Designing Work in the Age of AI: What Happens to the People Doing It?
The headlines have been asking a simple question for the past few years:
Is AI going to take our jobs?
The latest Australian evidence suggests the answer is more complicated.
Australia has not experienced mass AI-driven job losses. The overall labour market remains relatively strong, and there is no evidence yet of widespread disruption caused by AI.
That is the good news.
But there is another story underneath it.
The Australian Government’s AI and employment in Australia report finds a small but statistically significant shift in employment growth towards occupations that are less exposed to generative AI since the introduction of ChatGPT.
The effect is modest.
It is also sensitive to how you measure AI exposure.
And that last point may be more important than the headline number.
Because if we are trying to design better work, better support and better transitions for people, we need to know what we are actually measuring.
The story isn’t mass unemployment
Let’s start with what the report does not find.
It does not find evidence that AI has already caused widespread job destruction across Australia.
The report’s analysis uses employment data from the Australian Bureau of Statistics Labour Force Survey, covering 355 four-digit ANZSCO occupations from February 2015 to February 2026.
The researchers treat November 2022 as the point at which the labour market entered the ChatGPT era. They then compare the relationship between occupational AI exposure and employment growth before and after that point.
The result is interesting.
Before ChatGPT, the model found no statistically significant relationship between an occupation’s AI exposure and its employment growth.
After ChatGPT, that relationship became negative.
The core estimate of the change, Δβ, was -0.001418, statistically significant at the 5% level (p=0.018).
The post-treatment coefficient was also negative and statistically significant, while the pre-treatment coefficient was effectively zero.
In plain English:
Occupations with higher potential exposure to AI have been growing more slowly relative to less-exposed occupations than they did before ChatGPT.
That is a signal.
It is not a prediction of mass unemployment.
It is not proof that AI caused every difference in employment growth.
And the report is careful about that distinction. It explicitly describes the model as measuring a change in the relationship between AI exposure and employment growth, rather than a conventional treated-versus-untreated causal effect. pp. 67-70.
A small number can still tell us something important
The report’s event study adds another layer.
The researchers found that the pre-November 2022 estimates were generally small, statistically insignificant and showed no systematic trend.
After ChatGPT, the estimates became increasingly negative in the more recent quarters.
The final event-study estimate was approximately -0.017.
The report interprets this as broadly consistent with an occupation one standard deviation above average AI exposure experiencing employment around 2% lower by the end of the study period than would have been expected from its previous relationship.
That is a very different story from:
AI has destroyed 2% of Australian jobs.
It hasn’t.
The estimate describes a relative employment effect associated with differences in occupational exposure. That distinction matters. pp. 73-74.
And it gives us a more useful question.
Instead of asking:
How many jobs has AI destroyed?
we can ask:
Where is employment growth starting to behave differently, and what might that mean for the people doing that work?
The really interesting part: how do we know who is exposed?
This is where the report gets particularly useful.
There isn’t one universally accepted way to measure AI exposure.
The government researchers test three different approaches:
- JSA automation exposure, based on an assessment of how automatable occupational tasks are using generative AI.
- AIOE, developed by Felten and colleagues, which looks at the relationship between AI technologies and occupational abilities.
- Anthropic observed exposure, based on real-world usage of Claude and the tasks people are using it to perform.
They are measuring related things.
They do not produce the same answers.
And that creates a surprisingly important design problem.
Three measures. Three different pictures.
Look at the occupations that appear at the top of each exposure ranking.
Under the JSA automation measure, the five most exposed occupations include:
- Keyboard Operators
- Telemarketers
- Filing and Registry Clerks
- Human Resource Clerks
- Call or Contact Centre Workers
The AIOE measure produces a very different list:
- Actuaries, Mathematicians and Statisticians
- Purchasing and Supply Logistics Clerks
- Accountants
- Auditors, Company Secretaries and Corporate Treasurers
- Psychologists
Anthropic’s observed exposure measure produces yet another picture:
- Call or Contact Centre Workers
- Keyboard Operators
- Sales Representatives
- Software and Applications Programmers
- ICT Support Technicians.
p. 66.
This isn’t a technical footnote.
It changes who we think is at risk.
And this is where I think the question gets interesting
Imagine you are designing a government program to support workers through AI-related change.
You have a limited budget.
You need to identify occupations that might need support.
So you build an “AI exposure” score.
Seems reasonable.
But which score?
If you use one measurement framework, you might prioritise clerical workers.
Another might put psychologists and highly qualified cognitive occupations much higher on the list.
Another, based on observed use of a particular AI tool, might identify contact-centre workers, programmers and ICT support technicians.
The measurement changes the intervention.
That is a human-centred design problem.
The question isn’t simply whether the score is statistically sophisticated.
It is:
What happens to people when we turn a complex, changing experience of work into a single number?
Exposure isn’t the same thing as replacement
The report makes an important distinction throughout its methodology.
An exposure measure tells us about the potential for AI to affect the tasks within an occupation.
It does not tell us which jobs will disappear.
It does not tell us whether an employer will actually adopt the technology.
It does not tell us how quickly adoption will happen.
And it does not tell us how workers, organisations or regulators will respond.
