The Hon Andrew Leigh MP
Assistant Minister for Productivity, Competition, Charities and Treasury
Will AI Increase Inequality?
Third JPMorgan AI Conference,
Sydney / Melbourne
Online Address
Tuesday, 13 October 2026
I acknowledge the Ngunnawal people, on whose lands I am speaking with you today, and pay respects to all First Nations people present.
Thank you for the chance to address you today. For my topic, I’ve picked a question that’s been rattling around my head – and perhaps yours – over the past year. ‘Will AI increase inequality?’.
As someone who’s been working on inequality as both a professor and a politician, this strikes me as a question that matters economically.
It’s likely that AI will have substantial aggregate effects. The Australian Treasury (2026) says that ‘AI is the next major general-purpose technology that will push out the global productivity frontier and reshape the global and Australian economies.’
The Productivity Commission (2025) estimates that AI could lift the level of Australia’s multifactor productivity by more than 2 percent over the next decade.
But the question of who gets the benefits is crucial – whether you’re a citizen, a policymaker, or a business leader.
Let’s start with a practical example.
Imagine a customer support agent, two months into the job.
A small business owner appears in the chat window with a technical problem. The agent needs to diagnose what has gone wrong, understand the software, explain the solution clearly and deal with a customer whose patience may already have disappeared.
In front of the agent sits an AI assistant. It has learned from millions of previous conversations, including the responses of the firm’s best workers. It suggests questions to ask and ways to explain the answer.
Economists studied what happened when more than 5,000 customer-support agents were given access to such a tool, using the staggered introduction of the technology as a natural experiment.
Across the workforce, productivity rose by 15 per cent. Among less-skilled and less-experienced workers, the number of problems resolved per hour rose by around 30 per cent.
With AI assistance, an agent who had been in the job for two months performed about as well as an unassisted agent with more than six months’ experience (Brynjolfsson, Li and Raymond 2025).
Four months of experience, compressed into software.
That study gives us a sense of why the question of how artificial intelligence affects inequality isn’t an easy one.
In the past, a pretty solid economic rule of thumb was that new technologies tended to favour people who already had more education or expertise. Generative AI might not be like that. It may give novices a leg up, rather than cause them to fall behind.
Yet the distributional consequences don’t just depend on the productivity of those two workers. It could be that some novices earn more, but the firm hires fewer novices. It might also be the case that prices fall. And it’s possible that the owners of the AI system collect much of the gain.
That is why asking whether AI will increase inequality requires us to distinguish four channels.
There is inequality between workers.
There is the distribution of jobs and career opportunities.
There is the division of income between labour and capital.
And there is the distribution of economic rents between firms and between countries.
Each can move differently.
Inequality between workers
When labour economists think about the effect of new innovations on inequality, we tend to assume something called skill-biased technological change.
As computers appeared on office desks for the first time, they were good at following explicit rules. That made them useful for replacing routine clerical work and repetitive production tasks. At the same time, computers complemented people who could program them, analyse the information they produced or use them to solve complicated problems. Typists lost their jobs. Data analysts were in hot demand.
Technology tended to raise the relative demand for more educated workers. In many advanced economies, it contributed to widening wage gaps.
Generative AI has a different comparative advantage.
You do not have to tell a large language model every rule for writing a useful email, summarising a document or explaining a spreadsheet. It has inferred patterns from examples. That allows it to perform tasks which earlier computers struggled to codify.
That changes which workers get the biggest benefit.
Think of AI as a compression algorithm for expertise. Some of the accumulated advantage of years of practice is now in a tool available to everyone in the workplace.
Experience remains important. In the customer support study, the strongest workers still began well ahead. But AI allowed weaker workers to partly narrow the gap.
A randomised experiment provides further evidence.
Researchers gave 1,174 adults a workplace-style business problem, randomising whether or not they had access to AI. Among people working without AI, those with more education outperformed those with less education by 0.55 standard deviations. When both groups received access to generative AI, the gap fell to 0.14 standard deviations. About three-quarters of the education-based productivity gap disappeared for that task (Cruces et al. 2026).
