In the fifth of occupations most exposed to generative AI, hiring of master's-degree holders fell roughly 10% relative to the least exposed fifth after ChatGPT โ and the entire decline sits with candidates who already had work experience. First-time entrants saw no change at all (Cortรฉs, Dellarocas & Wang, SSRN, 2026).
That asymmetry is the finding. It is also the part that should worry anyone running a hiring funnel, because it means the market did not stop believing in graduate credentials. It stopped consulting them in exactly the roles you are automating, for exactly the candidates you hire most.
Meanwhile the supply kept climbing. US universities awarded roughly 949,000 master's degrees to the class of 2025, up from 815,000 in 2018 โ growth in every single year, straight through the arrival of generative AI (Dellarocas, The Credential Crisis, 2026). Employers did not run out of credentialed candidates. They started passing on them.
What the Boston University Study Actually Measured
Patricia Cortรฉs, Chrysanthos Dellarocas, and Qi Wang analyzed more than 100 million US job transitions between 2018 and 2025, using Revelio Labs data that reconstructs employment histories from public professional profiles. The design matters: it observes who employers actually hired, not what their job postings claimed to require (Cortรฉs, Dellarocas & Wang, SSRN, 2026).
Occupations were ranked using the "GPTs are GPTs" exposure index โ the OpenAI and University of Pennsylvania measure that scores each occupation by the share of its documented tasks an AI assistant could do at least twice as fast, at equal quality (Eloundou et al., 2023). Critically, that score was fixed the day the technology landed, before employers had time to react.
The team then compared the top exposure quintile against the bottom, year by year, in entry-level positions where a master's is a genuine choice rather than a licensing requirement: accountants, market research analysts, business intelligence analysts, paralegals, writers.
For four years before 2022, the two groups moved in lockstep. Then the lines split โ slowly, then faster, widening every year since.
Three details sharpen it considerably.
The decline is concentrated in non-STEM degrees, especially business master's โ the programs certifying synthesis, analysis, forecasting, and professional writing. STEM master's hiring did not budge. The devalued skills are precisely the ones a language model performs competently.
The decline is strongest at the firms where Census data confirms AI is actually deployed. This is the closest thing to a mechanism test the data allows, and it points the right way.
And it cannot be a supply shortage, because supply grew every year of the window.
One Credential, Two Signals โ and AI Broke the Bundle
The experienced-versus-entrant split looks strange until you stop treating a degree as one thing.
A master's carries two distinct pieces of information. The first is certification: this person can build the forecast, structure the analysis, write the report. The second is a signal of general ability: this person entered a demanding program, survived it, and finished it โ intelligence, persistence, follow-through. Economists have treated the two as welded together since Michael Spence formalized job-market signaling in 1973.
Generative AI is the crowbar that pries them apart. When a model produces a competent first draft of the analysis, the certification component loses value โ the skills are still real, just no longer scarce. The signaling component is untouched.
Now the asymmetry resolves. Evaluating a 24-year-old with a blank rรฉsumรฉ, an employer has almost nothing else to read, so the degree keeps its full weight. Evaluating an experienced candidate, the employer has something better: a work history, a record of shipped outcomes. Once that alternative exists, the degree is valued mainly for the skill certification AI just discounted โ so its weight in the decision falls (Dellarocas, The Credential Crisis, 2026).
The degree did not die. It got unbundled. And the market has already started pricing the components separately.
Your rubric almost certainly still prices them as one.
Why Your Hiring Funnel Still Overprices Graduate Credentials
Here is where this stops being an interesting labor-economics result and becomes an operations problem.
Most mid-market hiring processes still award graduate credentials structural advantage in three places, all of them invisible until you audit for them.
The screen. "Master's preferred" in the posting, or a recruiter heuristic that treats it as a tiebreaker at the top of the funnel.
The rubric. An explicit point or band in the scorecard, applied uniformly regardless of whether the candidate has ten years of relevant work history sitting directly above it on the same rรฉsumรฉ.
The offer. A tenure-and-education matrix that pays a premium for the degree โ a premium that, on this evidence, the wider market is quietly retracting in your most AI-exposed roles.
