Canaries in the Coal Mine
I’m writing this in light of the publication of a new paper from Stanford University: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.
As is often the case with AI research, the timing is striking. We’re still feeling the aftershocks of a Forbes article that leaned heavily on a (deeply flawed) MIT study claiming 95% of enterprise AI pilots fail. That claim has been torn apart, yet it lingers thanks to a mix of factors: uncertainty over US Federal Reserve rate decisions, the concentrated structure of tech markets, and Sam Altman’s own musings that we might be in a bubble.
At the same time, a steady drumbeat of reports has raised the possibility of AI-induced layoffs and hiring freezes, with entry-level roles hit hardest. If AI is allowing firms to substitute labour, even just at the more repetitive, junior end of knowledge work, that suggests it’s already proving useful.
The problem is that previous analysis has been blurred by overlapping shocks: pandemic labour distortions, mass tech layoffs, and remote work reshuffles. Earlier reports struggled to isolate an AI signal from all that noise. Were losses really about automation, or just the hangover from pandemic disruption?
The Stanford “Canaries” paper goes further. Using payroll data from ADP incorporating millions of monthly records across firms, occupations, and age groups, it provided a cleaner picture. The researchers matched this data to two measures of “AI exposure”:
A GPT-4 exposure index (from earlier research), estimating how many tasks in a given occupation GPT-4 could plausibly perform.
A novel Anthropic usage index, based on millions of Claude queries, which classified roles by whether AI was more likely to automate (replace humans) or augment (support humans).
This framework allowed the researchers to test hypotheses about automation potential with more rigour than we’ve seen before. The headline finding: since late 2022, early-career workers (aged 22–25) in AI-exposed occupations saw employment drop 13–16%. Meanwhile, older workers in the same firms gained jobs. Losses were concentrated in roles where AI is more likely to automate tasks than augment them.
By contrast, employment for less exposed roles, or more experienced workers in exposed roles, held steady or even grew. Crucially, wages didn’t fall, roles simply vanished, particularly for newcomers.
The results support a clear narrative: entry-level tasks are the most vulnerable. Documentation, coding support, customer service, data processing - repeatable tasks that tend to be delegated to juniors - are precisely the ones AI is swallowing first. And because early-career workers have less bargaining power, firms lean on AI rather than invest in training them.
Importantly, the methodology was robust enough to account for other factors that have clouded the picture before now; tech layoffs, remote work shifts, interest rates, post-pandemic quirks. None explained the pattern as convincingly as AI exposure did.
The implications for young workers are unsettling. This may be the strongest evidence yet of a structural break in traditional career paths. We’ve all heard the stories of graduates struggling to land their first role; the data confirms that AI is weakening the career ladder at its first rung.
For organisations, the strategy is understandable but short-sighted. Where will tomorrow’s leaders come from if fewer juniors are developing the skills to climb higher? It raises the possibility that future leaders may increasingly emerge from roles less exposed to AI automation, many of them more practical or vocational in nature.
Although the study is based on US data, the policy implications are global. The IMF explored options last year, rejecting the idea of an “AI tax” but suggesting governments might reduce corporate tax breaks for computers and software. That would only delay the inevitable, but it signals recognition that policy has a role to play.
What’s clear is that the fog is lifting. After a long period of pandemic disruptions and tech layoffs clouding the analysis, stronger data is emerging. And it tells a story many suspected all along: AI is already reshaping the labour market, starting at the bottom.
