Stanford’s Canaries Signal Trouble: AI Quietly Shrinks the Entry-Level Ladder

Erik Brynjolfsson saw the pattern first. Last summer his team at Stanford paired payroll records from ADP with measures of AI exposure. The result landed like a warning shot. Employment for workers ages 22 to 25 in the occupations most exposed to generative AI fell sharply relative to their peers. Critics called it noise. Interest rates. Tech-sector correction. Remote work. Anything but the new technology.

They were wrong. Brynjolfsson kept updating the numbers. He deepened the partnership with ADP Research. The latest figures, released through a live dashboard this month, show the gap has not closed. It has widened. Fortune reported the fresh data on June 27. The most AI-exposed occupations contracted 0.2 percent year over year as of April 2026. Least-exposed roles grew 0.1 percent. Those headline numbers hide a deeper split.

For early-career workers the picture turns stark. Employment in highly exposed jobs now shrinks at 3.8 percent per year. Least-exposed jobs for the same age group grow 2 percent annually. The decline sharpened over time. From a 2.8 percent drop through April 2024 it moved past 4 percent in the following period. Mid-career workers ages 31 to 34 also saw contraction of 1.7 percent. Workers 35 to 40 posted 2 percent growth. The technology does not eliminate entire jobs across the board. It removes the bottom rung.

This is no accident of the data. The original Stanford paper, titled “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” laid out the evidence in November 2025. Early-career workers in the most exposed occupations saw a 16 percent relative decline in employment after widespread generative AI adoption. The researchers controlled for firm-level shocks. Experienced workers in the same roles held steady or grew. Less-exposed fields showed no such pattern. Adjustments happened mostly through hiring, not pay cuts. The Stanford Digital Economy Lab published the paper here.

The six facts paint a consistent story. Declines concentrate where AI automates tasks rather than augments people. Software development and customer support illustrate the point. Young software developers saw employment drop nearly 20 percent from late 2022 peaks through July 2025, according to earlier ADP analysis. Young home health aides, whose roles carry low AI exposure, remained unaffected. ADP chief economist Nela Richardson, Brynjolfsson’s partner on the work, emphasizes the distinction. “In the aggregate, AI’s impact on jobs remains modest,” she wrote in a June 16 post. But break it down by career stage and “dramatic differences emerge.”

Richardson has stressed nuance. AI that amplifies human abilities tends to support employment growth. AI that substitutes for routine work does the opposite. Entry-level positions cluster in that second category. Juniors handle the summarization, formatting, basic coding, and data gathering that large language models now perform at scale. Seniors bring judgment, context, and client relationships that resist easy codification. So far the machines have not replaced the experienced workers. They have simply reduced the need to hire their replacements.

The Canaries Dashboard makes this visible in near real time. Built by the Stanford Digital Economy Lab and ADP Research, it draws on payroll information for 4.6 million workers across more than 730 occupations. The sample covers roughly one in six American workers. Data updates monthly. Visualizations track employment indices by AI exposure quintiles, age bands, and specific roles. It matches occupations to exposure scores from earlier research and to automation-augmentation ratios. Launched in June 2026, the tool stands as the lab’s highest-profile tracker of AI’s labor market effects. Brynjolfsson calls it essential. “We are flying blind into one of the most consequential periods in world history,” he said at the launch. “We need timely, trusted evidence.”

Critics have tested the findings repeatedly. Google economists pointed to interest rates. Brynjolfsson checked. The most rate-sensitive occupations, such as construction, show the lowest AI exposure. The pattern runs the opposite direction. Remove the entire tech sector. Exclude remote-friendly roles. Control for firm shocks. The early-career decline in exposed occupations persists. “If you take out the entire tech industry, or take out all tech-related occupations, or you slice it different ways, you still get this effect,” Brynjolfsson told Fortune. The effect has grown roughly half a percentage point per month since the original cutoff.