Education and experience are scarce assets partly because acquiring them consumes time. AI can make some elements of both available almost instantly. It is the chess-computer revolution, spread across thousands of occupations and available for the price of a streaming subscription. And it may already be showing up in the labour market, with the US college wage premium shrinking over the period 2022 to 2026 (Azar, Giné and Sanz-Espín 2026).
Not only do AI tools seem to offer a disproportionate boost to less educated and less experienced workers; they are also being adopted at unexpectedly high rates by lower-income Australians. Analysis by the e61 Institute finds that although higher-income Australians are more likely to pay for AI tools, the income gap in adoption is much smaller than it was for the internet at a similar stage (Thomas 2026). In 1998, the top 40 per cent of earners were 3.7 times as likely as the bottom 60 per cent to use the internet. For paid AI subscriptions in 2026, the ratio is closer to 1.5.
The distribution of jobs
Now consider a second mechanism.
The same AI that makes a junior worker more productive can reduce the number of junior workers a firm needs.
This is easiest to see through the task framework developed by Daron Acemoglu and Pascual Restrepo.
A job is a bundle of tasks. Technology can take some tasks away from workers. Another force expands human work because cheaper production raises demand or because new human tasks emerge.
Which force dominates depends on the technology, the industry and how customers respond to lower costs (Acemoglu and Restrepo 2018, 2019).
If you like your economic theories with a dose of alliteration, think of this as the distinction between augmentation and automation.
Consider an analyst whose job involves gathering information, cleaning data, constructing a model and exercising judgement about the result.
AI might help with the gathering and cleaning, freeing the analyst to spend more time interpreting the model.
That is augmentation.
Alternatively, the firm may conclude that one analyst with AI can perform work previously requiring two analysts.
That is automation.
The same software can produce both outcomes.
Entry-level jobs can be particularly affected because the work of younger employees often contains more codified knowledge and less accumulated organisational knowledge.
The junior lawyer only knows the law. The senior lawyer knows the clients.
The newly minted economist knows the theory. The experienced economist knows when its assumptions are about to cause trouble.
Generative AI is currently very good at the first type of knowledge. That could make entry-level tasks particularly vulnerable to automation.
There is debate about whether this is happening already in the United States. Work from the Stanford Digital Economy Lab finds a 19 percent employment gap among workers aged 22 to 25 in highly AI-exposed occupations (Brynjolfsson, Chandar and Chen 2026). However, other research finds that the main driver of the junior hiring slowdown is the shift towards working from home, which makes it harder to supervise junior hires (Lambert and Schindler 2026).
Complicating the picture, other countries are not seeing any slowdown in junior hiring. Research in Denmark (Humlum and Vestergaard 2026) finds no impact on employment of junior workers in AI-exposed industries. In Australia, analysis by the Department of Employment and Workplace Relations found that since the release of ChatGPT, employment among people aged 20 to 24 has grown slightly faster than employment among Australians aged 25 and over. The unemployment rate among young tertiary graduates was 5 per cent, low by the standards of the five years before COVID (DEWR 2026).
Interestingly, software development is routinely nominated as one of the occupations first in line for AI disruption. Yet in February this year Australia had around 199,000 software and applications programmers, about 40,000 more than immediately before ChatGPT was released. Employment in that occupation had risen by around 25 per cent.
The only hint of concern in the Australian job data came when researchers classified occupations by exposure to AI. Employment in the fifth of occupations most exposed to generative AI grew 5.6 per cent between late 2022 and early 2026, compared with 9.5 per cent in the least-exposed fifth. Statistical modelling finds a modest negative relationship between exposure and employment growth, although that result weakens under several alternative specifications and measures of AI exposure (DEWR 2026).
So the employment scoreboard currently reads: some concerning signals for young Americans, much less movement in Denmark, and Australian entry-level outcomes that remain solid. It is possible that the US is atypical, or that impacts there are driven primarily by the shift towards working from home. It is also possible that the US is experiencing AI impacts a year or two earlier than the rest of the advanced world.
Workers versus owners
Now we come to the distributional channel that concerns me more.
Imagine that AI ends up compressing wage differences among workers.
The junior accountant catches up with the senior accountant. The average programmer moves closer to the star programmer. The wage premium attached to education and experience falls.
