Daron Acemoglu

Institute Professor @MIT, @MITEcon. Co-Director of @MITShapingWork. Author of Why Nations Fail, The Narrow Corridor, and Power & Progress.

[Daron Acemoglu @DAcemogluMIT]

My new book #WhatHappenedToLiberalDemocracy is out on August 11. Here is a brief message about my motivation for writing this book and its central thesis. Pre-order links are available here: https://t.co/Ov218cLBqt. I hope you will join me in the discussion.

[Gabriel Zucman @gabriel_zucman]

Hi @elonmusk , we’ve conducted the most comprehensive study to date on tax unfairness in the United States.

We found that you, like your fellow billionaires, pay a lower effective tax rate than the average American.

So if you really want to help the American people, I have a better suggestion than charity: pay more taxes. 🧵

[Daron Acemoglu @DAcemogluMIT]

An excellent analysis of where Europe is right now.

It’s Not Just France. Europe’s Political Center Is Crumbling. https://t.co/WmqU6RDlGX via @NYTimes

[Daron Acemoglu @DAcemogluMIT]

A badge of honor

[Henry Gao @henrysgao]

The most intriguing part of the investigation readout on former Social Security Fund vice chairman Wang Wenling is the charge of privately bringing in and reading books with "serious political issues", rumored to be @DAcemogluMIT's "The Narrow Corridor".

[Timur Kuran @timurkuran]

Team Erdoğan’s corruption has become unsustainable. In a normal country, the leader would hit reset. Erdoğan may no longer be able to. Too many of his cronies now have vested interests in the status quo. A real cleanup would jeopardize the arrangements that keep them loyal.

[Daron Acemoglu @DAcemogluMIT]

This is interesting.

[Joshua Blanchard @philoshua]

A lot of philosophy-infused AI discourse is ridiculous because it boils down to this:
"Look at this dumb layperson who believes [contentious view held by some philosophers of mind]. As an expert, I can tell you [contentious view held by some philosophers of mind]."

[Yoshua Bengio @Yoshua_Bengio]

Over 400 researchers and I have signed this open letter calling for the AI and security communities to come together to improve alignment and containment approaches in order to protect our critical infrastructure and digital systems.

https://t.co/KieWSpurn9

[asli aydintasbas @asliaydintasbas]

Are you also worried about the future of democracy?

Come and join us for this conversation with @DAcemogluMIT on Tuesday October 13!

Prof. Acemoglu is asking the right question: Can ‘working-class liberalism’ heal liberal democracy? https://t.co/iu1IkuCz2h

[Daron Acemoglu @DAcemogluMIT]

Thanks to @joshgans for collecting my questions on AI and his responses in a single substack post.

[Joshua Gans @joshgans]

You asked for it. Here is the exchange between myself and @DAcemogluMIT collected in one post. The Acemoglu-Gans Exchange on AI. https://t.co/M4c3uRiRqV

[Daron Acemoglu @DAcemogluMIT]

This would be funny— if it weren’t our tragic reality

[Rutger Bregman @rutgerbregman]

So OpenAI and Anthropic expect a catastrophe by 2027, and they’re planning… the PR response.

Have they considered not causing the catastrophe?

[Daron Acemoglu @DAcemogluMIT]

Another thoughtful response from @joshgans

[Joshua Gans @joshgans]

In his 7th (and now final) question on AI, @DAcemogluMIT asks the question that he and I have been debating for years: will AI create as many jobs as it will replace?

We actually agree on some important parts. First, while AI may well transform jobs, this process will not be rapid. In my mind, that gives us time to adjust.

Second, in the long term, there may be labour reallocations, and they may impact regions differently. Indeed, back in the financial crisis, Daron gave a series of talks worrying about whether government actions may get in the way of a labour reallocation process he thought was necessary there.

Third, if it turns out that AI cannot easily automate jobs, then AI will turn out to be "more pro-Worker", and things will be fine. He reiterated this point in his post: "Of course, things may turn out differently. But if they do, it will be because AI becomes more pro-worker. I see this as unlikely, but it is useful to still consider the scenarios that would bring such an outcome. These are: (1) current AI technologies may become turn out pro-worker than I am anticipating; (2) in the future the industry will automatically and by itself change and focus on pro-worker applications; (3) based on AI-driven science and research, new sectors will be created where there will be need for new workers."

