Nobel Economist Daron Acemoglu on Why AI's Economic Impact Is Still Muted and What Workers Should Learn

The host introduces Daron Acemoglu as an MIT economist and winner of the 2024 Nobel Prize in Economics, and says he thinks people are overlooking some of AI's biggest opportunities. The host asks where people should be looking and what will keep them valuable.

Acemoglu says he can imagine a very scary future in which many people would not have meaningful work. He describes this as one of the two possible futures that the discussion revolves around, the other being one in which AI makes workers more productive.

When asked about agentic AI, Acemoglu says advances in coding have been very rapid. He explains that coding is done in offices with little direct interaction with the real world, and that the people supervising it are fluent in AI and technology.

He adds that, as with math and chess, coding has rules: a program either runs or does not run. Reinforcement learning works well when there is a ground truth that says an answer is wrong. He says that in many occupations this kind of ground truth is lacking.

Acemoglu says most people would have expected customer service to be the first occupation fully automated by AI. He says some automation is happening, but it is turning out to be very slow and very difficult. He says the AI systems of most organizations he deals with are so bad that calling customer service can feel like a nightmare.

He attributes this to a social element: the requests people make to customer service change over time and are person-specific, and different people use different language to ask the same question. There are also many gray areas and edge cases. He says this will require humans either to oversee the AI, work with it, or ideally to collaborate so that AI complements humans.

He adds that the path chosen will matter for how many jobs are displaced, how much productivity gain results, and who benefits.

Acemoglu says that if he were asked why AI's productivity effects are so small and what can be done, his answer would be one word: much better and more applications. He qualifies this, saying he is not inside the AI labs and does not know what they learn from their experiments.

He argues that Silicon Valley puts too much emphasis on foundation models and not enough on applications or on integrating applications into organizations. He acknowledges that the US leads China in large language models, but says Chinese models are catching up, sometimes using the lead of US models.

He contrasts this with Chinese companies, which he says put a lot of emphasis on bringing robotics and AI into production lines and throw a large amount of engineering talent at the problem. He says the US does not seem to have that much engineering talent, and what it has is mostly going into foundation models.

He also criticizes what he calls inevitabilism, which he says has not been good for the AI industry in his opinion.

Acemoglu says general purpose technologies such as electricity and the printing press spawned many other applications, and that there are many different ways of developing such technologies. He argues that these paths are not neutral: they bring different types of benefits and costs, so assuming the current approach is the only or best way is a mistake.

He also points to measurement as a problem, asking how one can measure AI's impact on personal or company productivity. He recalls talking with Erik Brynjolfsson, who was conducting a study, and says the economic impact of AI was basically zero when they spoke in May.

The host summarizes that raw capabilities are skyrocketing while economic impact is muted, and Acemoglu agrees that this gap is a big opportunity. He says Brynjolfsson is more optimistic than he is, and that his own view is that AI is not yet having the productivity effects expected but will, only somewhat more slowly than Brynjolfsson would say.

Acemoglu says healthcare and education involve a two-way or three-way street because most healthcare dollars involve government in some way, and teachers too, so government could play an important role there. For electricians, blue-collar workers, customer service representatives and scientists, he says the private sector is mostly responsible.

Even without government, he says two parties are needed: tech companies and business management. He says CEOs and high-level managers he has spoken with say that pro-worker AI, which makes workers more productive, is exactly what they need. He says he cannot test whether they mean it, because tech companies are not supplying such tools.

When he raises the same idea with people from tech companies, he says they either nod or tell him he lives in a different world and that they are going to create AGI.

Asked for good examples involving blue-collar work and China, Acemoglu says he has spoken to several manufacturing companies in China that are integrating both robotics and AI, naming Foxconn, BYD and other EV producers. He says these companies are very tech-savvy.

He says that from their experience the process is slow and requires more engineers than companies always think. He notes that China has an enormous supply of engineers, which is fortunate for those firms.

Acemoglu worries that the critical thinking problem will be harder to escape than the calculator problem. He says the calculator could only take over a very small part of thinking, and he knows of no one who used a calculator to do their thinking for them. He also knows of no school that told students they no longer needed to learn how to multiply, since calculators existed when he was a student.

He says he personally does not use AI for math, though AI is much better than him at it. He likes doing math himself, writing down models and doing calculations by hand, because that is what he is used to, lets him work quickly, and lets him learn from his mistakes.

The host asks where to stop using AI and how to still engage in key steps of human innovation and creativity alongside it. Acemoglu says he worries that AI's effect on critical thinking as a whole, not just a narrow function, is the key difference from the calculator, and that the boundaries are very gray.

Acemoglu says he is not against automation and does not think it can be stopped. He says he opposes anyone trying to reverse automation and does not want to return to carrying heavy things on our backs.

