Employment Effects of AI: The War For Tacit Knowledge

• A Stanford study from August 2025 shows that employment for 22–25-year-olds in AI-exposed jobs has dropped by about 13 percent.
• Experienced workers in the same AI-exposed roles have increased by 6 to 9 percent, indicating that experience and tacit knowledge are being rewarded.
• The shift is evident in hiring, not wages, raising concerns about a diminishing pipeline of future talent.

In 1986, hundreds of canaries were laid off when Britain officially outlawed their use in coal mines. For centuries, canaries had been the backbone of a sophisticated mining safety system, trusted by coal miners not only in Britain but around the world.

Three hundred to six hundred meters below ground, the buildup of harmful gases poses a major danger. Miners relied on canaries because the birds’ tiny lungs and rapid metabolism made them far more sensitive to poisonous gases like carbon monoxide. The canaries served as a living alarm system, often treated like pets by the workers.

In the 1980s, canaries were replaced with new technology, electronic gas detectors (“electronic noses”). The role of the canaries was … disrupted and automated.

“Canaries in a coal mine” has become an allusion or metaphor. It refers to “something that gives an early warning of danger or failure,” as defined in the Cambridge Dictionary.

It’s not by accident that a new study about AI’s impact on employment chose this metaphor for its title. The paper “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” was published on August 26 by Stanford economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen.

It makes six main observations about how AI is reshaping employment in the U.S.

The body of research on how AI has been changing the labor market on a macro level and the workplace on a micro level has been expanding. A Goldman Sachs analysis was able to show that “the labor market for recent college graduates has weakened at a time when the broader labor market has appeared healthy” (Fortune 2025).

 But the Stanford paper brings a new level of robust data and insights to the discussion.

The main finding of the study: AI is replacing entry-level workers whose jobs can be performed by tools like ChatGPT.

You can apply the canary metaphor here in two ways:

• Canaries = young workers in AI-exposed jobs. Their struggles are early warning signs of deeper shifts in the labor market. The miners represent the broader workforce, which may also be affected in the future.
• Canaries = companies that are reducing entry-level hires. They might be failing to replenish the pipeline of future senior talent by undermining the career ladder for junior talent. This could create a gap, potentially harming the entire company and even the economy (miners).

I will explain in this article why I find the latter interpretation compelling. However, quoting from the paper, it seems the authors had the first interpretation in mind when creating the title:

“Historically, technologies have affected different tasks, occupations, and industries in different ways, replacing work in some, augmenting others, and transforming still others. These heterogeneous effects suggest that there may be ‘canaries in the coal mine’ which are harbingers of more widespread effects of AI” (Brynjolfsson, Chandar, & Chen, 2025, p. 2).

In this article, I will outline the study’s methodology, summarize the main findings, and discuss possible implications for individuals and society.

Methodology

The study observes millions of workers, compares entry-level and experienced employees within the same firms, examines various definitions of AI exposure, and stress-tests the results against alternative explanations. That’s why the authors argue that the decline of junior positions in AI-exposed jobs is real and driven by AI, not just a coincidence.

The researchers pulled payroll records from ADP, the largest payroll company in America (processing paychecks for 25 million U.S. workers), from 2021 to 2025. They tracked specific workers over time at specific companies, not just aggregate statistics. They limited the analysis to firms that remained with ADP for the entire period to avoid distorting effects from companies joining or leaving the platform.

This dataset is powerful because it is payroll-based, not survey-based. That means:

• It avoids self-reporting errors.
• It captures actual headcounts and wages across a large sample.
• It allows tracking of employment by occupation and age group.

Furthermore, they measured AI exposure to assess how likely an occupation is to be affected by tools like ChatGPT. To do this, they built on the work of Eloundou et al. (2024), who had GPT-4 and humans evaluate which job tasks could be automated. They also used the Anthropic Economic Index, which analyzed millions of Claude conversations to see which occupations people were attempting to automate versus augment with AI. The distinction between automating and augmenting is important here (more about that later).

The authors divided all occupations into five equal groups from least to most exposed. For example, software developers and accountants are highly exposed, while plumbers and nurses are less exposed.

Without going into detail here, the authors also ensured they controlled for other shocks, such as changing interest rates or sector-related slowdowns. They compared entry-level workers versus experienced workers within the same firm at the same time. For instance, at Company A in June 2023, were entry-level workers declining while experienced workers were growing?

Of course, the paper cannot prove causality in the same way a randomized controlled trial could, but the authors make a strong case for the correlation as a causal story.

Main Findings

In this chapter, I quickly summarize the six main findings of the Stanford paper.

Entry-Level Employment Declines in AI-Exposed Jobs

The ADP payroll data shows that young workers (ages 22 to 25) in AI-exposed jobs experienced a 13% decline in employment since late 2022, while older workers in the same jobs continued to grow. 

In contrast, employment for workers in less exposed fields and for more experienced workers in the same occupations has remained stable or continued to increase.

The impact was most severe in jobs such as software development and customer service. By 2025, employment for junior software developers had declined by nearly 20% compared to its peak in late 2022.

Meanwhile, jobs with low AI exposure (such as nursing and plumbing) actually saw faster growth for young workers than for their older counterparts.

The Job Market Is Strong, But Not for the Young

Total employment in the US keeps climbing, and experienced workers are doing well. However, since late 2022, the numbers for entry-level workers have fallen. The gap between young and older workers is most evident in AI-exposed jobs.

