Speaking after Labor’s election victory, Treasurer Jim Chalmers said one of the government’s priorities for a second term was “to do more to embrace technology, particularly the AI opportunity”.
Australia has a mountain to climb
KPMG and the University of Melbourne surveyed more than 48,000 people in 47 countries on AI use and trust. Australia performed poorly across the board:
Over 60% of Australian respondents reported low knowledge of AI, compared to 48% globally and under 40% in Norway and under 45% in Italy and Israel.
Only 24% of Australian respondents had received training in AI, compared to 39% globally, 64% in China, 45% in Singapore and 60% in the UAE.
While 83% of respondents globally are interested in learning more about AI, 41% of Australians reported no or low interest in learning more about AI, ranking alongside Finland at the bottom.
Employer adoption of AI in the workplace jumped from 23% in 2022 to 65% in 2024 and over 50% reported improved efficiency, compared to 67% globally. However, 57% of Australian employees said they used AI output without verifying it and just over half candidly acknowledged they passed off AI output as their own work.
While over 70% of Australian employees use generative AI tools, only 30% say their employer has a policy on generative AI use.
Only 55% of Australian respondents reported experiencing benefits from AI compared to 73% globally and only 30% of Australians believe the benefits outweigh the risks, the lowest ranking of any country.
What’s the size of the prize?
The McKinsey Global Institute forecasts that generative AI could boost average annual GDP growth in advanced economies by between 1.5 to 3.4 percentage points over the coming decade. For comparison, Australia’s GDP grew by 1.3% over the year to December 2024.
Recent evidence suggests AI will have a modest impact on existing businesses. For example, the 2024 Stanford Global AI Index reported that while 48% of businesses reported cost savings in service operations from AI, nearly 60% reported those cost savings were 10% or less. Similarly, while 71% of surveyed businesses reported revenue increases through use of AI in sales and marketing, 60% of those revenue increases were 5% or less.
MIT Professor Daron Acemoglu published a widely debated study concluding that the macroeconomic effects of AI are “nontrivial but modest”.
Acemoglu identifies four possible ways AI may improve productivity:
Automation of human tasks: for example, large language models (LLMs) handling simple writing, translation and classification tasks or computer vision technologies taking over image recognition and classification tasks.
Task complementarity: AI can generate new task complementarities, raising the productivity of humans by better providing information to workers, such as diagnostic suggestions to health workers. Alternatively, AI could undertake some sub-tasks in a larger job, allowing the human to specialise in other sub-tasks, improving their skills and performance levels. Acemoglu observes “the new AI technologies would perform some of these subtasks and do so with sufficiently high productivity, so the subtask-level displacement would be weaker than the productivity gains, expanding the demand for labor and the marginal productivity of labor in these tasks”.
Automation deepening: for example, improved automated control of inventories.
Creation of new, labour-intensive products or tasks: including both good and bad (such as enhanced cyberattacks and deepfakes) tasks. Pointing to the impacts of social media, Acemoglu suggests that the negative effects from new bad tasks enabled by AI could be sizable.
Acemoglu argues that material AI productivity gains observed so far are from easy-to-learn tasks well suited to automation, but these gains cannot be translated across the whole economy because it also includes many hard-to-learn tasks.
An easy task is defined by two characteristics: there is a reliable, observable outcome metric and there is a simple (low-dimensional) mapping between action and the outcome metric. Acemoglu gives the following examples:
How to boil an egg (or providing instructions for boiling an egg), the verification of the identity of somebody locked out of a system or the composition of some well-known programming subroutines are easy tasks. The desired outcome—an egg that is boiled to the desired level, allowing only authorised people to access the system or whether the subroutine works or not—is clear. …With reliable, objective measures of success (well boiled egg, no security breach given the ground truth of authorised people or a subroutine that does not crash), AI models can learn to perform well in a relatively straightforward manner.
Writing is an easy task for LLMs given they are trained on vast amounts of writing. LLMs can deliver productivity gains in two ways:
Because AI models perform writing tasks more or less at the same level as expert workers, their use by lower skilled workers will improve their performance.
Because AI models are, in effect, vast correlation machines, they may discover action combinations that were not typically tried or known even by expert humans (although of course there are the costs of hallucinations).
By contrast, in hard tasks, “what leads to the desired outcome in a given problem is typically not known and strongly depends on contextual factors or the number of relevant contexts may be vast or new problem-solving may be required”. There is typically not enough information for the AI system to learn or it is unclear exactly what needs to be learned. In other words, correlation and probability calculations, even conducted within a vast data pool equalling a third or more of total content on the internet, cannot mimic or replicate human judgement.
AI can still learn about hard tasks from observing human experts, but because the outcome depends on many factors and a qualitative assessment by the human, there is no clear metric of success the AI can mathematically capture in its parameters. The AI ends up performing more or less on a par with expert humans, so there is limited scope for large productivity gains compared to easy tasks.