Wage inequality could decline.
Yet total income inequality could still rise.
The reason is ownership.
When AI performs a task that a worker used to perform, part of production moves from labour to capital. In the Acemoglu-Restrepo framework, that pushes down labour’s share of value added. Creating new tasks for people moves the share back towards labour (Acemoglu and Restrepo 2018).
This is the distinction between the distribution of wages and the distribution of income.
It is quite possible for the first to become more equal while the second becomes more unequal. That could occur if pay packets equalise, but the economy tilts in favour of capital.
Daron Acemoglu’s recent work on AI considers this issue. His assessment is that current AI looks less likely than earlier waves of automation to produce a large increase in wage inequality across demographic groups. Yet his model predicts a wider gap between labour income and capital income (Acemoglu 2025).
There is some supportive evidence for this theory. In the United States, productivity has grown strongly over the past decade, while the labour share of business output has recently fallen to its lowest level in nearly eight decades.
Studying Europe, Antonio Minniti, Klaus Prettner and Francesco Venturini examine AI innovation across 273 regions. They find that a doubling of AI innovation is associated with a fall in the labour share of between 0.5 and 1.6 per cent.
They also find that the negative labour share effect is larger for medium- and high-skilled workers. For lower-skilled workers, stronger employment partly offsets wage pressure (Minniti, Prettner and Venturini 2025).
Again, AI looks to have a different impact than industrial robots.
The industrial robot was an obvious substitute for physical and routine production work.
Generative AI substitutes for many tasks performed by accountants, programmers, analysts and lawyers.
Until now, I’ve been focused on income inequality. But it’s also worth considering what AI might do to wealth inequality.
Benjamin Moll, Lukasz Rachel and Pascual Restrepo develop an economic model in which automation raises the return to productive wealth. Those gains accrue disproportionately to households that already own more capital. Higher returns compound over time. So automation widens the gap between the have-nots and the have-lots.
The trio note that looking only at wages doesn’t give a full picture of AI’s impact on inequality. Automation can increase wealth inequality, even if it doesn’t increase wage inequality (Moll, Rachel and Restrepo 2022).
Imagine two economies with identical AI-driven productivity gains.
In one, ownership of the productive capital is widely spread and strong competition passes much of the efficiency gain through to households.
In the other, ownership is highly concentrated and the leading AI inputs earn large economic rents.
In the first economy, an AI-driven productivity surge doesn’t increase wealth inequality. In the second, the same AI-driven productivity surge exacerbates wealth inequality.
What matters is who owns the productive assets. It’s why our National AI Plan sets out not only to capture the opportunity of AI, but to spread the benefits.
Who owns the toll road?
AI can be a huge driver of productivity and economic benefit for Australia, but first people have to be able to trust that the technology is working for them, and not the other way around.
The National AI Plan is a clear roadmap for building public trust and safety measures to increase AI adoption across the economy – and realise lasting productivity gains.
When it comes to data centres, the Australian AI Standards announced by the Prime Minister will set clear national requirements for large data centres to make sure we capture the AI infrastructure investment boom in a way that’s in Australia’s interests.
We are working with states and territories to ensure there are consistent rules, clear investment signals and a simple compliance process that’s transparent for communities and proponents.
As the Prime Minister has said – this is the time to determine what AI looks like in Australia.
As Assistant Minister for Competition, you won’t be surprised to learn that I’m thinking hard about the effects of market structure.
It’s important because most previous workplace technologies were pieces of capital a business could buy. When electrification came, firms owned the machinery. When computerisation came, firms owned the desktops.
AI increasingly looks like a service a business can access.
This means that if every bank, law firm, retailer and manufacturer relies on a small number of upstream AI systems, then the model provider begins to resemble a toll road sitting across a growing share of the economy.
The economic question is the size of the toll.
Foundation models have considerable forces pushing towards scale (Leigh 2026). Frontier training requires enormous computing resources and highly specialised talent. Model improvement can benefit from scale, while distribution and integration can create further advantages. On one estimate, ChatGPT has three-quarters of the AI chatbot market.
At the same time, the industry currently features significant competition among major developers.
