The conversation about automation impact on jobs has been running for nearly a decade on the fuel of a single 2013 Oxford study, a paper that estimated 47 per cent of American jobs were at high risk of computerisation. That number has been quoted in a thousand headlines. What has been quoted considerably less often is the methodology behind it, specifically the part where the researchers assessed entire occupations rather than the tasks within them, a distinction that turns out to matter quite a lot. New Australian research published in 2025 does something the coverage rarely bothers with. It looks at what automation is actually doing, who it is doing it to, and where the real pressure points are. The picture is messier than the apocalypse. It is also more useful.
The number everyone cites and the study it actually came from
Forty-seven per cent. That is the figure that launched a thousand op-eds. Nearly half of all American jobs at high risk of automation within the next decade or two. If you have read anything about automation impact on jobs in the past ten years, you have almost certainly encountered it.
The number comes from a 2013 Oxford study by Carl Frey and Michael Osborne, “The Future of Employment: How Susceptible are Jobs to Computerisation?”. It is a serious piece of research, not a press release dressed as scholarship. The problem is not the study itself. The problem is the shorthand.
Here is what the methodology actually did. Frey and Osborne assessed 702 occupations and estimated the probability that each one could be automated based on a set of engineering bottlenecks, things like social intelligence, creativity, and manual dexterity in unpredictable environments. If an occupation scored above a certain threshold, the whole occupation went into the “high risk” column.
That is where the shorthand starts doing damage. Occupations are bundles of tasks. A radiologist reads scans, yes, but she also consults with patients, coordinates with surgical teams, and makes judgements in ambiguous situations that rarely fit neatly into a training dataset. Automating the scan-reading does not automate the radiologist. The study, by its own design, could not capture that distinction.
This is not a criticism that invalidates the research. It is a limitation the authors acknowledged. The issue is that almost none of the coverage did.
What augmentation means at the worker level
The honest framing is this: automation does not eliminate jobs cleanly. It eliminates tasks. And most jobs contain a mix of tasks, some of which are more automatable than others.
That sounds reassuring until you sit with it for a moment.
For a worker whose role is roughly half automatable tasks, augmentation means something specific. It means showing up to a job that now requires more of the harder, less routine work, probably alongside more digital tools, while the organisation captures the efficiency gains. The worker’s cognitive load may actually increase. The compensation does not always follow.
The distribution of these costs is where the automation impact on jobs debate gets uncomfortable. Higher-skilled workers tend to absorb technological transitions more readily. They have more transferable skills, better access to retraining, and critically, more financial runway when roles change. Lower-skilled workers in heavily automated roles often have none of those buffers.
OECD research consistently finds that workers without post-secondary qualifications are significantly more likely to be in high-automation-risk occupations. That gap does not close by itself.
So when someone says “AI will augment workers, not replace them,” they are not necessarily wrong. They are, however, describing a process that plays out very differently depending on where you sit in the income distribution. Augmentation for a knowledge worker might mean fewer tedious tasks. Augmentation for a warehouse operative might mean a faster pace of work set by an algorithm, with less margin for judgement and no obvious upward path.
Who bears the real costs
The transition costs of automation don’t fall evenly. That observation is not controversial. The question worth pressing is exactly how unevenly they fall, and on whom.
The pattern is consistent across research: workers in lower-paid occupations face higher automation exposure, have fewer resources to fund retraining, and tend to live in labour markets with narrower alternatives when their role disappears. The person whose job is augmented into extinction doesn’t keep the productivity dividend. That goes elsewhere.
What makes the current wave different is its reach. Previous automation shocks were largely sector-specific and, critically, played out slowly enough that adjacent industries could absorb displaced workers. Neither condition holds as reliably now. The automation impact on jobs has moved well beyond factory floors, into logistics, administrative work, customer-facing roles, and increasingly into entry-level knowledge work, the precise positions that historically opened the path to better-paid work.
The Australian research is useful here precisely because it maps exposure by income bracket rather than by occupation alone. Workers most at risk of displacement are disproportionately concentrated in the lower half of the income distribution, which makes the standard retraining response look considerably thinner than its advocates tend to suggest.
Retraining costs money. It takes time people don’t have. It requires access to options that aren’t evenly distributed. And it assumes the jobs being trained into will still exist by the time the training ends.
The productivity and wage story nobody is reporting
The standard reassurance goes like this: automation raises productivity, productivity raises wages, wages raise living standards, and so the disruption is temporary. History suggests a more complicated relationship between the first step and the rest.
The productivity-wage gap has been widening in most advanced economies for decades. In the United States, productivity roughly doubled between 1979 and 2020 while median hourly wages grew by around 17 per cent. The gains went somewhere. They did not go to the workers whose labour became more productive.
That history matters when evaluating the real automation impact on jobs. If displacement falls hardest on lower-income workers, as the evidence suggests, and if productivity gains accrue disproportionately to capital, then the “technology creates abundance for everyone” argument is carrying more weight than it can support. The distribution question is different from the efficiency question. Treating them as the same question is how comfortable narratives outlive uncomfortable evidence.