Acemoglu gives the following example of a hard task:
Diagnosing the cause of a persistent cough and proposing a course of treatment is a hard problem. There are many complex interactions between past events that may be the cause of the lingering cough and many rare conditions that should be considered.
Some developers of sophisticated medical AI models might object to this example, but the basic logic seems to hold given the current limitations of AI.
Acemoglu refers to earlier studies which find cost savings from AI of 14-27% of labour costs, which were (in his view) mainly on easy tasks. For the purposes of his calculations, he assumes that productivity gains should be discounted by 25% for the hard tasks factor, though he seems to consider that to be a conservative estimate.
When he runs these cost savings through his modelling to estimate GDP figures, he forecasts an increase in total factor productivity (TFP) over 10 years with no more than a 0.66% and possibly as low as 0.53% (applying his adjustment for hard tasks).
It has been argued that AI, in contrast with previous technology generations, will hit mid-level workers more than lower skilled workers because AI primarily displaces simple writing and administrative recording tasks rather than manual labour, while AI will be a tool for higher skilled workers.
Acemoglu sees it differently: yes, AI is different from previous technologies, but this is because AI exposed tasks (for example writing emails) are more equally distributed within the population than tasks exposed to pre-AI automation. Most workers also have a mix of easy tasks and hard tasks in their jobs, including because of previous waves of technology:
Many production workers today, including electricians, repair workers, plumbers, nurses, educators, clerical workers and increasingly many blue-collar workers in factories, are engaged in problem-solving tasks. These tasks require real-time, context-dependent and reliable information. For instance, an electrician dealing with the malfunctioning of advanced equipment or a short-circuit on the electricity grid will be hampered from solving these problems because he or she does not have sufficient expertise and the appropriate information for troubleshooting. Reliable information that can be provided quickly by generative AI tools can lead to significant improvements in productivity.
Consequently, in his view, predicted wage impacts do not appear to have a big impact on inequality between education groups.
The good news is the inequality consequences of AI will not be as adverse as pre-AI automation. However, his modelling shows that GDP increases substantially more than average wages and as a result, the capital share of national income increases over labour. Women with lower levels of education will do more poorly than men.
Conclusion
Acemoglu’s paper has been hotly debated. Some argue that he seriously underestimated the productivity gains to be realised.
From AI deepening automation:
“This kind of productivity improvement can have huge growth effects. The second industrial revolution was mostly 'deepening automation' growth. Electricity, machine tools and Bessemer steel improved already automated processes, leading to the fastest rate of economic growth the US has ever seen.”
From AI enabling entirely new tasks we (or AI Models) have not yet conceived:
“There is zero argument or evidence given for why we should expect the harms from AI to be similar to those from social media or why we should expect new bad tasks to outnumber and outweigh new good ones.”
Acemoglu’s model places a strong emphasis on quantifying the value of process innovation, that is, improving efficiency with existing resources, rather than focusing on the product innovation of AI itself, such as the creation of new markets and meeting emerging needs. Further, AI markets are driving significant externalities such as investments in better energy generation, data storage and fibre transmission – transforming other sectors. In its wake, all boats will rise with the tide.
Conversely, other critics have argued that Acemoglu underestimated the costs of AI because he failed to account for its unique risks to humanity.
AI could shift intelligence from humans to machines, creating a different kind of productivity dynamic—one that may reduce human cognitive engagement, potentially stunting long-term growth…This phenomenon mirrors the consequences of the industrial revolution, where reliance on machines led to a decrease in physical activity and overall health. As AI continues to play a more prominent role in society, we risk undermining the cognitive capacities of young people, which could hinder their ability to participate in the workforce. The rise of mental health issues, educational dropout rates and youth unemployment are symptoms of this larger societal issue, which Acemoglu’s model does not fully address.
However, it would be wrong to view Acemoglu’s analysis as suggesting that AI will only ever have a moderate impact on productivity. Rather, he argues that rather than focusing on developing ever more human-like conversational tools, there needs to be a fundamental reorientation of the purpose and architecture of generative AI models in order to focus on reliable information that can increase the marginal productivity of different kinds of workers. As he puts it:
It remains an open question whether we need foundation models (or the current kind of LLMs) that can engage in human-like conversations and write Shakespearean sonnets if what we want is reliable information useful for educators, healthcare professionals, electricians, plumbers and other craft workers.
While Australia will continue to be a taker of AI technology as it was with earlier generations of technology, the difference this time is that we have scope onshore to adapt, modify and fine-tune these systems – something not possible with previous, rigidly programmed generations of technology. We also can gather, pool and manage our own localised, safe and culturally resonant training databases, as South Korea has done.
Peter Waters
Consultant
