The most misread number in the AI debate

Almost every widely circulated statistic about AI and employment is a measure of exposure. Almost every headline treats it as a measure of risk. These are different things, and the researchers who produce the figures say so explicitly in the same documents that the headlines are drawn from.

Exposure asks a narrow technical question: how much of what this occupation currently does could, in principle, be performed or assisted by the technology? It says nothing about whether an employer will adopt the tool, whether adoption saves money, whether the role is redesigned or eliminated, or whether the worker is trained to use it. Those are separate questions with separate answers, and they are the questions that determine what actually happens to a job.

What the ILO measured, and what it said about it

ILO Working Paper 140, published 20 May 2025, is the most detailed global index of occupational exposure to generative AI currently available. It scores occupations across four progressively increasing exposure gradients, built from the 29,753 tasks in the Polish occupational classification, a survey of 1,640 workers yielding 52,558 usable data points, and a panel of national and international experts who validated a representative subsample.

Its findings, in the paper's own terms:

  • One in four jobs globally, 24%, has some degree of exposure to generative AI.
  • In high-income countries that rises to roughly one in three, with total exposure ranging from 11% of employment in low-income countries to 34% in high-income countries.
  • Just 3.3% of global employment falls into the highest exposure gradient.
  • Clerical occupations remain the most exposed category, with data entry clerks and typists at the top of the index.

The paper then does something the coverage of it generally does not. It states the limit of its own measure: such exposure "does not imply the immediate automation of an entire occupation, but rather the potential for a large share of its current tasks to be performed using this technology". It goes further, noting that whether exposure leads to an occupation disappearing "will depend on the initial decision to adopt the technology, but also the extent to which individuals in these occupations are given opportunities to learn to work with these technologies and adapt to the evolving nature of their tasks".

That second clause is the one worth sitting with. In the ILO's own framing, the outcome is partly determined by whether employers invest in capability. It is a variable, not a constant.

The direction of travel in the evidence

One detail rarely reported: the ILO's 2025 estimates are lower than the estimates in its own 2023 paper. Refining the method and validating it against human assessment moved the numbers down. This is worth knowing because the popular impression runs the other way, that each new study finds the threat larger than the last. In this case the most methodologically careful revision available reduced the estimate.

What Australia's national study found

Jobs and Skills Australia published its overarching report, Our Gen AI Transition, on 14 August 2025, with supporting analysis papers on 2 September 2025. It is the first whole-of-labour-market national study of its kind in Australia, and it adapted the ILO's method, scoring each task twice: once for whether generative AI could assist it, and once for whether generative AI could perform it.

Scoring the entire Australian workforce against 998 occupations at the four-digit ANZSCO level produced this distribution:

  • Automation exposure: 79% of the workforce low or very low, 17% medium, 4% high.
  • Augmentation exposure: 13% low or very low, 56% medium, 31% high.
  • The largest single group: 49% of workers, nearly half, sit in occupations that are low automation and medium augmentation.

JSA's reading of that middle group is the important one. Those occupations, it says, would "more likely experience change rather than disruption", because most of their tasks could not be automated while some could be assisted. The study's summary conclusion is that "Gen AI is more likely to augment human work than replace it", and that "the higher potential for automation is concentrated in routine roles".

Exposure is not evenly shared

Averages conceal the part that matters most to individuals, and both studies say so.

The ILO finds a persistent and widening gender pattern: 4.7% of female employment sits in the highest exposure gradient against 2.4% of male employment, and in high-income countries the gap widens to 9.6% against 3.5%. The driver is compositional rather than anything intrinsic to the work: clerical and administrative roles are the most exposed category and are disproportionately held by women in high-income economies.

JSA makes the corresponding point about aggregation directly, warning that "it is important that individuals are not lost in this aggregate picture of the labour market", that some people will be affected more acutely than the averages suggest, and, in the same breath, that "these acute challenges aren't generalised across the board, and to all occupations". Both halves of that sentence are doing work.

One structural finding cuts against the intuitive picture: JSA notes that automation potential is highest for middle-skill occupations, particularly ANZSCO skill level 4, which includes several forms of clerical work, while higher-skilled occupations show more exposure to augmentation. Many clerical tasks that were untouched by previous waves of automation are now within reach, while the occupations most exposed to automation "account for significantly smaller workforce numbers".

What has actually happened so far

Exposure studies are projections of potential. JSA also looked at observed outcomes, and its finding is more restrained than the discourse around it: "Gen AI adoption is reshaping skill demand and roles. Large-scale job displacement is yet to emerge, with current impacts limited to early-adopters."

The study records reduced demand in some narrow, task-specific occupations, giving voice-over work as an example, but concludes that "most observed impacts involve the evolution of roles, upskilling, and redeployment of workers, rather than widespread job loss". On entry-level work, where concern is most concentrated, it found that such roles "may be more likely to transform than diminish", with no current evidence of widespread displacement in Australia, while noting this may partly reflect how early adoption still is.

Why the distinction changes what an organisation should do

If exposure is read as risk, the implied response is defensive: forecast headcount reductions, and treat affected staff as a cost line. If exposure is read as what it is, a measure of how much of a role's task content overlaps with what the technology can do, the implied response is different. High augmentation exposure with low automation exposure, the profile of nearly half the Australian workforce, describes a job whose content changes while the job persists. The organisational task is redesign and capability, not attrition.

That has practical consequences for governance. Exposure scores are a reasonable input to workforce planning and to deciding where training investment goes first. They are not a defensible basis for restructuring decisions about individuals, and using them that way carries real legal exposure under employment and discrimination law, particularly given the gendered concentration the ILO documents. Organisations making decisions that affect people on the basis of AI-related analysis should also understand their obligations around automated decision-making transparency.

The limits of both studies, stated plainly

Neither study is a forecast, and neither claims to be. Some limits worth holding onto:

  • Exposure is not quality. JSA is explicit that its scores "cannot be used to infer the quality with which Gen AI might acquit a task", and that exposure "does not by itself predict job quality or complexity".
  • Task-level scoring is a model. Both studies decompose occupations into tasks and score the tasks. Real jobs contain judgment, relationships and accountability that resist that decomposition.
  • Long-range economic modelling is scenario work. JSA's projections to 2050 are computable general equilibrium modelling against a base case. They illustrate direction and composition under stated assumptions. They are not predictions of what any particular labour market will look like.
  • The technology is moving. JSA notes that its scores "would be expected to shift as Gen AI technologies evolve, especially with more integration of robotics and complementary technologies", and observes separately that agentic AI so far appears most likely to automate tasks that are "structured, predictable, and low risk".

The honest summary is that the best available evidence supports a picture of widespread task change, concentrated automation pressure in routine clerical work, and no observed large-scale displacement to date. It does not support the numbers most often quoted, and the researchers producing it are the first to say so.

Figures in this article are modelled estimates of technical exposure published by the ILO and Jobs and Skills Australia. They are not forecasts of employment outcomes, and they should not be used to make decisions about individual roles or individual people.

Related reading

Sources: ILO Working Paper 140, Generative AI and Jobs: A Refined Global Index of Occupational Exposure (20 May 2025) · Jobs and Skills Australia, Our Gen AI Transition: Implications for Work and Skills (14 August 2025) · Jobs and Skills Australia, Analysis Paper A: Exposure (2 September 2025)