The Economic Policy Institute has tracked this decoupling in detail, showing how the link between productivity and typical worker pay weakened significantly from the 1970s onward. None of this is irreversible. Policy shapes who captures the gains from technological change. But that sentence is harder to write than “innovation creates prosperity,” which probably explains its relative absence from the discourse.
What Australia has done, and what it has not
Australia entered this debate early. The 2015 CEDA report that put roughly 40 per cent of Australian jobs in the high-risk category made international headlines and seeded a decade of policy anxiety. What got less attention was the methodology it borrowed from Frey and Osborne’s 2013 study at Oxford, which assessed occupational susceptibility based on task composition rather than actual adoption rates. Whether a job could theoretically be automated is a different question from whether it will be, under what economic conditions, and on what timeline.
More recent Australian research has refined that picture considerably. The uneven distribution of risk is now well established: workers in regional areas, in routine-heavy roles, and without post-secondary qualifications carry disproportionate exposure. The aggregate headline figure obscures that concentration almost completely. Saying “40 per cent of jobs face automation risk” tells you almost nothing useful about labour market policy. Saying “these specific cohorts, in these sectors, face transition risk within this window” tells you quite a lot.
Australia has mapped the automation impact on jobs with reasonable rigour. The policy response has not matched the quality of the diagnosis.
Closing / key takeaways
The automation panic and the automation shrug are both wrong in roughly equal measure. The actual automation impact on jobs is real, specific, and distributed very unevenly. It falls hardest on people who are already least insulated from economic disruption. That is the version of this story that deserves serious attention.
A few things worth taking from the research:
- Aggregate automation risk figures are almost meaningless without demographic breakdown
- Regional workers, lower-qualification workers, and routine-heavy roles carry significantly higher exposure than the headline numbers suggest
- Australia’s diagnostic work is solid; the policy response has not kept pace
- The window for effective transition support is not indefinite
None of this requires catastrophising. It does require treating labour market analysis as something that drives policy, not just headlines.
Frequently Asked Questions
Didn't we already establish that automation would wipe out 47 percent of jobs?
That figure comes from a 2013 Oxford study by Frey and Osborne, and it has had a remarkably long life for a number built on a methodology its own authors later qualified heavily. The study estimated susceptibility at the occupational level, treating an entire job title as automatable if its most routine tasks were. What it did not do was account for the fact that most jobs contain a mix of tasks, that organisations change alongside technology, or that labour markets adapt in ways that are genuinely hard to model. A decade of subsequent research has consistently produced lower, messier, and more sector-specific estimates. The 47 percent figure is not wrong so much as it is incomplete, and the mainstream coverage that ran with it never really explained the difference.
What does the 2025 Australian research actually say?
It says the picture is uneven rather than catastrophic, which is both more reassuring and more useful than the headline version. The research finds meaningful automation exposure concentrated in specific occupations and industries rather than spread uniformly across the workforce. Routine, codifiable tasks in particular sectors face genuine pressure. But the blanket prediction that a near-majority of jobs simply disappear does not hold up to the task-level analysis the newer research applies. What it does surface, and this is the part worth paying attention to, is that exposure is not evenly distributed by income, age, or geography. The people most exposed are often the least positioned to move into whatever comes next.
Which workers are actually most at risk?
The honest answer is that it depends heavily on the intersection of sector, task composition, and individual circumstances. Clerical and administrative roles with high routine-task content show consistent exposure across multiple studies. So do parts of manufacturing, transport logistics, and entry-level data processing. What makes the 2025 findings particularly pointed is the demographic pattern: workers in mid-skill, mid-wage occupations, older workers with highly task-specific experience, and workers in regional areas with limited labour market alternatives face disproportionate transition costs. That is not a reason to catastrophise. It is a reason to take seriously who absorbs the adjustment burden, because right now the answer is largely the people least equipped to do so.
Is it not true that technology always creates more jobs than it destroys?
Historically, yes. The industrial revolution, electrification, and computing all eventually produced more employment than they displaced. The "eventually" is doing a lot of work in that sentence. Transitions take decades, the new jobs often require different skills in different places, and the workers displaced are not always the ones who end up filling the roles that emerge. The pattern holds at the aggregate level over long timeframes. It provides limited comfort to a 55-year-old accounts clerk in a regional town in 2026. The optimistic framing and the pessimistic framing are both technically defensible depending on which timescale and which population you choose to look at. The more useful question is what the transition actually costs and who pays it.
What should someone actually do with this information?
Probably not panic, and probably not ignore it either. If your work is heavy on routine, codifiable tasks, it is worth understanding where that sits in the exposure picture for your specific sector, not because a robot is arriving next Tuesday, but because the direction of travel is reasonably clear. Upskilling matters, but the evidence on which skills transfer well is more limited than the "learn to code" discourse suggests. The more durable play is developing judgment-intensive and interpersonal capabilities that are genuinely harder to automate at scale. At the policy level, the research argues fairly clearly for targeted support mechanisms aimed at high-exposure workers rather than broad reassurances that the market will sort it out. Markets do sort things out. They just do so on their own schedule.