The first item is one that @avicgoldfarb and I developed in our paper on O-Ring Automation; that when tasks are highly complementary, automating most of them doesn't eliminate a job or even lower worker productivity. This is something @DarioAmodei has also recently argued.

When it comes down to it, Daron and I have the same economic model in mind. Where we differ is the probability weights we are putting on uncertain parameters. I tend to be more optimistic. He tends to be less. There is no resolution to this other than hard evidence, and that doesn't exist quite yet. We are at a stage where any glimmer of hard data is playing an undue role in updating people's beliefs here, and we should be more Bayesian about that.

There is another aspect where we differ. Daron is of the opinion that governments can move the needle on pro-Worker AI by, say, reforming capital treatment in the tax code and giving people strong property rights over data.

However, I think this is all going to require a ton of innovation and adjustment, with its own uncertainties, and so I am sceptical that governments can really help here. Nonetheless, if there are major disruptions and my optimism is misplaced, I think there will be government responses, but not much in that scenario is going to be fun.

[Daron Acemoglu @DAcemogluMIT]

As usual, @Yoshua_Bengio is raising central issues and speaking with great moral and intellectual clarity. Getting the direction of AI right is a responsibility for all of us.

[Yoshua Bengio @Yoshua_Bengio]

We have a choice in front of us. In my latest op-ed for Transformer, I invite employees of frontier AI companies to ask themselves if their work is truly leading us toward a safer future. Reflecting on these questions is hard, but we no longer have the luxury of time. We must act now.

Read my full op-ed: [https://t.co/ruFBouQJbk]

[Daron Acemoglu @DAcemogluMIT]

