The August 2026 revision of "Canaries in the Coal Mine" runs 140 pages. Page 140 is Appendix L. It is titled "Changes from the August 2025 Version," it is about two hundred words long, and it is where the authors report that their most carefully controlled estimate got smaller and stopped being statistically significant.

Nobody is going to read page 140. The abstract is on page 1, and the abstract says 19 percent.

This is not an accusation. Appendix L exists because Erik Brynjolfsson, Bharat Chandar and Ruyu Chen wrote it, and they did not have to. Most working papers get quietly reposted with a new date and a bigger N. This one carries a section whose entire function is to tell you which of its own numbers moved in the wrong direction. That is the behavior you want from people holding the only high frequency payroll panel that can answer the question at all.

But the two numbers travel at different speeds. The 19 percent has a Stanford press release, a dashboard, and a headline. Appendix L has a page number.

The data are a balanced panel of firms drawn from ADP payroll records, between 3.5 and 5 million full time employees observed monthly from January 2021 through June 2026, with a longer 2018 panel used for pre-trends. ADP serves firms employing more than 26 million US workers. The analysis sample is a subset of that, and the distinction matters more than it sounds.

Fact 4 is the one most likely to be misread, and it is the one that should change how you think about the problem. A falling employment stock can come from letting people go or from not bringing people in. The paper finds the second. Separation rates fell across the board after 2022, and among young workers they fell at least as much in the most exposed occupations as in the least. Displacement would show the reverse.

The whole divergence sits in the hiring rate. Firms are not clearing out their juniors. They are declining to create the role in the first place, which is quieter, involves no severance, and generates no news story. It is also much harder to reverse, because the person who was never hired has already gone somewhere else.

The paper survives a long robustness list. Drop technology firms, drop computer occupations, control for occupational interest rate exposure, swap in five alternative AI exposure indices, include part time workers, allow firm entry and exit. The pattern holds each time. Two things do move it.

Education. Adding occupational college share to the long difference regression cuts the most exposed quintile coefficient for 22 to 25 year olds from 0.179 down to 0.091, significant only at the 10 percent level. With all three controls together it falls to 0.080 and loses significance entirely. The authors handle this honestly. College share may be a confounder, in which case the smaller number is right. It may also be the channel itself, since generative AI substitutes best for exactly the codified knowledge that formal schooling produces, in which case controlling for it discards the effect. They present both and call it a bracket. That is the correct move and it is also an admission that the true value is somewhere in a range that includes numbers small enough to be uninteresting.

The national benchmark. This is the larger problem.

Then the sector breakdown, which is where this stops being an abstract methods argument.

The mechanism the authors propose is that generative AI substitutes for codified knowledge, the formal standardized content you can learn from a textbook, and complements tacit knowledge, the kind acquired through practice and repeated exposure to real situations. They build crude proxies for both. Occupations heavy in codified knowledge show slower entry level growth. Occupations heavy in tacit knowledge show faster growth for mid career and senior workers. The codified gradient does not survive a control for college share. The tacit gradient for experienced workers does.

Anyone who has trained a resident will recognize the distinction immediately, because graduate medical education is an institution built entirely around it. Four years of medical school delivers the codified layer. Then we take people who have already passed every written examination we can devise and put them on a ward for three to seven years, supervised, because we have never found another way to transfer the part that is not written down. The whole apparatus exists on the premise that codified knowledge is necessary, insufficient, and cheap relative to the other kind.

Brynjolfsson's paper says AI is now very good at the cheap part. Medicine has known that the cheap part was cheap for a century. What medicine has not had to think about is what happens to an apprenticeship when the tasks used to justify paying a junior person to be present are the tasks that got automated. Residency is protected here by accreditation rather than by economics, which is a real protection and an accidental one.

The entry level job and the training pipeline are the same object viewed from two sides. The paper measures one side. Nobody is measuring the other.