Automation vs. Augmentation

Where AI automates, young workers lose jobs. Where AI augments jobs, the data shows employment growth for young workers: “While we find employment declines for young workers in occupations where AI primarily automates work, we find employment growth in occupations in which AI use is most augmentative” (Brynjolfsson, Chandar, & Chen, 2025, p. 3).

AI affects jobs in two ways. It can replace people (automation) or help them perform their jobs more effectively (augmentation). 

It is Not Just Hiring Freezes

The decline in entry-level positions is not solely due to firms slowing down overall. The study compares entry-level and experienced workers within the same firms. Even when mid-career workers and seniors increased, the number of young workers decreased. This indicates that the change results from how firms utilize workers in AI-exposed jobs, rather than from a broader slowdown.

Employment, Not Wages

Wages have not fallen. The adjustments come from employment levels, not salaries. Entry-level positions are being hired less frequently, but those who remain in their positions are not earning less than before. Companies are cutting positions without reducing pay.

The Pattern Holds

As mentioned in the methodology section, the authors checked whether other factors could explain the decline of entry-level workers. For example, they excluded remote jobs. They also ruled out tech firms that might be impacted by higher interest rates after 2022. They even examined potential COVID-era education effects.

The result remained consistent. The decline only begins after late 2022, when generative AI became widespread. That timing strengthens the link to AI adoption.

Questions and Implications

In this chapter, I will address possible implications, a couple of which are also raised by the authors themselves.

Rational Adaptation or Careless FOMO?

This point revisits the question from the introduction: Who is the canary and who is the miner?

In other words, are companies rationally adapting to an exogenous shock, or are they irrationally succumbing to the ChatGPT trend due to FOMO?

By neglecting their future senior talent, companies might be setting themselves up for failure. To stay with the metaphor: By cutting young workers, they are effectively “killing their own canary,” destroying their future talent pipeline.

Right now, experienced workers are thriving. However, without entry-level positions feeding into the pipeline, firms may face a future shortage of experienced workers.

In this sense, companies are not responding rationally to AI; they’re “poisoning their own mine.” If firms over-adopt AI for entry-level tasks, society may end up with a missing generation of skilled professionals (engineers, accountants, managers). The warning is less about “everyone’s about to die” and more about long-term self-sabotage.

By automating away entry-level jobs too quickly, firms may be reacting to hype rather than engaging in long-term planning. It reminds me of the over-outsourcing in the 2000s: many firms saved a lot of money in the short term but lost in-house expertise in the long term.

Seek Augmentation

AI reduces jobs when it can automate tasks. It creates jobs when it augments human work (also for entry-level positions). At least, that’s what the study suggests. Following this, it would help younger workers to demonstrate how AI, as a tool, supports their work. The skill is not just “knowing AI exists” but being able to show concretely how they use AI to do more, not less.

• Young workers in automation-heavy roles risk being pushed out before they can climb the ladder.
• Young workers in augmentation-heavy roles can still thrive if they can demonstrate that AI enhances their value rather than replaces them.

That means choosing roles where human judgment, tacit knowledge, or interpersonal skills still matter, and then learning how to use AI tools to extend their own output.

Of course, this doesn’t guarantee safety, and the study is careful not to predict the future. But based on the data, I’d argue that young workers, at least in some fields, will have to prove that they think “AI first” and how it adds value.

Knowledge ≠ Knowledge

The authors argue that one main difference between young workers and experienced workers is the tacit knowledge versus the codified knowledge (“book learning”).

• Codified knowledge: Tasks that can be written down, standardized, and turned into explicit instructions. This includes coding, accounting entries, or customer service scripts. AI excels at these tasks, making it easier to automate junior roles that rely heavily on codified tasks.
• Tacit knowledge: This is know-how that comes from experience, judgment, and human interaction. It’s harder to document in a manual or incorporate into a model. Roles that rely more on tacit skills (managing teams, providing health care, or coordinating complex operations) are much harder for AI to replace. In these jobs, AI tends to augment rather than replace.

The study links the jobs of more experienced workers to greater resilience because of their “tacit knowledge,” which AI can’t fully replicate.  Meaning: If your work is mostly codified, you need to demonstrate how you can use AI to achieve better results, rather than just being replaced by it.

Build the skills that AI cannot easily codify – judgment, leadership, context, and creativity.

Bottom Line

In 1986, Britain retired its last canaries from the coal mines and replaced them with technology. Today’s canaries are fresh out of college, while AI takes over tasks. Forty years after the canaries lost their role, the danger is not invisible but is expanding in plain sight.

This is according to a Stanford study from August 2025, which shows that the employment of entry-level workers has decreased due to AI, while the number of experienced workers remains stable or even grows.

The lesson is twofold. For workers, the skill lies in demonstrating how AI makes you stronger rather than replaceable. For companies, the risk is cutting too deep at the bottom and starving the future pipeline of tacit knowledge.

The danger is not an immediate collapse but a slow erosion that leaves firms without the talent they need in the future.

Literature

• blog.scienceandindustrymuseum.org.uk/canary-resuscitator/
• Brynjolfsson, E., Chandar, B., & Chen, R. (2025, August 26). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab. digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
• dictionary.cambridge.org/dictionary/english/canary-in-a-coalmine
• Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science. https://doi.org/10.1126/science.adj0998
• fortune.com/2025/07/14/how-easy-to-find-job-college-graduate-safety-premium-gen-z/
• smithsonianmag.com/smart-news/what-happened-canary-coal-mine-story-how-real-life-animal-helper-became-just-metaphor-180961570/