Looking for for evidence of AI impacts in the age structure of occupations: Nothing yet

TL;DR:

Using data on occupational employment by age group, I look at the share of young people (age <25 and 25-34) within AI-exposed occupations. For one story about how AI automation could progress, young-age shares should drop as AI disproportionately automates entry-level tasks. I didn’t find any evidence that this had started happening yet as of 2024^[1].

Background

As AI systems move into the realm of being able to take on substantive tasks, I’m interested in identifying signals that might show early signs of AI having an effect in the economy. If early evidence can be found, it can then be used to help change the minds of AI skeptics in the policy community, and hopefully spur a more forward-looking policy response.

One such metric arises from the corporate pyramid argument presented by Luke Drago. In a nutshell: AI will start by automating simpler, more clearly defined, entry-level tasks in a given field^[2]. At the same time, higher-level workers doing AI-complementary tasks will become more productive. In response, firms will hire fewer (or maybe fire) low-level workers, while working harder to retain more experienced talent.

At the macro level, we can use age as a proxy for work experience, resulting in a concrete prediction: as AI gets adopted in a given occupation, we should expect the fraction of younger people in that occupation to decline. In this post, I’m presenting measures that can be used to test this prediction, and watch for it arising over time.

Data and Methodology

I use aggregate data on employment by occupation and age group for the US, available from the Bureau of Labour Statistics. These are derived from the Current Population Survey(CPS), as are most other employment statistics that get reported (this is referred to as the Household Survey in some contexts). Unfortunately, at the age-group and occupation level I need for this analysis, these aggregates are only given at an annual frequency^[3].

To define AI-exposed occupations, I use results from Elondou et al (2023). They don’t share a full dataset of their results, but Table 4 in the paper lists the most exposed occupations according to each of the methods they use. I take all of the occupations listed in their Table 4 and use them as my set of AI-exposed occupations.

The O*NET database used by Elondou et al is based on SOC codes, which in some cases are finer-grained than the Census codes used in the CPS^[4] . In cases where the listed occupation was too fine-grained, I took the census occupation grouping which contained it.

The resulting list of “AI Exposed Occupations” I used is shown in the table below. Note also that a handful of them only go back to 2020.

Occupation	Data available starting in
Tax preparers	2014
Other financial specialists	2014
Web or digital interface designers	2020
Computer occupations, all other	2014
Mathematicians	2014
Other mathematical science occupations	2014
Agricultural and food scientists	2014
Environmental scientists and specialists, including health	2020
Survey researchers	2014
News analysts, reporters, and journalists	2014
Public relations specialists	2014
Writers and authors	2014
Interpreters and translators	2020
Court reporters and simultaneous captioners	2020
Correspondence clerks	2014
Legal secretaries and administrative assistants	2020
Proofreaders and copy markers	2014

Results

As seen in the charts below, there has not been any significant reduction in the share of younger age groups in AI-exposed occupations as of 2024. This isn’t particularly surprising, given that AI agents have only really been around starting in 2025. I’m interested to see how these series evolve over the next few years.

We can also look at single occupations. Here I’m showing all of the occupations that have total employment of at least 40,000 in all periods. With smaller occupations, age shares get too noisy due to sampling variation. On most occupations, there doesn’t seem to be any shift in age composition, with the exception of “News analysts, reporters and journalists”, who have seen a shift down since 2021/2022.

Conclusions and Next Steps

Overall, this analysis, perhaps unsurprisingly, doesn’t find major signs of labour-market disruptions from AI. Given the richness of the CPS dataset, a lot more could be done to expand on these measures and find other metrics to watch. First of all, this approach could be extended to become more sensitive to changes by calculating higher-frequency versions of these measures from the microdata, as mentioned above. Similar analysis could also be extended to industry, geography, or other characteristics of workers. If you are interested in working on this topic, please don’t hesitate to reach out

^
Unfortunately, aggregates released by the BLS only provide annual estimates of employment by age group and occupation. It should be feasible to calculate higher-frequency metrics from public use microdata files, at least for larger occupations. I might do this next year once there’s been more time for effects to have started happening.
^
For example, common prompt engineering guidance is to treat the AI as a smart intern or research assistant.
^
Future work could take advantage of the CPS public use microdata files to calculate these at a higher frequency for larger occupations. For microdata, harmonized extracts are available from IPUMS.
^
This document shows the correspondence between Census and SOC codes