Canaries in the Coal Mine
A new paper out of Stanford University, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, has arrived at a telling moment.
We are still dealing with the fallout from a recent Forbes article built on what I consider a deeply flawed MIT NANDA study. That report sparked headlines claiming 95% of enterprise AI pilots fail, though the finding actually concerned a lack of measurable financial return in its specific sample rather than a universal failure rate across all initiatives. Despite being thoroughly picked apart, the narrative has stuck around. It feeds into a broader mood of market nervousness driven by uncertainty over US Federal Reserve interest rate decisions, heavy concentration in tech markets, and Sam Altman’s own public musings that we might be in an AI bubble.
At the same time, there has been a steady stream of reports pointing to AI-linked layoffs and hiring freezes, particularly at the entry level. If companies are already using these tools to substitute for human labour, even if only for repetitive junior knowledge work, then the technology is clearly doing something useful in practice.
Until now, proving that link has been difficult. The data has been muddied by overlapping shocks: pandemic labour distortions, broader tech layoffs, and the shift to remote work. Earlier analyses struggled to separate an actual AI signal from the wider post-pandemic hangover.
The Stanford paper gets a much cleaner readout by using ADP payroll data covering millions of monthly records across different firms, occupations, and age groups. The authors cross-referenced those payroll records against two distinct measures of AI exposure:
An existing GPT-4 exposure index, which estimates the share of tasks within a given occupation that GPT-4 could plausibly perform.
A new Anthropic usage index derived from millions of real Claude queries, categorising occupations by whether users were leaning on AI to automate work (replacing human effort) or to augment it (supporting the human in the loop).
Matching payroll records to real-world usage gives us a much sturdier framework for testing what automation is actually doing to headcount. The standout finding from the August 2025 analysis is that since late 2022, workers aged 22 to 25 in the most AI-exposed occupations have seen a 13% relative decline in employment. Meanwhile, older, more experienced workers within the very same firms and exposed occupations saw their employment hold steady or even grow, as did workers in less-exposed roles. The job losses were heavily concentrated in occupations where AI tends to automate tasks rather than augment the people doing them.
Interestingly, the adjustment is happening through headcount rather than pay. Wage movement was limited; instead of driving down salaries across the board, the entry-level roles simply shrank in number.
It is easy to see the operational logic behind this. The repeatable tasks traditionally handed to juniors, such as basic documentation, coding support, customer service, and data processing, sit right in the sweet spot of what current models handle well. Because early-career staff also have the least bargaining power, many firms are choosing to lean on software rather than invest the time and money required to train up new recruits.
Because the study’s design controls for firm-level shocks and other confounders like interest rates, remote-work shifts, and general tech layoffs, it strips away much of the noise that dogged earlier research. While an observational study cannot rule out every conceivable alternative explanation or prove that AI caused every single lost job, none of those other factors accounted for the age-and-exposure pattern nearly as convincingly.
For anyone trying to start a career right now, this is grim reading. We have all heard plenty of anecdotal stories about graduates struggling to land a first job, and this data offers strong evidence that the bottom rung of the traditional career ladder is being sawn off.
For businesses, cutting junior intake makes short-term financial sense, but it creates an obvious pipeline problem. If you stop hiring and training juniors today, where do your senior practitioners and leaders come from in ten years' time? We may well end up seeing tomorrow's leadership cohorts drawn increasingly from practical or vocational fields that are much harder to automate.
Although Stanford’s numbers come from the US payroll system, the policy problem is international. The IMF looked at this last year; it advised against bringing in a dedicated "AI tax", but suggested governments reconsider corporate tax incentives for hardware and software that actively encourage rapid, labour-displacing automation. Tweaking capital allowances will only slow things down at the margins, but it shows policymakers are at least recognising the problem.
After years of pandemic distortion and noisy tech-sector corrections, we finally have decent empirical data on what generative AI is doing to the workforce. The impact is real, and it is falling squarely on the people at the very start of their careers.