Thinking like an economist about AI, labour markets, and AGI

You can listen to this podcast on Spotify, Apple Podcasts, or wherever else you get your podcasts. You can also watch this conversation on YouTube, and we have started a Substack for this podcast.

Recently, we’ve seen many dire predictions about the impact of AI on labour markets. We’ve heard, for example, from Dario Amodei, CEO of Anthropic, that AI could wipe out half of all entry-level white-collar jobs. And in 2016, Geoffrey Hinton, one of the pioneers of AI, said “we should stop training radiologists now”. It seemed like algorithms were about to take over radiologists’ jobs, with AI models able to detect pneumonia on chest X-rays more accurately than expert doctors.

Yet here we are in 2026, when hospitals are hiring more radiologists than ever, and their salaries are higher than before. While the algorithm performed well on benchmarks, it could not yet replace the human.

So what’s going on? In this episode of Ideas in Development, we are joined by Anton Korinek to think like an economist about AI and labour markets, and discuss whether our current economic paradigm will hold up as AI moves from being a tool to a coworker. In the first half of the episode, we go through some rules of thumb to cut through the noise.

1. Big User Numbers ≠ Big Economic Impact.
2. Benchmarks ≠ real world performance.
3. Correlation ≠ Causation.
4. Consider the source.

Why widespread AI adoption ≠ big economic impact

One of the most common metrics in AI reporting is adoption. It’s tempting to think that if an AI tool has hundreds of millions of users, or writes a huge percentage of code, the economy must already be transforming.

But Anton points out that since the 1990s the digital economy has produced many services with enormous user bases that have only had modest impacts on standard economic aggregates. Social media is one example which is economically harder to see in the numbers.

This is not to say that nothing is happening. Anton argues that the mismatch partly reflects the limits of what GDP and other statistics are designed to measure.

“what our economic statistics measure, and what matters for us and what matters for our lived experience, it's two different things.”

GDP can miss genuine value creation in the digital economy, since it largely captures market transactions, the priced output of goods and services, not the consumer surplus created when services are free or extremely cheap. Search, social platforms, and many large language model (LLM) tools may improve lives or work processes without showing up proportionately in headline figures.

We need better measures of where value is being created, who captures it, and what changes inside firms actually translate into productivity gains.

Benchmarks meet the real world

A core driver of some of the hype we see about AI comes from different model’s performance on benchmarks. Benchmarks allow progress to be tracked and optimised against standardised tests, but economists ask a different question: does AI change what workers and organisations can do, at scale, in routine settings?

Anton argues that the gap between benchmarks and economic outcomes is often a matter of diffusion. A technology can exist and still take years to become productivity in practice. Firms must redesign workflows, workers must learn new tools, and complementary investments (data systems, governance, training, regulation, integration) must catch up.

“even if we stopped all progress in artificial intelligence today, and we spent the next five or ten years figuring out how can we use the systems that we have right now throughout our economy in a productive way. We would probably experience very significant productivity gains.”

Looking forward, for many everyday uses, improvements in new models have become hard to notice because performance is already high. The frontier is shifting most visibly in demanding tasks, such as solving complex dynamic economic models or coding autonomously for long stretches.

Correlation, causation, and what the current evidence really says about AI and jobs

A ‘think like an economist’ piece would be incomplete without mentioning that correlation does not equal causation. It is tempting to blame AI for worrying labour market trends, especially when the timing lines up, but actually establishing causal links is challenging.

There is a strong intuition that AI is reshaping the job landscape, but limited conclusive evidence at the macro level. Where evidence is clearer is in narrow categories, particularly tasks that can be cleanly substituted by generative AI today. But this is not the same as proving AI is driving aggregate unemployment or economy-wide wage shifts.

Anton highlights an important channel that doesn’t show up neatly in simple ‘jobs destroyed’ narratives: uncertainty. If firms believe AI capabilities are improving quickly, they may delay hiring, especially for entry-level workers who require training. Korinek suggests companies might ask whether to invest in training a fresh graduate now, or wait a year and hire later once the technology (and the job design around it) is clearer. That kind of behaviour could plausibly depress entry-level opportunities even without dramatic job losses showing up in official statistics.

Who to believe? Incentives, hype, and whether lab leaders can be trusted

Another rule of thumb is about information sources. Those leading AI labs dominate the public debate and have their own incentives to sell a compelling story for fundraising, market positioning, and influence.

But while Anton does believe there is a little bit of hype, his view is that the biggest claims are not purely marketing spin. Based on his interactions with staff inside labs, he reports a broadly similar internal view: progress is rapid, the ceiling is unclear, and many researchers are genuinely worried about what the technology implies for society.

“It must be similar to people who have been working on the Manhattan Project and they know that what they're working on is something really big and they're scared of the power of what they're developing.”

Jevons paradox and the future of work

The second half of the episode turns from our rules of thumb, to how economists model the labour market effects of automation. We focused on two key concepts: Jevons paradox and comparative advantage.

In relation to AI, Jevons paradox captures the idea that making something more efficiently and cheaply with AI could actually increase total demand for that good enough to offset reduced labour needs in producing it. This is sometimes offered as a reason not to worry about job loss from AI, but Anton warns against using this logic to assume everything will be fine.

Past technological waves still produced painful transitions, and with AI, he expects similar patterns in the short run: displaced workers move into other tasks as long as AI has not reached general intelligence. He emphasises that transitions are difficult but possible – and raises the idea of policy support, explicitly mentioning an AI adjustment programme analogous to trade adjustment assistance, to help workers retrain and relocate.

AI as a tool, AGI as a worker. Will comparative advantage save us?

Where Anton’s train of thought gets more worrying is in the scenario where AI approaches AGI, i.e. becomes capable of performing virtually all economically valuable work. In that world, the central question becomes: what is left for humans to do, and on what terms?

Comparative advantage might seem reassuring. Even if AI is good at everything, humans could still specialise in what they are relatively least bad at. Korinek agrees that the logic can still hold even under absolute machine advantage. Yet he stresses that this efficient allocation could still be a grim one for humans, because wages may be pushed down towards the machine cost level.

“you could imagine a world in which AI systems and robots and so on can produce everything so cheaply that humans who have to compete with them won't even be able to afford a subsistence income.”

This is a speculative vision of the future, but Anton’s message is that neither jevons paradox, or comparative advantage, should make us feel particularly shielded if AGI is achieved. Even if economies as a whole become richer, workers can lose badly, and then the distribution of gains becomes the decisive political and moral question.

Can economics keep up with AI? Scenario planning, research methods, and a call to the profession

Economists value rigour and tend to avoid speculation, but thinking seriously about AGI requires engaging with uncertain futures. Anton feels that the probability of transformative AI is now high enough to justify speculative research which comes with very clear caveats.

One approach Anton is keen to see more of is scenario planning. This involves building and stress-testing multiple futures for technological progress, diffusion, labour markets, and policy responses. Economics and the social sciences can add serious value to these exercises, by offering tools for understanding incentives, institutions, and distribution rather than just engineering feasibility.

And AI is already changing economic research itself. Tasks that previously required graduate research assistants can now be accelerated by AI systems, and Anton believes that economists may increasingly use frontier models as junior co-authors. The frontier, in his view, is moving from helping with execution (editing, coding, summarising) towards shaping research questions, which is arguably the distinctive human contribution that has defined scholarship.

Anton hopes that economists treat this moment seriously and contribute proactively to guiding society through it:

“to hopefully come out of it at the other end in a way that AI has fulfilled all those utopian promises of making us all wealthier and happier and more fulfilled and has avoided the potential dangers of severe disruption and undermining of livelihoods and so on.”