Seventh and last question on AI.
Can AI really help us create as many jobs as it will replace?
The AI industry offers two narratives at once.
The first emphasizes that because of AI’s capabilities, there will be a lot of job loss, and soon. Anthropic’s co-founder and CEO Dario Amodei is particularly vocal on this: https://t.co/hcOmaSZIKD
The second claims that AI will also enable lots of new jobs, so we should not fear joblessness, higher unemployment, and humans being sidelined (see, for example, Jensen Huang, https://t.co/6ZO0eGWvuX or Marc Andreessen, https://t.co/JrKsPIpTvh).
A lot hinges on this question. Sizable job displacement is costly for workers and could destabilize our already ailing democracy. Displacements from robots and cheap Chinese imports in the 1990s and 2000s in the United States, which were smaller than the projected job losses from AI, contributed significantly to declining trust in government and institutions (see my book What Happened to Liberal Democracy, https://t.co/7xaaR1eyuw, on this). Our democracy may not be able to survive another massive job dislocation.
I am worried about job loss, for two fundamental reasons. AI is currently being developed as an automation technology and in addition, I do not find the arguments about large numbers of jobs being created by AI particularly realistic. (I’ll come back below to how AI could create jobs if it weren’t developed as an automation technology).
Proponents of the job creation side fall back on four general arguments, none of which is convincing in my assessment.
First, new technologies have always created lots of jobs. True, but that does not guarantee that this time will be the same. And more importantly, this account neglects the richer details about different historical episodes. In a nutshell, automation technologies by themselves need not create jobs (though they can, depending on circumstances and how productive they are). They involve capital substituting for tasks previously performed by workers. The way they would create jobs is that the displaced workers would go and take non-automated jobs in other parts of the economy. This can happen, but it also depresses wages (or at the very least the labor share), and consequently, some of the workers may be discouraged from looking for work or may not have the skills that remaining non-automated tasks require. This, in any case, is what task-based models predict. See, for example: https://t.co/czsxeCUjjZ and https://t.co/zlb8XL7vfC.
The empirical evidence also casts doubt on plentiful job creation taking place automatically. For example, the introduction of robots led to many fewer jobs in local economies that depended on blue-collar jobs in robot-heavy industries (and this effect was quantitatively large; one more robot per thousand industrial workers reduced the employment population ratio in the local economy by about 2 percentage points and local wages by 0.42% between 1990 and 2007). See https://t.co/Lteugf7ckL.
Second, AI will increase productivity and higher productivity will lead to more jobs. But this argument suffers from the same weakness as the previous one: higher productivity will trigger job creation only in remaining non-automated tasks (unless there are new tasks, which I discuss below). If wages are depressed and workers don’t have the skills for these non-automated tasks, even with higher productivity, there will not be AI-driven job creation.
Third, AI will need a lot of supporting roles, such as people to train the AI models. This one is the weakest argument of all. If AI is indeed an automation technology, then the whole point is to save labor costs. By the same logic, the labor needed to train AI must be much less than the labor AI ultimately saves. For this reason, I see essentially no chance that there training and maintaining AI will create anywhere near as many jobs as AI replaces.
Fourth, AI is already complementing workers. This argument is based on some evidence that many workers are using AI in their jobs – see, for example, https://t.co/P70kGyTvuO. However, the mere fact that workers are using AI in their jobs doesn’t establish that new jobs will be plentiful.
The theory here is again quite clear (see the detailed treatment of this issue here, https://t.co/AbYZ2zaU49). If AI simply makes workers more productive in the tasks that they used to perform already, it will drive down task prices. If you are the only worker using AI, this will benefit you. If all workers like you start using AI, then the skills on which you depend become commodified. This is what happened, for example, to London black cab drivers, whose highly-valuable knowledge of London streets was commodified when GPS technology enabled minicabs, and then later Uber drivers, to perform the same tasks as effectively. To put it simply, it isn’t enough for AI to make workers more productive at the tasks that they are performing already.
This brings me to how AI could create a lot of jobs. Again, the task-based framework provides the answer here: new technologies create jobs, and in fact well-paying jobs, if they enable new tasks (either in the same industry or in new industries). I have argued elsewhere that AI has this potential as a powerful information technology. See, for example, https://t.co/aDd8GLNIkN. But this isn’t currently the direction in which AI is being developed. Hence, on its current trajectory I remain quite concerned about future job losses (and wage declines) from AI.
Of course, things may turn out differently. But if they do, it will be because AI becomes more pro-worker. I see this as unlikely, but it is useful to still consider the scenarios that would bring such an outcome. These are: (1) current AI technologies may become turn out pro-worker than I am anticipating; (2) in the future the industry will automatically and by itself change and focus on pro-worker applications; (3) based on AI-driven science and research, new sectors will be created where there will be need for new workers.
I’m not optimistic about any of these possibilities, and therefore my conclusion is that AI is likely to lead to more jobs being lost than gained. (But I do think there are policies that would encourage pro-worker AI; see, for example, https://t.co/yIjrtqmEeJ).
That being said, I don’t think the job loss will happen very quickly, because as I discussed in an earlier post (third question), the diffusion of AI is relatively slow. But this also implies the productivity gains from AI will be slow in coming.

[Daron Acemoglu @DAcemogluMIT]

Sixth question on AI.
Does AI need to be a zero-sum race between the US and China?
Many influential voices in tech as well as the current administration argue that it does. (See, for example, https://t.co/63jNKZIGON, https://t.co/82WRjDEklx, https://t.co/3isNlCWI2U and this from the White House, https://t.co/roDrwiWjIj).

There are many aspects of these arguments, but I would like to focus on three that have received less attention than they deserve.

The first aspect is that this framing typically assumes that there is a single direction of progress for AI, and the US and China are racing along this dimension toward superintelligence or at the very least some sort of powerful AI that will confer huge military and strategic advantages to the winner. If, on the other hand, the US and China are developing these technologies in different directions, there is much more room for coordination and even for each country learning from the other’s best practices.

My reading of the evidence is that there are huge differences in the American and Chinese approaches. The US is focused on foundation models, such as frontier large language models. China is also investing – and arguably catching up – in these models, but there is much greater emphasis in China on applications of AI (with the government setting targets for the spread of AI across industries, https://t.co/1yimDjGmzK).

