RBA note: Job postings fall steepest in occupations with the highest AI exposure
The Australian central bank's new monitoring tool points to weaker outcomes for young workers in AI-exposed occupations — but the author himself cautions against causal conclusions, and a US analysis dates the problems to a time before ChatGPT.
The Reserve Bank of Australia (RBA) published an Insight note on 9 November 2026, written by Mike Major, which introduces a tool for monitoring the early effects of artificial intelligence on the labour market. According to summaries from LDS, the note finds "tentative signs" of weaker outcomes in AI-exposed occupations, particularly for younger workers, while concluding at the same time that little so far suggests real labour market disruption (LDS).
The gap between the RBA's own wording and the headlines around the note is wide. The Decoder framed the coverage as "AI hitting young workers first" (The Decoder), while the note's own formulations — descriptive analysis, no causal assessment, purely theoretical exposure — are far more restrained. That distance is what this story is really about.
How the Tool Works
The note ranks occupations by how exposed they are to AI, measured through the task composition of each occupation. The benchmark date used is November 2022 — the month ChatGPT brought generative AI into widespread public use. Developments in job advertisements and employment are then compared against this point in time, grouped by exposure level (LDS).
According to the ranking, the most exposed occupations include software developers and clerical and administrative workers. Among the occupations in the most exposed quintile contributing most to the decline in new advertisements at Jobs and Skills Australia since November 2022 are general clerical workers, software and applications programmers, ICT business and systems analysts, customer and contact centre workers, and receptionists (LDS).
What the Numbers Show
According to the analysis as reported by The Decoder, job advertisements have fallen in all exposure groups since ChatGPT launched in late 2022 — but the decline is steepest in the group with the highest exposure. In that group, workers aged 15 to 24 and young candidates have lower employment rates and higher unemployment than older colleagues in the same occupations (The Decoder).
This is what constitutes the "tentative signs": a pattern consistent with a story in which AI hits entry-level jobs first, but one that the note itself does not interpret as evidence of any such thing.
The Caveats Are Substantial
Major explicitly warns against drawing early conclusions. According to The Decoder, the analysis establishes no direct causality, and the exposure measure captures only theoretical potential — not actual AI use in firms, and not real changes in hiring or pay practices. Alternative explanations have not been ruled out: shifts in the post-pandemic business cycle or in the extent of remote work could also be driving the numbers (The Decoder).
At the same time, data from the job search engine SEEK, cited in the coverage, point in a different direction than pure substitution: they suggest AI is complementing work and creating new demand for AI skills (The Decoder).
The note's analytical lens is thus doubly cautious. It measures exposure, not adoption. And it describes correlations in time without claiming cause.
The data basis is also limited: the most recent complete occupational data in the note covers February 2026, with partial data for May 2026. Upcoming changes to the Labour Force Survey and the transition from the ANZSCO occupational classification to OSCA will require a revised mapping between the exposure ranking and the new data (LDS).
It is also worth noting what kind of document this is. The RBA describes the Insight series as notes in which staff present their own analysis, with a disclaimer stating that the views are the author's and should not be read as those of the RBA or the Bank's boards (LDS). The note is therefore not an official central bank position on AI and employment.
The International Evidence Pattern
The Australian note lands in the middle of an ongoing debate in which studies point in partly opposite directions. The figures below are known via The Decoder's summaries of the studies, not from the originals:
- A Stanford study from August 2025, based on millions of US payroll records, found that employment among 22- to 25-year-olds in AI-exposed occupations had fallen 13 percent in relative terms since late 2022 — and by almost 20 percent among young software developers.
- An Anthropic study from March 2026 found signs of weaker hiring among 22- to 25-year-olds in exposed occupations, with the probability of finding a job down around 14 percent. But it found no systematic increase in total unemployment. An internal Anthropic review is reported, according to the coverage, to show that the share of junior technical hires fell from 27 to 9 percent since early 2025 — but this is a company claim, not independently verified.
- An analysis from January 2026, led by Morgan Frank of the University of Pittsburgh and based on 10.6 million LinkedIn profiles and US unemployment data, warns against attributing the entire decline to AI. The research team found that problems in exposed occupations began in early 2022 — months before ChatGPT launched — suggesting macroeconomic factors may be the larger driver.
The Pittsburgh finding is the most important counter-explanation in the evidence, and it cuts hard against all the studies that use the ChatGPT launch as a natural dividing line. If the decline in exposed occupations began before generative AI was in widespread use, the timing alone cannot be used as evidence of an AI effect. That applies in principle to the RBA note as well, which uses November 2022 as its benchmark.
Does that mean the Stanford and Anthropic findings are disproven? No — but it means timing alone does not settle the question. The Stanford study is based on payroll records and finds a pattern in age and occupation consistent with entry-level substitution; the Pittsburgh study is based on profiles and unemployment data and finds a different timeline. The disagreement is partly about method and partly about interpretation, and it is not resolved by what is available.
What Could Settle the Question
The note itself points to some of the levers: a revised mapping of the occupational classification when ANZSCO is replaced by OSCA, and updated data beyond February 2026. Beyond that, it will be decisive whether the time pattern holds as more years of data arrive — whether the decline in advertisements and employment actually breaks around the ChatGPT launch, or continues steadily across the divide as the Pittsburgh analysis suggests. Another open question is whether theoretical exposure actually translates into real AI adoption in Australian firms, something the tool cannot measure today.
Until then, the most precise summary is the note's own: tentative signs of weaker outcomes for young people in AI-exposed occupations, little evidence of real disruption — and no basis for saying AI is the cause.

