In jobs exposed to AI, the first rung of the career ladder is getting thinner, mostly because employers are hiring fewer young people, not firing more of them. The judgment that rung used to teach still matters. Our view is that new workers can build it now by checking AI's work, explaining it, and owning the result, and that small businesses can train for exactly that.

Last updated: November 1, 2026

In the Stanford Digital Economy Lab's "Canaries in the Coal Mine?" study, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen found that employment of workers aged 22 to 25 in AI-exposed jobs is 19% below where it would otherwise be. That is the August 2026 update. The August 2025 version found a 13% relative decline in the most AI-exposed jobs, and the authors say the gap has widened steadily since. The declines are concentrated in jobs where AI mostly substitutes for tasks people did. The gap comes mostly from fewer young workers being hired, not more being let go. The authors call their findings early, descriptive indicators, not proof that AI caused the drop.

The New York Fed puts unemployment for recent college graduates, ages 22 to 27, at about 5.6% in the second quarter of 2026, and underemployment at 42%. Underemployment means a college graduate working in a job that typically does not require a college degree.

We read those numbers as a door that opens less often, not one slamming shut. So the question is how the next hires learn the job once they get in.

Julia McCoy notes in a September 7 video on marketing with Claude, an AI model, that the research to brief step she demonstrates used to take up a junior marketer's whole week. In our view, that tedious week was the training: it is where a junior learned which numbers look wrong, which client always changes the brief, and what a finished job looks like.

Nate B Jones asks the harder question in a September 6 video: if AI does the junior work, where does a junior person pick up the judgment to become senior? He also wonders whether managing an AI agent becomes a junior skill in its own right. We think it does, and that the managing is mostly checking.

After Jake Moffatt's grandmother died in 2022, Air Canada's chatbot said a bereavement fare could be claimed after traveling. The airline's own policy page said it could not. Air Canada argued it wasn't liable for its chatbot, and in February 2024 a British Columbia tribunal rejected that: "It should be obvious to Air Canada that it is responsible for all the information on its website." The award was C$812.02. The sum is small; the point is who paid it.

In Mata v. Avianca, a federal lawsuit in New York against the airline Avianca, two lawyers filed a brief citing court opinions ChatGPT had invented, and kept standing by them after the court questioned them. In June 2023 the lawyers and their firm got one joint $5,000 penalty and were ordered to write to their client and to every judge falsely named as the author of a fake opinion.

In May 2025, a summer insert in the Chicago Sun-Times ran a reading list of 15 books, and 10 of them didn't exist. The syndicator that produced it dropped the freelance writer, who wrote: "I'll take responsibility and again, will blame my AI naivety and lack of oversight."

Our read: in none of these cases was the tool held responsible. An airline, two lawyers and their firm, and a freelancer were. Whoever puts their name on the work owns the mistake, which is why checking is real work, not busywork. If a landscaping company's chatbot tells a customer estimates are free when the owner charges $75 for one, the owner either honors it or argues with a customer holding a screenshot. Air Canada argued, and lost.

These are our examples, not research findings. In each job, checking means going back to the source, not rereading the draft.

None of these takes a week. Our read is that it teaches the same thing the week of grunt work did, faster, as long as the junior has to explain what they found and put their name on the result.

Nate B Jones argues in his AI Bottleneck video from September 27 against making someone early in their career build a spreadsheet by hand just to prove they can. He would rather they use AI to build it, use AI to check it, and then explain it one cell at a time.

We agree, and we would push it past spreadsheets. A paralegal explains why each case supports the argument. A dispatcher explains each line of the quote to the customer on the phone. A junior who can't explain a number doesn't understand it yet, whoever produced it. Explaining is also what a manager can actually observe, which makes it easier to judge than a finished draft that may have come straight from the AI.

This is our recommendation, not a finding from the research. A small shop has no training department, so the training has to live inside the daily work.