On July 24 we covered Stanford’s ADP payroll study in Entry-level jobs are down 13% since ChatGPT, working off the then-current data vintage.
On August 12 the same three authors posted a revision. The number moved. More usefully, they finally named a mechanism.
Six facts, and the one everyone skips
Erik Brynjolfsson, Bharat Chandar and Ruyu Chen are still working from ADP administrative payroll records, still splitting the workforce by age and by how exposed an occupation is to generative AI.
Their first stated fact is that they see no widespread, economy-wide job displacement associated with AI. That is fact one in the paper, not a disclaimer bolted to the end.
Fact two is the divergence. Workers aged 22 to 25 in highly AI-exposed occupations now sit about 19% below where they would be had they kept pace with same-aged workers in less-exposed occupations. Experienced workers show no comparable gap.
Fact three gives the slope. At the July 2025 vintage the shortfall measured 15%. As of June 2026 it measures 19%. The line did not peak and settle.
The levels are in a footnote and read harder than the ratio. Between November 2022 and June 2026, employment of 22-to-25-year-olds in the two most exposed quintiles fell about 11%, while the same age group in the three least-exposed quintiles grew about 10%. Same cohort, same window, opposite signs.
Fact four names the channel: the adjustment runs through reduced hiring, not increased separations.
Fact five splits the exposure. Declines concentrate in occupations where AI usage automates human tasks. Where AI is used to complement workers, employment is flat or rising, and rises most for experienced staff.
Fact six: the adjustment shows up in employment, not in base pay.
Codified knowledge and the stuff you cannot write down
The revision’s real contribution is the mechanism section.
The authors split knowledge in two. Codified knowledge is formal, standardized and documented, teachable through education, textbooks and written procedure. Tacit knowledge is acquired through practice, mentorship and repeated exposure to real situations.
The data lands cleanly on both sides of that line. Employment fell among young workers in occupations that lean on codified knowledge. Employment rose among experienced workers in occupations that lean on tacit knowledge. The authors’ reading is restrained: generative AI is particularly effective at reproducing and applying knowledge already encoded in text, while experience-based knowledge remains harder to replicate.
Applied to an actual 23-year-old, that is blunt. Nearly everything a new graduate brings is codified. It is what the coursework taught, what the manual documents, what onboarding walked through. It is simultaneously the reason they got hired and the exact material the model learned best. They hold no tacit knowledge yet, because tacit knowledge only accrues through time in roles that are now getting scarcer.
One more finding from this revision deserves its own line: women face greater AI exposure on average. The authors say they intend to monitor it.
The authors argue against themselves
This version is more cautious than the one we wrote up in July, and that matters.
Start with what they ruled out. The divergence survives excluding technology firms and computer occupations. It survives controls for interest-rate exposure and remote work. It survives including firms that enter and leave the sample. It survives alternative measures of AI exposure. And a 2026 Census Bureau working paper using government administrative data shows broadly consistent raw patterns by age and industry.
Now their own three cautions. Account for education and the gap between more- and less-exposed young workers shrinks. Some differential trends predate widespread generative AI use. And the estimated gaps run larger in the ADP analysis sample than in national survey benchmarks, with the discrepancy concentrated in education, health care and public administration.
A fourth is more technical and easy to miss. After improving their data pipeline, the raw patterns held, but estimates that account for overall firm-level hiring changes became more sensitive to specification choices. The authors say this raises legitimate questions both about how much of the pattern AI actually causes and about how well the ADP sample generalizes.
They state outright that they do not view this paper, or any single study, as definitive.
Where it sits against the rest of the year
The revision earns its weight when you stack it against the other entry-level evidence.
On July 5 we covered Entry-level postings are down 35%, but the Fed blames something else: the hiring side is visibly collapsing, and the Fed’s work could not pin it on AI. On July 21, Entry-level jobs in law and consulting fell 35% since 2023 gave the sector cut, with the two most codified-knowledge professions in the economy losing their bottom rung fastest.
Put together, the three say something narrower than the headlines and more useful: entry-level work is shrinking, the shrinkage concentrates in codified-knowledge occupations, and nobody has yet established the causal link to AI.
200 economists, 16 Nobels, four sentences of maybe covered the same reluctance from the consensus side. Stanford’s revision is that same caution, with the mechanism advanced by one notch.
What a 23-year-old should actually take from it
First, what the paper does not say. It does not say AI took your job. The authors describe patterns and refuse the causal claim.
Now what it does say.
The damage is on the hiring side, not the firing side. People already in seats are far less affected, which prices moving jobs higher than staying in one. That spread is new.
Exposure tracks the task, not the industry. Two people with “analyst” on the badge sit on opposite sides of this line if one produces standardized reporting and the other holds a client relationship. The question worth asking about a job offer is how much of the work could be written into a procedure manual, not whether the sector sounds AI-proof.
Whether AI automates or complements sets the direction for the whole occupation. In complementary occupations, employment for experienced workers is still rising. That is fact five, not reassurance.
And accumulate tacit knowledge as fast as the job allows. Mentorship, live client contact, situations with no documented answer. This used to be called paying your dues; it is now the only stable side of the line. The catch is that the roles which supply that exposure are exactly the ones contracting fastest. The paper does not resolve that, and does not pretend to.
One scheduling note. Stanford’s Digital Economy Lab launched its AI Economic Indicators project alongside this release, and its Canaries Dashboard will refresh the paper’s key results monthly. The gap is a monthly series now, not an annual paper. When the next reading lands, and which way it moves, will tell you more than this 19% does.
Sources
- No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19% (Stanford Digital Economy Lab, August 12, 2026)
- Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence (revised paper, August 2026, PDF)
- Tucker 2026, CES-WP-26-27 (U.S. Census Bureau working paper)