His objection is to doing only automation. He argues that this would not deliver full productivity, because not everything is easy to automate. It would also fail to create shared prosperity, because pure automation eliminates jobs without creating meaningful high-paying, high-productivity jobs.

He says AI should also be used to create new things, such as new work for nurses and electricians and many occupations that cannot be imagined right now.

Asked how to explain this to a small business owner whose only metric is increasing profits, Acemoglu says productivity is the answer. He says that in his experience most businesses understand that their workers are their greatest asset, so making them more productive is fantastic for the business.

He presents a comparison: with 100 employees, automating 5% of jobs and cutting costs by half in those jobs, versus making workers 5% more productive. He says the latter will get a business much more.

He adds that successful companies do something new rather than just cut costs. He cites Ford, General Motors, General Electric, IBM and Microsoft, and quotes Henry Ford saying that if he had asked consumers what they wanted, they would have said faster horses.

Asked how to know which tasks to automate, Acemoglu notes that he mentioned 5% of tasks are automatable. He says that if costs fall by 5%, GDP typically rises by 5% times the sector's share of GDP. He calls that arithmetic not hard.

He says the hard parts are that changing quality complicates the calculation, and that even a 5% cost cut can take a lot of work and years to achieve. He has seen businesses take the automation plunge because they think it will be easy or because they feel pressure, and then fail to make it work because it is not that easy and must be integrated into the company.

On the phrase 'jagged frontier,' he says he likes the image but that it is often a code word for admitting AI was overestimated. Businesses that keep the same number of workers and do a bit of AI are, in his words, just a mélange.

Asked whether university will still be worth a lot of money in 10 years, Acemoglu says he thinks so but offers caveats. He says he has never believed everybody should go to college, because people have diverse skills and interests, and that lifelong learning should be encouraged.

On which skills to acquire, Acemoglu says he does not believe true AGI is just around the corner, so human creativity is needed more than ever. He says human creativity without strong foundations is an oxymoron. He acknowledges he is not a mathematician or computer scientist but says he would recommend studying math, physics, economics and engineering, and learning flexibility, since even within one occupation people will need to change what they do as technology and social context change.

He says human ingenuity and creativity will be critical in the next 10 years. The host agrees that changing careers over a lifetime will become the new normal. Acemoglu adds that starting in middle school, students should learn to work alongside AI and also place more emphasis on civic responsibility and a sense of community and the common good.

The host cites Dario Amodei saying there is a 25% chance that AI goes really badly and a 75% chance it goes really well. Acemoglu describes himself as relatively an optimist and says he thinks people in leading AI labs who are worried about some implications of AI should voice that, and he welcomes their courage.

He says he is more worried about a general loss of control by society over AI than about existential risk. He means that the public, workers, students and parents could lose control of their own agency because of AI. He attributes this partly to how other companies and frontier labs develop, use and roll out AI, rather than only to AI becoming an independent superintelligence. He says the conversation should happen but other risks should not be ignored.

The host notes that Acemoglu once rated AI at minus 6 on a scale from minus 10 to plus 10 and asks whether it has gotten worse since 2025. Acemoglu says he does not think so, because in 2024 and 2025 he thought AI could do a lot of damage and was not clear about what bottlenecks would face its improvement.

He says improvements have been fast, so the dangers have increased, but the good things that can be done with AI are also becoming clear. He argues that AI must be redirected and steered, and that it is too important to be decided by a handful of smart, well-meaning people.

On how to foster that at an individual level, he says the process starts with information and narrative. He criticizes the media for failing to cultivate a true information debate about AI, saying that for years people either treated figures like Elon Musk as beyond criticism or cast AI as an enemy, leaving little middle ground for public engagement.

Asked what a typical workday of a knowledge worker might look like in 2035 if everything goes positively, Acemoglu says he does not know. He describes three possibilities: a future in which people work a few less hours and are still very productive; a very scary future in which many people lack meaningful work; and a future in which people still go to the office or combine office and home work differently, with human innovation and talent central.

He says the world is at a uniquely turbulent time even apart from AI, citing climate change, an aging population, shifting geopolitical balances, and a huge middle class emerging from the developing world. He argues these require new goods, services, tasks, ways of approaching markets and ways of organizing companies, and that the people who will do this are humans, not AI.

Acemoglu says he does not see how foundation models will make the large sums of money that would justify the investment in frontier labs, so he expects some failed IPOs, though not immediately because enthusiasm is still high. He points to possible bottlenecks in model development or diffusion, and notes that open-source models in the next layer are going to be very good, so monetization opportunities will be limited.

He says the money is in high-quality data and applications. Companies with a lot of data should cultivate and use it, and compensate the workers who generate that data. He adds that applications, especially creative ones that fill a niche, will be much harder for open source to replicate.

He also stresses that workers are a source of data and tacit knowledge. Developing good applications requires going to workers, getting hands dirty, and talking with them, rather than sitting in an office with a powerful AI.