Anton Korinek and Jai Vipra model these forces and show why economies of scale could produce greater concentration over time, including the possibility of market tipping or vertically integrated firms controlling important inputs and downstream applications (Korinek and Vipra 2025).
For inequality, concentration affects who captures the productivity gains.
Susan Athey and Fiona Scott Morton model AI as an upstream input purchased by firms.
Suppose AI makes an Australian business more efficient. Some of that efficiency can appear in lower domestic production costs. But if the AI supplier possesses substantial market power, part of the gain is extracted through access and usage charges.
For a small open economy, some of those payments flow overseas.
Their model therefore produces the result that labour-displacing AI can reduce some workers’ wages while market power in the AI input transfers an additional share of the surplus to the upstream supplier (Athey and Scott Morton 2025).
Creating value with AI doesn’t mean you’re capturing value from AI.
A business may achieve a productivity improvement yet retain only part of it if suppliers possess pricing power and competitors quickly copy the innovation.
Conversely, an AI provider with market power may collect rents across downstream industries.
As my late coauthor Tony Atkinson noted, industrial organisation can affect income distribution (Atkinson 2015).
The global distribution
Until this point, we have been focused on inequality within one country, but another aspect of inequality is how it affects the gaps between countries.
The customer-support worker that I started with was likely outside the United States. In the study, most agents were based overseas, predominantly in the Philippines.
For decades, countries such as the Philippines and India have built export industries around relatively abundant educated labour, but AI changes that comparative advantage.
A call centre interaction that once had to be performed by a person in Manila can increasingly be assisted, shortened or partly performed by software running in a data centre.
That also applies to transcription, coding, bookkeeping and parts of professional services.
Anton Korinek and Joseph Stiglitz point out that labour-saving AI could weaken one of the development paths used by lower-income economies: combining relatively inexpensive labour with access to wealthy export markets. If the most valuable AI assets are simultaneously concentrated in advanced economies, the technology can shift income towards countries that own the intellectual and physical capital (Korinek and Stiglitz 2021).
Here the evidence is thinner, partly because the global deployment of generative AI is still so recent.
AI substantially lowers the price of certain forms of expertise. That is good news for a worker gaining access to expertise.
It has a different implication for a country whose export advantage consisted of selling that expertise cheaply.
Conclusion
So how can we sum up the effect of AI on inequality? I would draw two conclusions: one positive, the other a little less optimistic.
First, generative AI appears capable of compressing productivity differences among workers.
Studies repeatedly find larger productivity gains among people who begin with less experience, less education or lower measured performance.
That distinguishes this technology from many earlier waves of computerisation.
The second conclusion is a smidgin more challenging.
The risk of AI on inequality is not primarily through its effect on the wage distribution.
It lies in who gets hired, who owns the productive assets, and how much of the surplus is captured by firms with market power. It also lies in which countries are able to deploy AI throughout their economies while retaining a substantial share of the returns.
This is why aggregate productivity forecasts tell us little about inequality.
Even Acemoglu, whose forecast of AI-driven productivity growth is towards the conservative end of the economics literature, concludes that the technology could widen the gap between capital and labour income (Acemoglu 2025).
A modest productivity gain can have a large distributional effect if it changes who gets paid for doing the work.
In the area of productivity policy, this underscores the value of our work on competition, technological diffusion and worker mobility. It also raises the importance of broad participation in capital ownership. Universal superannuation means that many Australian workers have part of their savings invested in technology firms, making it more likely they will share in the gains.
Competition affects whether lower costs become lower prices for the many or larger rents for the few.
Diffusion affects whether AI capability reaches regular suburban small businesses or remains the preserve of a few technological leaders.
Worker mobility affects whether people displaced from one set of tasks can move into expanding parts of the economy.
And patterns of ownership determine who receives the capital income generated by automation.
These are old economic institutions facing a novel technology.
For investors, it may be worth not only thinking about how exposed a company is to AI, but also what happens to its business model as AI becomes ubiquitous. Firms that are able to use AI to scale quickly are likely to be especially attractive to investors, particularly if they are able to avoid the situation in which upstream providers capture much of the return.
Let me finish where I began.