The JSA exposure scores themselves have limitations. They were based on an earlier generation of large language models. They focus on potential automation rather than actual adoption. They are based on tasks currently undertaken by workers, even though those tasks can change. And they do not fully account for the relative importance or substitutability of individual tasks within an occupation.
pp. 63-64.
This is worth sitting with.
A job isn’t a fixed object.
It is a collection of tasks performed by a person, within a particular organisation, using particular tools, under particular conditions.
Change the tools and the job can change.
But that doesn’t necessarily mean the person disappears.
The measurement problem becomes a design problem
This is why I think page 66 is one of the most useful parts of the report for anyone thinking about the future of work.
The three frameworks disagree substantially about which occupations are most exposed.
That doesn’t mean the research is failing.
It tells us something about the problem.
AI exposure isn’t a single observable characteristic of a job.
Different frameworks capture different dimensions of exposure.
- The JSA measure focuses specifically on automation.
- AIOE considers relationships between AI and occupational abilities.
- Anthropic’s measure draws on actual usage of one AI system.
Each tells us something.
None tells us everything.
The report itself concludes that the alternative measures produce different results and that this gives reason for caution when interpreting the main estimates as definitive. pp. 65-75.
For human-centred design, that should change how we approach intervention.
Instead of asking:
“Which occupations are high risk?”
we might ask:
“Which tasks are changing?”
“Who is experiencing that change?”
“What is becoming easier?”
“What is becoming harder?”
“What work is disappearing?”
“What new work is appearing?”
“Who has the opportunity to adapt?”
“Who doesn’t?”
And perhaps the most important question:
What support do people actually need?
Because the impact of AI may arrive quietly
A worker does not need to be made redundant for their experience of work to change.
- A task can disappear.
- A workflow can be redesigned.
- A team can become smaller.
- Performance expectations can increase.
- A junior worker can lose opportunities to practise basic tasks because AI now performs them.
- A worker can become more productive.
- Or they can become more closely monitored.
- Their job title might remain exactly the same.
Their job might still feel completely different.
The government’s statistical model cannot capture all of these experiences.
That is not a criticism of the model. It is a reminder of what labour-market statistics are designed to do.
The report measures employment across occupations.
It does not tell us what happens inside individual jobs.
And that gap matters.
The report itself notes that its approach does not measure AI adoption at individual firms, and its exposure measures capture potential exposure rather than implemented use.
So what should we pay attention to?
The temptation with AI and employment is to look for one big number.
- How many jobs?
- How many losses?
- How many exposed occupations?
But the Australian evidence suggests we need to watch the direction of change, not just the size of today’s disruption.
The report finds that the relationship between AI exposure and employment growth has changed.
It finds that the change is small.
It finds that the effect becomes more negative in the more recent data.
It also finds that the result changes depending on how AI exposure is measured.
And that combination is much more interesting than a simple “AI is taking jobs” headline.
Because it tells us that we are potentially at the beginning of a transition whose shape is still emerging.
What this means if you’re designing for people
For human-centred designers, policy makers, employers and educators, there is a practical implication.
Don’t design the intervention around the score. Design around the person and the work.
An exposure score can help you identify where to look.
It shouldn’t tell you what to do.
If an occupation scores highly, investigate the tasks.
- Talk to the people doing them.
- Understand what AI is actually being used for.
- Find out which parts of the job people value.
- Identify what they are losing.
- Identify what they are gaining.
Then design the support around that reality.
- That might mean training.
- It might mean redesigning workflows.
- It might mean giving people time to experiment.
- It might mean protecting opportunities for early-career workers to build experience.
- It might mean redesigning roles rather than removing them.
- And sometimes it might mean accepting that some tasks or roles genuinely will decline.
The point is to understand the change before deciding what the response should be.
The question we should be asking next
Australia’s AI employment story is not currently a story of mass unemployment.
It is a story of early signals.
Some occupations appear to be growing more slowly.
The differences are small.
The evidence is still developing.
And even the way we measure exposure changes the story.
That uncertainty is uncomfortable.
But it is also useful.
Because we still have an opportunity to design the transition rather than simply react to its consequences.
The question is no longer just:
“Will AI take my job?”
A more useful question is:
“How is AI changing the work inside my job, and what do I need to do that work well?”
And then there is a bigger question for the people designing workplaces, education and workforce policy:
“How do we make sure people are supported through that change?”
That’s where the conversation needs to go next.
Sources and further reading
Key sections for this article:
- pp. 63-67: How AI exposure is measured, limitations of exposure measures, and comparison of JSA, AIOE and Anthropic approaches.
- pp. 67-70: Statistical model, data, treatment date and interpretation of the AI exposure/employment relationship.
- pp. 70-72: Model assumptions and limitations.
- pp. 73-75: Event study results and alternative exposure measures.
- pp. 76-78: Why the researchers use PPML rather than log-linear OLS.
- pp. 79-83: Robustness tests and sensitivity of the results to different modelling choices.
The report’s wider labour-market monitoring framework is consistent with Jobs and Skills Australia’s role in providing Australian labour-market data and analysis.
The key page to return to: p. 66.
If you only have five minutes, read that page.
Discover more from ElevatePx
Subscribe to get the latest posts sent to your email.