This reflects several factors, in my assessment. First, Chinese AI investments started with government demands for surveillance technologies and government-supplied data to AI companies (e.g., https://t.co/GXWDWukQty). Second, China is facing a huge demographic transformation, which will create scarcity of labor in the near future. This encourages companies to integrate robotics and AI much more actively into the production process (and China accounts for 59% of all robot installations in the world, https://t.co/YQjXLWZwrs). Third, applications and integration of these technologies into the production process requires a lot of engineering. China has a large and very well-trained supply of engineers (whereas the US has far fewer engineers and a lot of engineering talent is currently going to Silicon Valley, which hasn’t focused on applications). Fourth, China was relatively late in catching the “artificial general intelligence,” AGI, bug, although now there is more emphasis on this in China as well.

Whatever the exact causes, this difference in approach implies that there is plenty for the US to learn from China as well as the other way around.

The second aspect is that the zero-sum framing ignores or downplays the crucial need for global coordination. This framing claims that the strategic and military advantage that powerful AI will create is all that matters. But the race to AGI also poses huge risks (and this is true regardless of whether this race is truly advancing toward superintelligence or just producing distorted intelligence, as I argued in my first post in this series). And these risks are global. If the US and China are racing toward AGI with no guardrails and coordination, even the follower (probably China) will have models capable of creating havoc. In this case, racing without any global coordination is a recipe for disaster.
The third aspect of the problem is that this framing narrows the scope for any kind of coordination on other global issues between the US and China – including climate change, pandemics, nuclear nonproliferation and world peace. Tellingly, climate cooperation between the US and China has now largely stalled).

Let me add that I’m saying all of this not as an admirer of the Chinese system. On the contrary, I have been a vocal critic of both the Chinese Communist Party and the Chinese development strategy since Why Nations Fail (https://t.co/NMlp4VprG7) and I have argued that the Chinese success has some of its roots in severe labor repression (https://t.co/14SDaw2Y8H).

Indeed, this argument does not ignore the fact that China has an authoritarian, repressive regime, and it is using AI for surveillance and to stamp out dissent. Global coordination and efforts to set up reasonable rules that apply to all parties do not require the counterparty to be fully democratic or ruled by angels. In fact, sometimes it is more important to set up such a global framework when dealing with dangerous adversaries (as the arms-control treaties that the US and the Soviets signed demonstrate).

In sum, my assessment is that it is a colossal mistake to frame the AI race as a zero-sum one between the US and China. It poisons the much-needed dialogue and coordination between the US and China on broader issues; it makes the AI race more reckless; and it discourages the building of reasonable guardrails and regulations.

The last point is particularly relevant, and makes me sometimes say: if China did not exist, Silicon Valley would have to invent it.

[Daron Acemoglu @DAcemogluMIT]

And this

[60 Minutes @60Minutes]

Economist Daron Acemoglu says AI should be designed to help people do their jobs better, not make them expendable. He also suggests interventions like a tax code that incentivizes hiring real people. https://t.co/uFxSKrVli9

[Daron Acemoglu @DAcemogluMIT]

Dear followers,
I’m delighted to share this interview with 60 minutes on AI and jobs. Thank you

[60 Minutes @60Minutes]

Nobel Prize-winning MIT economist Daron Acemoglu fears unemployment could triple over the next decade if nothing is done and we stay on our current course. https://t.co/4sahxyfsxj

[Daron Acemoglu @DAcemogluMIT]

And please see this somewhat more optimistic take on my third question from @joshgans

[Joshua Gans @joshgans]

Today, I am going to look at @DAcemogluMIT 's third question on AI.

It is about the AI investment boom and its consequences. There is a lot to think about with regard to this, but Daron focuses on: "whether the AI boom can continue without leading to a massive increase in inequality."

The premise this question is based on is whether an AI boom will continue and earn 10 per cent or more to its investors per annum over the foreseeable future. Daron doubts these returns will be realised (competition, speed of adoption, etc., all constrain it), and I share those doubts. But what if the dreams of investors are realised?

Stijn Van Nieuwerburgh has crunched the numbers (https://t.co/w1BiW6oBVL), and it is daunting. The revenues required for a 10 per cent return would take up over 10 per cent of national income per annum. (Not because they are the same thing; that is just how this kinda worked out.)

Daron worries about where those returns will go. He argues they will go to capital owners, and if that happens, there will be a massive rise in wealth and income inequality.

But I don't think you can answer this question without first asking where the revenue will come from. AI investments can create AI, but the revenue still has to come from somewhere.