Each of those is a small bet on the certification component. In finance, analytics, marketing operations, and internal reporting โ the mid-market functions with the highest exposure scores โ you are paying full price for a signal whose informational content has measurably declined.
The sharper version: for an experienced candidate in an AI-exposed role, the degree is now largely redundant with information you already have. You hold their work history. The credential's remaining unique contribution is the general-ability signal, and you are paying a skills premium for it.
Removing the Filter Is Not the Same as Replacing It
The obvious move is to strike the degree requirement. The evidence says that alone accomplishes close to nothing.
The Burning Glass Institute and Harvard Business School studied 11,300 roles at large firms before and after degree requirements were removed. Nearly 40% of firms dropped them. The result: roughly 97,000 additional skills-based hires out of some 77 million annual hires โ fewer than 1 in 700. Almost all of the real change came from just 37% of the companies that made the change on paper (Burning Glass Institute & Harvard Business School, 2024).
Removing a filter does not create a decision rule. It creates a vacuum, and the vacuum fills with the same heuristic operating informally.
What replaces it has to be measured. The strongest available meta-analytic evidence puts structured interviews at r = .42 for predicting job performance, against r = .19 for unstructured ones โ the single largest validity gap available to a hiring team, and it costs nothing but discipline (Sackett, Zhang, Berry & Lievens, Journal of Applied Psychology, 2022).
That is the substitution the Boston University result argues for: not credential-blind hiring, but credential-right-sized hiring, with a validated capability measure carrying the weight the diploma used to.
Where This Argument Could Be Wrong
I would not act on the decimal. I would act on the direction. The difference is worth being precise about.
The paper is an SSRN working paper, not yet peer-reviewed. The pre-trend is clean and the design is careful, but the magnitude can move under review.
Exposure is also not replacement. A high index score means a role's documented tasks overlap with what a model can do โ not that a model did them. It is a measure of technical feasibility, fixed at one moment in time.
And profile-derived data carries selection. Revelio reconstructs employment histories from public professional profiles, and who maintains one, how completely, is not random. That could correlate with both credentials and occupation in ways the design cannot fully absorb.
Now the part that holds regardless. A credential premium that tracked steadily for four years diverged the moment a general-purpose text model arrived, in the occupations that model handles well, for exactly the candidates about whom employers already hold better evidence. Halve the effect and the operational conclusion is unchanged.
What Changes in a 50โ500 FTE Company
This is a process change, not a spend. It fits inside a quarter and needs no approval beyond your own.
Audit where the credential still carries weight. Pull your open reqs and scorecards for the roles with high AI exposure โ finance, analytics, marketing ops, internal reporting, technical writing. Find every place a graduate degree earns a screen advantage, a rubric point, or a pay band. Most operations leaders have never listed these in one place and are surprised by the count.
Split the rule by candidate stage. The research supports keeping credential weight for first-job entrants, where you have little else to read, and cutting it sharply for experienced hires, where you already hold the better signal. One rule for both stages is now demonstrably wrong in one direction or the other.
Put a measured instrument in the gap before you open it. A structured interview with a fixed question set and an anchored rating scale, or a scored work sample built from the actual job. Removing the filter without this is the 1-in-700 outcome (Burning Glass Institute & Harvard Business School, 2024).
A note on the internal version of this problem
The same logic runs inward. If your promotion criteria or high-potential program awards points for a graduate degree, you are applying a decaying signal to people whose entire work history you already own โ the population where its informational value is lowest of all. That is the cheapest place to fix this, and the least likely to be on anyone's list.
One Decision This Quarter
Open the scorecard for your most AI-exposed open role and find the line where a master's degree earns its points. Then answer one question: what does that line tell me that this candidate's last three years of work does not?
If you cannot answer it in a sentence, you are not screening for capability. You are paying a premium for information the market has already repriced โ and the Boston University data says the discount on graduate credentials in AI-exposed roles is widening every year.
The degree did not stop meaning something. It stopped meaning what your rubric says it means.
