Liberalism Died of the Educated Elite, Not Automation
Acemoglu dismantles three myths: liberalism is dying of the educated elite rather than automation; AI productivity will not break out for a decade; regulating algorithms is not a restriction on speech, and pro-worker AI is the direction.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Conflicts of value cannot be settled by converting everything into utility
Asked why he does not use utility or cost-benefit analysis to adjudicate conflicts of value, Acemoglu says such methods would themselves require a whole book to specify how everything gets converted into utility and who decides what counts as an externality. He rejects both the absolute priority of a Rousseauian general will and the common denominator of utilitarianism, arguing instead that on individual liberty as the single foundation, one returns to the social level to seek a provisional consensus. Morality, he holds, evolves gradually and under uncertainty, just as science does; it is not a fixed external standard.
— AcemogluThe problem is not that elites have power but that technocracy is complacent
Acemoglu traces the crisis of liberalism to two forces: the post-industrial economy has made the educated elite both more powerful and more numerous, while a philosophical turn has assigned them higher priority. He warns that any hierarchy drawn along lines of education, race or income obstructs real freedom; although he himself values expertise, the United States and Europe have developed a complacent technocracy that looks down on ordinary people. What is needed, he stresses, is to avoid that technocracy, not simply to strip elites of power.
— AcemogluAutomation's first-order effect is to push down the labor share
Responding to the counterexamples Tyler cites, Acemoglu insists on the micro-level mechanism: automation converts labor tasks into capital, so its first-order effect is to lower the labor share. That does not mean wages or employment must fall, because new tasks can absorb labor, but it is very hard to recover the original share. He cites his own robots paper: blue-collar workers displaced by robots in the 1990s and the 2000s, in jobs such as welding and painting, saw clear wage declines. He objects to treating technology as a black box and argues for distinguishing automation from the creation of new tasks.
— AcemogluEven a free-speech absolutist should want algorithmic recommendation regulated
When Tyler draws an analogy between social media and the printing press spreading Marx and Lenin, Acemoglu replies that human beings have already adapted to carriers like books, but an algorithm can pinpoint Tyler's weaknesses precisely and fan the flames, which is an entirely different world. So he calls himself a free-speech absolutist while supporting regulation of social media's algorithmic recommendation, because he is not restricting what Tyler says, only how the platform exploits Tyler's words. He proposes non-algorithmic timelines, such as showing the accounts you follow in chronological order, as a possible way out.
— AcemogluAmerica has already tacitly given up on half its secondary students
Tyler pushes back on the charge that education has been abandoned by pointing to New York City's high per-student spending. Acemoglu answers that roughly half of American secondary students read and write several grade levels behind, and society as a whole has come to accept this by default. In his view the elite class fills the gap by importing skills rather than being willing to educate its own people, and this is precisely the rupture in the "social contract" he criticizes. He argues for updating that contract so that politicians, journalists and technology leaders genuinely commit to educating Americans, and for restructuring teachers' unions that have become ideological and dysfunctional.
— AcemogluAI will not be a game changer within the next ten years
Acemoglu concedes that if he rewrote his 2023 paper today he would raise his productivity estimate for AI, because models have advanced faster than expected, but progress at the application layer is slow, which drags out adoption. He expects one or two years of rapid growth over the coming decade, with a limited average growth rate: not the game changer McKinsey or OpenAI imagine. On Tyler's proposal of at least half a percentage point of growth, as in the IT period of 1995-1998, he says it will at least be no worse than that, while refusing any utopian or doomsday estimate.
— AcemogluPro-worker AI comes from changing the narrative, not from appointing a regulator
Asked what regulatory form pro-worker AI would take, Acemoglu explicitly opposes setting up a regulator to decide which kinds of AI are permitted. The pro-worker AI he has in mind lets workers draw on information and tools to do new and complex things, expanding the capabilities of nurses, journalists and engineers. Cultivating it requires changing the narratives and ambitions of technology companies; public policy is only a supporting player. Automation that merely replaces tasks is pro-capital rather than pro-worker, whereas new machines combined with information systems and retraining are the direction to go.
— AcemogluEconomists use AI to save labor, not to generate new ideas
On whether AI will raise his own output of papers, Acemoglu says economists at present use it only to save labor; nobody is using it to generate ideas they could not otherwise have had. He grants its value in cleaning data, in meta-analysis, and in turning qualitative historical material into quantitative data, but it has not yet delivered a breakthrough. He offers the Beatles and Beethoven as examples that quantity and quality often go together, so writing more papers raises the probability of top-tier output, but as things stand AI has not changed how he innovates.
— AcemogluIn their own words · checked verbatim
I think non interference.By itself, isn't enough.
Acemoglu6:27
We cannot have true freedom.And through flourish ing of a population.When there is a clear hierarchy, whether it is by education, it is by race, it is by income.
Acemoglu11:35
But once.Algorithism can see how Tyler has a weakness for this specific thing. I can. Fund the flames.I think that's a completely new world.
Acemoglu25:00
I completely reject the crazy utopian or dystopian estimates.
Acemoglu34:56
I do not believe that we need to have regulator that says oh only pro work AI is allowed absolutely not
Acemoglu41:46
Figures
| Share of US labor-market tasks accounted for by the creation of new tasks | 40% to 50% | 19:23 |
| Annual H-1B visa cap | 85,000 | 27:37 |
| Acemoglu's probability for the scenario in which the AI stack does not empower foundation models | about 30% | 37:58 |
| Brazil's total fertility rate (mentioned by Tyler) | about 1.36 | 45:00 |
Glossary
- non-domination
- A republican ideal: the individual is free from the arbitrary power of others, which requires the protection of the community.
- task-based framework
- A theoretical framework that breaks production down into tasks and studies how capital and labor divide them.
- pro worker AI
- Artificial intelligence that augments workers' skills and problem-solving capacity rather than simply replacing labor.
- agentic AI
- AI able to plan and execute multi-step tasks autonomously; Acemoglu believes it brings greater uncertainty.
How to listen
Economics researchers who hold a position on the relationship between automation and the labor share, investors skeptical of the AI productivity narrative, and public policy practitioners who care about social media regulation.
The stretch on Armenian food and travel advice is small talk and can be skipped; the rest is core argument.