One possibility is that it comes straight out of the labour share of income because AI replaces jobs with machines, and then we get a straight-out redistribution of income.

Another possibility is that it comes from higher-value goods and services and, more generally, from higher productivity. Some of that productivity dividend goes to capital owners but, at the same time, it can also go to labour income, and it can also come to people in the form of consumer surplus.

How do we delineate between the two in terms of plausibility? Well, for the straight-out redistribution story where the machines replace the people, the 10 per cent comes out of the labour share. The US labour share has been declining and is now about 54% of income. If the 10 per cent comes out of it, that falls to 44%.

That sounds bad. But does it make sense? The trillions of dollars in AI investment didn't just pop out of nowhere. I would gather they do not come out of the luxury yacht market either. Instead, it came from other investments -- that is, the part that was producing the existing 46% capital share. Thus, what we have done here is potentially boost the return on capital by some amount. How much? Over the past decade, the NASDAQ has given investors an 18 per cent rate of return. This makes you wonder if Van Nieuwerburgh's 10 per cent return is reasonable. That said, that is 'real money' rather than 'book money.'

For the growth scenario, with a 1 per cent boost to growth per annum, I calculate that the labour share would fall to about 47.5% a decade out. So about half the loss that the redistribution scenario has. This is because growth compounds, which softens the impact.

But we can go further. Keeping the 10 per cent return (I know it's not the most reasonable ceteris paribus here), if the boost in growth was 2.1% per annum, then the labour share would be unchanged in a decade, allowing the capital owners to get their 10 per cent due. [By the way, the whole 10 per cent number is doing lots. It is the rate of return, and it is also the share of income going to AI investors, but that is just some funny miracle, and no one is really doing anything more than back-of-the-envelope scratches here.)

I am not so optimistic that such a growth rate can be achieved, but at the same time, I suspect that the 10 per cent return assumption hinges on that. If it is a lower rate of growth, the premise of Daron's question also doesn't hold.

So my answer, therefore, is that there is a good argument that we can do this without the massive increase in inequality. Simply because AI might be a lucrative investment isn't enough to cause us to worry, precisely because we need to answer why it is a lucrative investment in order to provide a more fulsome picture.

[Daron Acemoglu @DAcemogluMIT]

I think this is well worth reading.

[John Tasioulas @JTasioulas]

https://t.co/M1VTVvB7Oc

[Daron Acemoglu @DAcemogluMIT]

Reactions on my second question from @joshgans

[Joshua Gans @joshgans]

My response to @DAcemogluMIT 's second question on AI:

This is a tough one to answer, in part, because Daron doesn't quite ask it. There are sub-questions at the end.

So let me try to pose it: if AI is going to be so transformational, how do we ensure that it is governed and regulated in a way that takes into account the values and preferences of all people?

That is an important question. Indeed, it is precisely the question at the heart of political economy writ large. Its most salient branch appropriate for today's AI moment is whether we leave this to the experts (Daron calls them the technocracy) or let the people have much more of a say than they seem to have now.

Daron argues, not surprisingly, that we should let everyone have a say and, more controversially, that we can do this. The trade-off is familiar. The technocracy has important knowledge but is no angel. The people are more sensible and able to grasp the issues (at least in aggregate) than we often give them credit for. To this last part, I would add another: the people will be happier to buy into things where they have a say.

My issue with Daron's answer is not his analysis but instead the way it implicitly confines the questions and the options. In reality, it is not about this versus that but instead about delegation.

To give people their due, I think (in aggregate) they are a good judge of when they don't know enough. The Technocracy, should that be a thing, could exist as the appointment of a few rich people, say, or it could exist as the appointment of the elected representatives of the people. When it comes to addressing market failures -- which is basically what the issues with AI that we worry about mostly are -- then democratic values hand the management of the experts to elected officials. In the market, that management is handed to, let's lean in here, the owners of capital. But even that permission is supposed to be on a leash that is ultimately held by the people.

That is how things ought to be. They are far from it, and so we are already in a second-best world for considering these questions. Basically, the people have less say (especially worldwide), and the owners of capital seem to have more say, although even amongst their lot it is hardly a free and open club.

We can lament this all we want, but the real question is how we would push the needle here with respect to AI. Many countries around the world are doing that by, ahem, appointing a panel of experts to advise governments on how to do it. Canada's was announced yesterday (https://t.co/OUOJgDmejx).