A technology that lets a worker with two months’ experience perform like a worker with more than six months’ experience might be a powerful equalising force.
But suppose the firm subsequently needs fewer new workers. Suppose much of the saving accrues to the owners of capital. Suppose the AI provider is able to charge more for every transaction.
The productivity gap between the two workers may have narrowed, while the income distribution across the economy has widened.
So I think the most likely answer to the question ‘Will AI increase inequality?’ is that different measures of inequality may head in opposite directions.
AI could narrow the gap between the novice and the expert, but at the same time, it could widen the gap between people who earn mainly from their work and people who earn heavily from owning productive capital.
Which force dominates will depend partly on what the technology can do.
The rest is economics: competition, ownership, institutions and the way firms choose to deploy it.
AI will boost productivity. Sharing the benefits fairly is the real challenge.
Our government will apply the enduring values of fairness to make AI work for Australians.
Thank you.
References
Acemoglu, D. (2025), ‘The simple macroeconomics of AI’, Economic Policy, 40(121), 13–58.
Acemoglu, D. and Restrepo, P. (2018), ‘The race between man and machine: Implications of technology for growth, factor shares, and employment’, American Economic Review, 108(6), 1488–1542.
Acemoglu, D. and Restrepo, P. (2019), ‘Automation and new tasks: How technology displaces and reinstates labor’, Journal of Economic Perspectives, 33(2), 3–30.
Athey, S. and Scott Morton, F. (2025), ‘Artificial intelligence, competition, and welfare’, NBER Working Paper No. 34444.
Atkinson, A.B. (2015), Inequality: What Can Be Done? Harvard University Press, Cambridge, MA.
Australian Treasury (2026), ‘Economic Implications of Artificial Intelligence’, Ministerial Submission MS26-001548, 27 August (publicly released).
Azar, J., Giné, M. and Sanz-Espín, J. (2026), ‘The College Wage Premium in the Generative AI Era’, Revised 10 September. Available at SSRN: http://dx.doi.org/10.2139/ssrn.7364838
Brynjolfsson, E., Chandar, B. and Chen, R. (2026), ‘Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence’, Stanford Digital Economy Lab Working Paper, revised 12 August.
Brynjolfsson, E., Li, D. and Raymond, L. (2025), ‘Generative AI at work’, Quarterly Journal of Economics, 140(2), 889–942.
Cruces, G., Fernández Meijide, D., Galiani, S., Gálvez, R.H. and Lombardi, M. (2026), ‘Does generative AI narrow education-based productivity gaps? Evidence from a randomized experiment’, NBER Working Paper No. 34851, revised May 2026.
Department of Employment and Workplace Relations (DEWR) (2026), AI and employment in Australia: Monitoring framework and evidence to date, July, Canberra.
Humlum, A. and Vestergaard, E. (2026), ‘Still waters, rapid currents: Early labor market transformation under generative AI’, NBER Working Paper No. 33777, revised March 2026.
Korinek, A. and Stiglitz, J.E. (2021), ‘Artificial intelligence, globalization, and strategies for economic development’, NBER Working Paper No. 28453.
Korinek, A. and Vipra, J. (2025), ‘Concentrating intelligence: Scaling and market structure in artificial intelligence’, Economic Policy, 40(121), 225–256.
Lambert, P. and Schindler, Y. (2026), ‘The Broken Ladder: AI, Remote Work, and Early-Career Hiring’, Working Paper, London School of Economics.
Leigh, A. (2026), ‘Competition Policy for Invisible Agents’, Address to ‘The Invisible Mind: The Economics of AI’ conference, Economic Society of Australia (NSW Branch), Sydney, 17 June
Minniti, A., Prettner, K. and Venturini, F. (2025), ‘AI innovation and the labor share in European regions’, European Economic Review, 177, 105043.
Moll, B., Rachel, L. and Restrepo, P. (2022), ‘Uneven growth: Automation’s impact on income and wealth inequality’, Econometrica, 90(6), 2645–2683.
Productivity Commission (2025), Harnessing data and digital technology, Inquiry Report No. 111, Canberra.
Thomas, H. (2026), ‘AI versus the internet’, plus61, e61 Institute, 10 July.
ENDS