Recent studies add texture. A Swiss analysis of more than 7.3 million job advertisements found AI adoption linked to fewer junior postings, especially in marketing. Harvard researchers observed that firms using generative AI saw junior employment decline while senior roles held or expanded. These accounts align with the payroll data. They do not replace it. ADP’s sample offers high-frequency, administrative precision that surveys and postings cannot match.

Yet the debate refuses to settle. Daron Acemoglu, the MIT economist and Nobel laureate, remains a prominent skeptic. He and Brynjolfsson have sparred publicly for months over productivity estimates. Acemoglu sees smaller gains. Brynjolfsson expects more. They exchanged messages the morning of Brynjolfsson’s recent interview. Both agree AI should complement rather than replace people. Their disagreement centers on magnitude and speed. “I don’t get how he has such low productivity numbers,” Brynjolfsson said. “Time will tell. Pretty soon we’re going to see who’s right.” Acemoglu has called much of the productivity discourse speculative.

The comparison Brynjolfsson reaches for is the Industrial Revolution. That era automated and augmented muscle. This one targets the mind. He expects the transformation to prove larger and ten times faster. His friendly wager with Northwestern economist Robert Gordon on longbets.com bets that labor productivity growth will exceed 1.8 percent annually through 2030. Brynjolfsson believes he leads already. The J-curve he has described for years suggests gains arrive back-loaded. The early pain falls on those entering the workforce.

Business leaders wrestle with the consequences. Fewer entry-level roles mean thinner talent pipelines. Companies gain short-term efficiency. They risk long-term skill development. Internships and projects that once built experience now compete with AI outputs that arrive faster and cheaper. Schools and universities face pressure to adapt curricula. Some programs already emphasize prompt engineering, AI oversight, and higher-order problem solving from day one. Others lag. The gap between what graduates know and what employers now demand widens.

Richardson has spoken at the World Economic Forum about rethinking early-career development. The traditional ladder compressed. Roles once reserved for learning now demand skills that used to come after years on the job. Workers with AI-related competencies command wage premiums. PwC research cited in recent analyses puts that premium at 56 percent in some cases. The pace of skill change accelerates. Young people sense both opportunity and risk. Surveys from the World Economic Forum show entry-level workers more optimistic about job security than executives, yet uncertainty remains high.

Not every sector or role follows the same script. Marketing and sales have shown relative resilience in some breakdowns. Occupations heavy in codified knowledge, by contrast, contract faster at the junior level. Software engineering offers the clearest example. Employment for young developers fell dramatically after ChatGPT. Older developers in the same firms maintained or increased headcount. The code still gets written. Fewer novices write the first drafts.

Economists continue to argue over causation. Some point to work-from-home patterns as a stronger predictor of reduced junior hiring. Training and supervision prove harder remotely. One study found the remote-work effect dominated when both variables were included. Others maintain the AI signal survives those controls. The Canaries Dashboard will keep producing fresh numbers. Monthly updates should clarify whether the trend reverses, stabilizes, or deepens as agentic systems and newer models roll out.

Brynjolfsson refuses panic. He also rejects complacency. The dashboard exists to replace guesswork with evidence. Its name carries intent. Canaries warned miners of danger without stopping the coal from being dug. They simply told everyone the clock was running. The labor market’s canaries have sung for nearly four years now. The song has not stopped. Employers, educators, and young workers hear it. The question is who acts on the signal before the air grows thinner for the next generation of talent.

Recent coverage reinforces the urgency. A June 2026 analysis from Stern Strategy Group examined the talent pipeline risk and noted that the Stanford findings now rest on actual payroll data rather than forecasts. Stern Strategy Group discussed the implications here. Swiss evidence of falling junior advertisements arrived just days ago, adding international weight to the U.S. payroll patterns. The conversation has moved past whether AI affects entry-level opportunities. It now asks how societies adjust the gateway to professional careers when that gateway narrows.