AI Cuts Entry-Level Jobs 16%, Spares Experienced Workers

💡 TL;DR - The 30 Seconds Version

📉 Stanford researchers found employment for workers aged 22-25 dropped 16% in AI-exposed jobs since late 2022, with young software developers hit hardest at nearly 20% decline.

👴 Workers over 35 in identical roles saw stable or growing employment, revealing AI targets codified knowledge from school rather than tacit skills from experience.

🤖 Companies using AI to replace workers cut more entry-level jobs, while firms using AI to augment human capabilities actually increased hiring in some cases.

📊 The study analyzed millions of real payroll records from ADP through mid-2025, providing the first large-scale evidence of AI's employment impact beyond surveys or job postings.

⚠️ The findings create a training paradox: if AI eliminates entry-level roles that teach job skills, companies risk losing institutional knowledge as current experts retire.

🔍 Researchers plan real-time employment tracking to monitor whether automation spreads to experienced workers as AI capabilities advance beyond routine tasks.

A new analysis of millions of U.S. payroll records finds that generative AI is already reshaping who gets hired in exposed fields. The sharpest hit is to young workers breaking into software development, customer service, and similar roles, while employment for more experienced colleagues is steady or rising.

What’s actually new

Stanford researchers examined anonymized data from ADP covering tens of thousands of employers from late 2022—when ChatGPT kicked off the current AI boom—through mid-2025. They isolated occupations most exposed to automation by large language models and compared outcomes by age and experience.

The headline result is stark: employment for 22- to 25-year-olds in AI-exposed jobs fell by roughly 16% from late 2022 levels. Among entry-level software developers, headcount is down nearly 20% from its peak. The blow is concentrated among those at the very start of their careers. That’s the news.

The experience premium

The same analysis shows little to no damage for older workers in the very same occupations; in many cases, employment continued to grow. Why the split? Current systems excel at codified, textbook knowledge—the kind taught in school and embodied in routine tasks. They struggle with tacit know-how accumulated on the job: navigating messy edge cases, coordinating across teams, and delivering products that meet business needs.

That gap maps cleanly onto career ladders. Seniors keep their footing. Newcomers don’t. It’s a tough asymmetry.

Automation vs. augmentation

The researchers see a second divide inside firms: where AI is used to replace routine work, headcount for young workers falls faster. Where AI is used to assist humans—drafting code, summarizing logs, suggesting responses—hiring is more resilient, and in some cases increases.

The policy implication is straightforward. Tools framed as substitutes shrink entry-level demand. Tools framed as complements can support it. Choice of deployment matters.

Evidence beyond anecdotes

Unlike prior studies that inferred effects from job postings or surveys, this work uses payroll records tied to occupations, ages, and firms. That allows the authors to separate AI exposure from macro noise like higher interest rates, a slowdown in tech hiring, or the rise of remote work. It’s not proof beyond doubt. But it is the clearest signal yet.

Crucially, the wage picture looks different from the hiring picture. Pay hasn’t fallen materially for those who remain; firms appear to be adjusting primarily through fewer entry-level hires rather than lower salaries. That suggests productivity gains from AI are offsetting wage pressure—at least for now. Watch that “for now.”

The training paradox

The findings surface a structural problem employers can’t ignore: if AI strips away the rote work that used to train juniors, where do tomorrow’s seniors come from? Software offers a clean example. Routine coding is the first to automate, but the ability to design systems, reason about trade-offs, and communicate with stakeholders is learned by doing. Remove the on-ramp and the pipeline starves.

Some firms are adapting. They are formalizing apprenticeships, pairing juniors with seniors on “centaur” workflows that require human judgment, and evaluating performance on human-AI collaboration rather than solo output. It’s slower in the short run. It preserves capacity in the long run.

What to watch next

Two questions will shape the next phase. First, does the shock stay contained to entry-level roles, or do advances in multimodal reasoning and tool use push automation up the ladder? Second, do firms move from substitution to augmentation as tooling and management practices mature?

The authors plan a near-real-time “AI labor dashboard” to track hiring and wages by occupation and exposure. That kind of instrumentation would help policymakers and employers adjust training and incentives before damage compounds. Early warning beats late repair.

Limits and caveats

This is a working paper, not yet peer-reviewed. It leans on one dataset, albeit a large one, and on an exposure taxonomy that reasonable people can debate. Payroll records can show who is on staff; they can’t show how much AI each team actually uses or how tasks are reallocated inside roles. Still, triangulation across multiple outlets points to the same pattern: early career workers in automatable roles are taking the hit first, while seniors in those roles are mostly insulated.

Why this matters:

The on-ramp is breaking. If entry-level roles vanish, industries risk a hollow middle in five to ten years as today’s experts retire without replacements.
Choices now set the curve. Deployment and policy that favor augmentation—training, apprenticeships, tax treatment—can blunt the shock and preserve capacity.

❓ Frequently Asked Questions

Q: How is this study different from previous research on AI and jobs?

A: This uses real payroll data from millions of workers rather than surveys or job postings. Previous studies relied on theoretical predictions or small samples. The ADP dataset covers tens of thousands of companies from late 2022 through mid-2025, providing the first large-scale evidence of actual employment changes.

Q: Which specific jobs are getting hit the hardest by AI?

A: Software developers and customer service representatives show the steepest declines for young workers—nearly 20% and 16% respectively since late 2022. Administrative assistants and accountants also face significant drops. Meanwhile, health aides, nursing assistants, and physical labor jobs are growing faster for younger workers.

Q: What's the difference between "codified" and "tacit" knowledge?

A: Codified knowledge is what you learn from books and classes—programming syntax, customer service scripts, accounting procedures. Tacit knowledge comes from experience—knowing which clients are difficult, debugging complex systems, or navigating office politics. AI excels at codified tasks but struggles with tacit ones.

Q: How fast is this happening compared to previous technology disruptions?

A: The speed is unprecedented. Previous technology shifts took decades to reshape employment patterns. This AI impact became measurable within just 2-3 years of ChatGPT's launch in November 2022. Even remote work during COVID took longer to show clear employment effects across industries.

Q: What can companies do to avoid losing their future talent pipeline?

A: Some firms are creating structured apprenticeships, pairing junior workers with seniors on AI-assisted projects, and evaluating employees on human-AI collaboration skills rather than solo work. The key is using AI to augment rather than replace entry-level workers, preserving learning opportunities.