AI is hitting entry-level jobs hardest, Stanford study finds

Employment for workers aged 22 to 25 in the most AI-exposed occupations now sits 19 percent below their peers in fields less touched by AI disruption — a gap that measured just 13 percent one year ago, according to an August 2026 update to a Stanford University study tracking how artificial intelligence is reshaping the labor market. The research, titled "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," draws on one of the most reliable datasets anyone has applied to this question: actual payroll records, not surveys.

What Changed

The paper, authored by Stanford economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, was first published in 2025 with data through mid-year. The August 2026 update revises and extends those findings with fresher numbers — and the picture it paints is noticeably worse than the original. The 13-percent employment gap for early-career workers that the team identified last year has widened to 19 percent in the year since, suggesting this is not a one-time shock that the market is absorbing but a trend that is still accelerating.

The industries seeing the sharpest declines for workers aged 22 to 25 are software engineering and customer service, where entry-level employment has fallen by nearly 20 percent since late 2022. Accounting, auditing, and secretarial and administrative roles also appear in the data as significantly impacted. What makes the pattern striking is the contrast with older workers in the same sectors: their employment in those same roles has held steady or, in some cases, increased. The problem is not that entire industries are contracting. The problem is that the bottom rungs of the ladder are disappearing while the upper floors remain intact.

Fields with lower AI exposure — production supervision, stock clerks, health aides — show no meaningful impact and continue to grow, giving researchers a useful control group against which to measure the AI-exposed occupations. The divide is sharp enough that it rules out broader economic explanations like a general youth employment slowdown.

Multiple corroborating reports confirm the core finding. Open Data Science and VKTR, both covering the earlier 2025 version of the paper, independently flagged the same 13-percent baseline figure and the same sector concentration. The August 2026 update's 19-percent figure represents the first major revision to that baseline and comes from the same research team using the same methodology — not a different group replicating the work.

How It Works

To measure AI exposure, the Stanford team used two complementary tools. The first is a potential labor market impact index developed by prior researchers; the second is the Anthropic Economic Index, which tracks how real workers actually use the Claude model in their daily jobs. Google published a similar report based on Gemini usage patterns around the same time. The combination matters: one measure captures theoretical vulnerability, the other captures observed behavior. Using both reduces the chance that the findings are an artifact of how "AI exposure" was defined.

The data itself comes from ADP, the HR and payroll management company. The researchers used a large anonymized subsample of ADP's high-frequency payroll records across tens of thousands of US firms. Payroll records are significantly harder to manipulate or misremember than survey responses — a worker either received a paycheck or did not — which gives this dataset a credibility edge over most prior AI-employment research, which has leaned heavily on self-reported job postings and worker surveys.

The mechanism the authors propose for why younger workers are disproportionately affected comes down to what Brynjolfsson described in an interview with CBS MoneyWatch as "book learning" — the formally teachable, codified knowledge that people acquire in universities and structured training programs before entering the job market. That is precisely the category of knowledge that large language models encode at scale. A recent graduate applying for a software development or customer support role arrives with skills that overlap heavily with what a well-prompted LLM can now produce. An experienced worker in the same role has accumulated something different: tacit judgment, institutional context, and the ability to handle situations that were never written down anywhere. That kind of knowledge is harder to replicate, which is why employment for older workers in the same occupations is holding.

What It Means for Developers

The automative versus augmentative distinction the researchers draw is the most practically useful thing in this paper for anyone building or adopting AI tools. Automative AI — tools that replace discrete tasks outright — correlates with the employment declines. Augmentative AI — tools that make experienced workers more effective — correlates with flat or growing employment in those same roles. That distinction is not just academic framing. It maps directly onto decisions about how AI gets deployed in software development workflows.

Our read is that this creates a specific kind of pressure on developer tooling vendors. Products that market themselves as replacements for junior developers — autonomous coding agents that close tickets without human review — are building on the automative pattern that the data associates with hiring freezes. Products that position themselves as acceleration tools for experienced engineers are on the augmentative side, where the data shows no negative employment signal. Whether that distinction holds as the tools improve is an open question, but as of the data the Stanford team has in hand through mid-2026, the line is real and measurable.

There is also a pipeline problem worth naming directly. Entry-level roles have historically been where experienced workers get made. A software engineer who never spent two years doing code review, debugging other people's work, and getting things wrong in recoverable ways is a weaker senior engineer five years later. If the hiring data continues on its current trajectory — a 6-percentage-point widening of the gap in a single year — the industry will eventually face a shortage of the experienced workers that augmentative AI depends on to function well. The canary metaphor in the paper's title is apt: the signal is visible now, in a population small enough that most organizations can ignore it, before the downstream effects reach anyone with organizational power to act.

Sources