure Open reporting of failed experiments and reusable artefacts rests on people whose stated reasons for staying are non-monetary, so it is exposed to any change in how those reasons hold up.
- precedent If a dual company post becomes the ordinary way to keep a senior AI researcher, the independence questions that used to sit between universities and industry move inside the individual appointment.
Matching on specialism is the part of this analysis that deserves respect. A raw comparison of academic and industry pay in AI conflates the person, the subfield and the job; the National Bureau of Economic Research instead identified researchers with similar specialisms on both sides of the line before comparing what they earn [1]. So the $1.5 million is a gap between people whose work is recognisably alike [2].
Then the denominator. The figure belongs to the top 1% of industry authors [2], and that makes it a tail figure, not a retention budget. Nature's account gives no median, no mean and no sample size, and does not say which years the analysis covers [12]. A head of department deciding whether to counter-offer needs the gap facing the person sitting across the desk, and that number is not in the piece.
The tail figure is still large in the currency academics actually work in. Stewart Clark, a computational physicist at Durham University, said that in academia a project can run for a decade [8]. Ten years at a $1.5 million annual gap is $15 million forgone [11], by the individual who would have landed in that top 1%, which is the condition carrying most of the weight in that sentence.
Nature asked 14 academics in AI and computational science why they stay [3] and got research freedom and the pleasure of training students, alongside complaints about unstable funding [4]. Damen said that three former PhD students becoming faculty members is a greater achievement than the research, because research is state of the art today but not a legacy [10]. Every person quoted had stayed. That is selection on the outcome, and it means the interviews cannot estimate how many comparable people left, or what offer moved them. The former students now at Anthropic and OpenAI earning millions a year appear as anecdote, mentioned by the people who did not follow them [7].
Which leaves the hybrid appointment. Nature reports that such arrangements, splitting time between a technology company and a university, are becoming more common [5], and the quoted material names one holder: Dima Damen, a computer scientist at the University of Bristol and at Google DeepMind [6]. A stated trend and one named case do not establish that university AI labs are unable to hold senior staff without a company affiliation attached. They establish that some people have found the arrangement worth having.
What is missing is any measure of whether the pay gap is affecting the science itself. The capability at stake is the one Noah Smith of the University of Washington describes: somebody reporting experiments in the open, including what did not work, and sharing artefacts others can reuse [9]. That output is not measured anywhere in this analysis, so no thinning of it is shown here either. The number that would settle the retention question is an attrition rate for senior AI faculty, not a pay level.
What to watch
- Whether the NBER working paper reports median and quartile gaps, and whether industry figures include equity, which is what a counter-offer turns on.
- Any institutional count of hybrid university-company AI appointments, with the fraction of time each involves.
- Departure rates for senior AI faculty by year, which is the outcome the pay gap is currently being used to predict.