Kookaburras in the Coal Mine?
Cancelling the AI apocalypse
Today I am putting out a new working paper that uses high-frequency, timely Australian administrative data to ask whether AI is already disrupting work.
The answer, at least up to October 2025, is: No.
For a headline summary, see Shane Wright’s write-up of the main results in the The AI jobs apocalypse hasn’t landed in Australia yet
The working paper is available here. But I thought I would summarise the results here:
I find no evidence of large aggregate labour-market disruption as of October 2025. Gross earnings have continued to rise strongly, and the occupation-level relationship between AI exposure and gross-earnings growth is only mildly negative at most.
That conclusion comes from linking three unusually useful Australian data sources. The first is the 2021 Census, which gives occupation codes before the AI shock. The second is Single Touch Payroll, which provides daily employer-reported payroll records for the formal workforce. The third is linked firm data, which allow workers to be connected to employer characteristics such as reported AI use.
That combination matters. Most of the existing evidence on generative AI comes from surveys, online work platforms, or case studies. Those are helpful, but they are not the same as watching an entire national labour market move month by month.
Why Australia is a clean test case
Australia is an unusually useful place to look for early AI disruption.
First, according to Anthropic, Australia is one of the highest per-capita users of AI, especially for workplace tasks. That matters because low adoption would make a null result uninteresting. Australia is not a place where nobody is using the technology.
Second, unlike the United States, Australia has had little of the AI data-centre boom and no frontier model development. That makes it a cleaner setting for studying downstream AI adoption: the labour-market effects of ordinary firms using generative AI are less likely to be mixed up with the labour-demand effects of producing AI itself.
Third, Single Touch Payroll gives near-real-time employer-reported payroll data for the near-universe of Australian employees. It records employment relationships, earnings, job starts, and job changes at high frequency, with only a short lag. That is exactly the kind of data needed to study a technology that diffuses quickly and unevenly.
Fourth, these payroll data can be linked to Census occupation data and firm-level administrative data, including whether firms had prior AI or digital investment. For example, the data allow me to take someone who was working as a software engineer before ChatGPT, track their gross labour earnings through job changes, control for age, and ask whether their employer had pre-existing AI use, all the way to October 2025.
Pay up?
The main result is simple: across occupations, there is basically zero relationship between AI exposure and growth in gross labour earnings since 2022. Highly exposed occupations are not seeing earnings collapse relative to occupations with very low AI exposure.
Across low-, medium-, and high-exposure occupation groups, average pay continues to rise after 2022 and ends up in a fairly tight band by late 2025. Those lines remain very flat regardless of whether the data are broken down by occupation, age, firm size, new workers versus incumbents, or firms’ survey-reported use of digital products. A weighted scatter of occupation-level pay growth against AI exposure is slightly negative, but only modestly so: a 0.1-point increase in exposure is associated with 0.83 percentage points less gross-earnings growth since 2022. Even if that whole gradient were attributed to AI, the implied aggregate effect is tiny: about a 0.03 percentage-point reduction in annualised gross-earnings growth.
There are several possible reasons for that. Firms may be using AI to raise productivity without cutting pay. Adjustment may be happening through hiring instead of incumbent wages. Or the gains and losses may be concentrated in narrower jobs than broad occupation groups can capture.
What about hiring?
Hiring falls after 2022 for the labour market as a whole, so the key question is whether exposed occupations weaken more than that general slowdown would imply. The answer is yes, but the effect is not uniform and the aggregate implication remains small. The main hiring scatter slopes down: a 0.1-point increase in exposure is associated with a 0.91-point lower hiring index relative to 2022. The clearest weak group is a set of clerical and administrative occupations near the top of the AI exposure ranking: accounting clerks, payroll clerks, general clerks, keyboard operators, human resource clerks, information officers, and telemarketers.
So the broad hiring result looks similar to the pay result: there is no large across-the-board collapse in high-exposure occupations. The interesting evidence appears only after zooming in on narrower margins.
These jobs have something in common. They involve drafting, form processing, data entry, scheduling, and routine customer-contact work: exactly the kinds of tasks large language models can plausibly absorb or streamline without requiring a full redesign of the firm.
That does not mean every exposed office job is collapsing. Some highly exposed non-clerical occupations, such as software and applications programmers or multimedia specialists and web developers, do not look like the weakest clerical series. Call or contact centre workers and receptionists sit somewhere in between. The early Australian pattern is therefore narrower than a generic story about “AI hitting white-collar work.” The first canaries look more like back-office clerical roles than broad professional classes.
One of the stronger patterns in the current data is the age split. The negative hiring gradient with respect to AI exposure is much steeper for younger workers than for older workers. That fits a particular mechanism: firms may be slowing recruitment into exposed junior office jobs before they make large adjustments to incumbent staffing or pay.
This is important because it changes the policy and research question. If adjustment begins through entry hiring, headline employment and wage statistics may understate what is happening. A labour market can look fairly stable in aggregate while first-rung jobs quietly become harder to get.
The firm split is interesting too. Workers attached to firms that report AI use show stronger overall pay growth in levels. But within those firms, the more exposed occupations are not the occupations leading the gains. In the cross-occupation comparison, the gross-earnings slope is substantially negative among workers attached to AI-adopting 2022 employers, while it is positive among workers attached to non-adopting employers. In other words, AI adoption at the firm level does not mechanically translate into a wage premium for the most exposed jobs inside those firms.
That result should be treated cautiously, but it is a useful reminder that the incidence of AI may run within firms across occupations, not only between adopting and non-adopting employers.
Have we cancelled the apocalypse?
Clearly AI has not affected the labour market as a large broad-based pay shock. If there is an early labour-market signal from generative AI, it is most visible in hiring, in exposed clerical occupations that often serve as entry points for younger workers, and in how the gains from AI-adopting firms are distributed across occupations inside those firms.
That is exactly why administrative data are so useful here. They make it possible to see adjustment on narrow occupational margins before the story is obvious in aggregate wage data. For a technology that diffuses quickly and unevenly, that is the margin worth watching.