AI adoption and jobs – positive for productivity?

A new academic paper from Stanford uses a large administrative dataset (ADP) to examine employment trends by age and exposure to AI. The researchers find that employment is falling for ‘early career’ workers (defined being age 22-25) in AI-exposed jobs such as software development and customer services. However, while early-career jobs that AI can automate are declining, those where AI can ‘augment’ labour (e.g. maintenance and repair workers) are not.

The researchers’ theory is that young people are struggling because they have ‘textbook’ knowledge that AI can reproduce, while older workers have more ‘on-the-job’ knowledge (e.g. how each firm works) and so less affected. However, they also find that less-educated occupations vulnerable to automation are seeing problems even up to age 40, as the benefits from experience are less relevant in such roles.

The authors also suggest firms could be in an evaluation phase. It’s easier to pause hiring and test AI than fire experienced workers (we note that half of job losses in recessions are typically due to less hiring) so firms could be experimenting with AI. This seems likely to lead to broader job losses down the line.

Overall, the paper concludes that job losses in highly AI-exposed roles explains the stagnation of young-people employment, as shown in the chart below.

For younger workers (higher quintile = more AI-exposed). Source: Standford paper: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.

How does this compare to previous research?

In a previous blog, we evaluated some earlier research, and suggested that the impact of AI on the graduate market may be more gradual than feared. So how does this new paper differ?

For starters, the latest research uses employment data rather than vacancies. We were previously unsure how to interpret unfilled vacancies falling from ‘war-like’ highs following COVID.

Moreover, it has a much bigger dataset. It uses administrative data from ADP on around 3.5 million workers. This is 58x bigger than the 60,000 people sampled in the monthly labour-force survey. This allows for better-quality granular analysis: the labour force survey becomes erratic even just looking at recent graduate unemployment, let alone by occupation.

We suggested a general hiring freeze due to tougher macro conditions would also explain an underperformance of recent-graduate employment. However, the authors control for macro shocks using ‘firm’ and ‘time’ fixed effects (e.g. how an individual construction or nursing-home company behaved as monetary policy tightened from 2022). The authors still find a relative decline of 12% of most-to-least AI exposed youngest workers within each individual firm.

We found similar struggles for both recent college graduates and school leavers. This paper ducks that issue by only focusing on workers aged at least 22. However, it does find less-educated workers are struggling up to age 40 in careers exposed to AI automation.

Overall, this fascinating paper uses a larger dataset to find ‘canaries in the coalmine’ (early-career workers) struggling to find work in areas most exposed to AI ‘automation’. This should, over time, lead to a notable productivity boost - which is much needed given demographic pressures from ageing.