This makes me wonder whether this question is really a problem for AI. My -- surely very uninteresting take -- is that governments are on it, slower than we might have liked, and will be influenced more by political interests than we might have liked, but ultimately will do things that avoid disaster and allow the voice of the people to influence the process. This is at least for democracies. For the others, I am left with the usual but useless conclusion that it would be better if they were democracies.

[Daron Acemoglu @DAcemogluMIT]

Fifth question on AI.

AI is currently viewed as a centralizing technology – with ever-larger models, trained on much of the world’s knowledge, running on massive, concentrated compute, overseen by a few companies, and aiming to control and deploy all of that knowledge

But many early computer scientists and pioneers thought of digital technologies as powerfully decentralizing, as Simon Johnson and I discuss in our book Power and Progress (https://t.co/WROlEDnvYx).

One such pioneer, Lee Felsenstein, wrote “The industrial approach is grim and doesn’t work: the design motto is Designed by Geniuses for Use by Idiots.” In the words of one of his collaborators, Bob Marsh, “We wanted to make the microcomputer accessible to all human beings.” And this wasn’t just an aspiration; they believed it was both feasible and natural that computers would empower individuals.

The first newsletter of People’s Computer Company also captures this optimism:

“Computers are mostly; used against people instead of for people; used to control people instead of to free them; time to change all that – we need a…PEOPLE’S COMPUTER COMPANY”.

Should we accept centralizing AI? Or is there a way in which the promise of artificial intelligence can be realized while data, capabilities and control are much more decentralized both across domains and across people? Wouldn’t decentralized AI be more consistent with human agency and control?

(Some of the ways in which greater decentralization can happen are discussed in my first post in this series.)

[Daron Acemoglu @DAcemogluMIT]

Fourth question on AI.

One of the great promises of AI is the discovery of new drugs and cures, so that we can live longer and healthier lives. If and when this becomes a reality, it would indeed be a big achievement.

These potential health benefits are often invoked to justify current investments (and lack of regulations). Some even argue that slowing down AI would harm humanity by delaying these benefits. (See https://t.co/3D0lTFCwmA).

This raises another uncomfortable question: can we justify rapid and large AI investments based on health benefits?

Here is why I think this is an uncomfortable question.

First, despite significant effort, AI has so far produced few gains in drug discovery or new cures. A fascinating new paper by Ryan Hill and Carolyn Stein documents significant (downstream) scientific work on proteins whose structures AlphaFold predicted. But the authors conclude “we find no evidence so far that more applied, early-stage drug development is targeting these proteins.” (See their paper here https://t.co/yWHpDCNgOF).

This thought-provoking post by Daphne Koller explains the difficulties that AI is currently facing in drug discovery. Koller emphasizes that AI-based research is focused on the last layer of drug development (creating new molecular entities for already well-understood therapeutic modalities), rather than targeting new disease mechanisms (see https://t.co/uQMWlQ02Ul). Understanding disease mechanisms is harder because there is much less data on it, and this endeavor requires more innovative approaches. It remains an open question whether current AI models are capable of doing this.

Second and more importantly, if you want to improve life expectancy and health in the United States, there is plenty of low hanging fruit and diverting some of the huge investments going to AI for this purpose would likely do much more for health outcomes.

The United States has the lowest life expectancy at birth among large rich economies. Americans live about five years less, in expectation, than the citizens of Switzerland, Sweden, Japan, Italy, South Korea and several other industrialized countries, and most of the gap comes from deaths before age of 70 due to chronic diseases, overdoses and other preventable causes (see, for example, https://t.co/LndeZHpj60).

These largely reflect failures in US public health: preventable problems, unhealthy diets, insufficient immunization rates, poor access to primary care; and poor information about health. The country spends much more than other peer economies on healthcare, but too little and too ineffectively on public health (here is a reference to my own work on this: https://t.co/VZI2JbO4Gf).

Could we, and should we, divert some of the massive investments in AI towards public health if what we want is better health and longer lives for Americans? (And yes, new drugs will benefit citizens of other countries as well, but they can also invest more in public health and the bulk of global AI spending is currently in the United States).