The problem they were solving

Before this work, the dominant model was skill-biased technical change (SBTC). Technology raises the productivity of skilled workers more than unskilled ones; inequality rises because the skill premium rises; everyone’s wages still rise, just unevenly. It fit the 1980s well.

It fit the 2000s badly. Real wages for non-college men fell. Employment polarized — growth at the top and bottom, hollowing in the middle. SBTC has no mechanism for a group’s wages going down in absolute terms, because in that model technology always raises the marginal product of labor.

The task framework’s core move: technology does not augment workers, it reallocates tasks. If a task moves from labor to capital, the workers who specialized in it lose the task itself — not merely relative standing. Their wages can fall absolutely. That is the thing SBTC could not produce and the data demanded.

The three effects

“So-so automation” is their term for the worst case and it is the key concept for the AI argument: technology just good enough to replace a worker, not good enough to generate much productivity gain. Self-checkout is the standard example. You get displacement with a small productivity effect and no new tasks — the mix Acemoglu fears AI is producing.

The papers, in order

The robots number

The headline result from the 2020 JPE paper, using variation in robot adoption across US commuting zones 1990–2007:

One additional robot per thousand workers reduces the employment-to-population ratio by about 0.2 percentage points and wages by about 0.42%.

Reported ranges across specifications: −0.18 to −0.34 pp on employment, −0.25% to −0.5% on wages. The identification rests on areas most exposed to robots after 1990 showing no differential trend before 1990.

How the position evolved

This is the part worth noticing, because it is not a straight line.

1. 2011–2018 — building the machine. Establish that tasks are the right unit and that displacement is real.
2. 2018–2022 — the pessimistic empirical phase. Robots lower wages. Task displacement explains most of rising inequality. This is the Acemoglu most people know.
3. 2022 onward — the conditional turn. Demographics and Automation shows automation triggered by labor scarcity is good. The 2026 Baby Busts paper finds falling birth rates associated with higher GDP growth per working-age adult, via labor-saving innovation. Cowen catches this and asks whether it contradicts the Restrepo work. Acemoglu says no, and he is right: both use the same framework. The framework was always conditional. Automation responding to scarcity is a different animal from automation replacing available workers.

That is why he can say “I am not, 100 percent not, against automation” without contradiction, and why he sounds pessimistic anyway. The model is neutral. His read of current conditions is not.

Where Cowen attacks

Cowen’s move in this conversation is to grant the micro and deny the macro: fine at the firm and sector level, but the aggregate labor share barely moved — “62 to 60” adjusting for equity compensation — and automation since the Industrial Revolution has plainly made workers richer. Acemoglu’s reply is that the macro series is not a test of automation because it bundles automation with everything else, including new task creation, which he estimates accounts for 40–50% of what the US labor market currently supports. Full treatment in Automation and the Labor Share.

From the book

What Happened to Liberal Democracy? carries this machinery without the notation, mostly in Chapter 6, Pathway to Crisis, and it supplies the numbers the interview only alludes to.

The 50–70% figure is in the book, not just in the Econometrica abstract:

“Recent estimates suggest that between 50 and 70 percent of the surge in U.S. wage inequality over the last four decades can be accounted for by the effects of automation, and that much of the decline in the wages of low-education demographic groups, such as young men with a high school degree or less, can be explained by the spread of automation tools.”

And the robot adoption series it rests on: the US went from about 3.5 robots per 1,000 industrial workers in 1993 to 15 in 2014 and 25.5 by 2020. Germany reached roughly twice the US density.

The displacement mechanism, stated plainly — this is the task framework in prose:

“Recall that in the industrial compact, as companies reached larger markets and increased their production volume, they also hired more workers, generating both jobs and, for your labour market competition, higher wages. With automation, this latter link was broken. Firms could expand and output could multiply without any need to hire more workers.”

The carmaker claim is verified — with corrections. The example that this page called “doing enormous work in his argument” and worth its own page in draft 2 is in Chapter 6, and three details in the interview version don’t survive:

The core of it holds, and is stronger than the interview version because the comparison is embarrassing for the simple technological story:

“German manufacturing firms have introduced industrial robots considerably faster than their American counterparts… But German manufacturing companies must negotiate with worker representatives and unions when making these choices… robot adoption in Germany was not associated with large employment drops. Blue-collar work declined, but there were concurrent increases in technical occupations.”

More robots, fewer layoffs. And the within-Germany variation is the part that makes it evidence rather than anecdote: “in establishments where labor was better organized, there were less steep employment declines, and more pronounced efforts to reallocate blue-collar workers to technical tasks.” Same country, same technology, same decade — the difference is who was in the room. That is the strongest empirical support for the “choice” claim anywhere in the conversation, exactly as this page suspected.

On the American side, the book gives the autos numbers the interview does not: the sector paid “about 40% higher wages to Americans without a college degree than the rest of the economy,” and employment of workers with a high-school degree or less fell from 625,000 in the 1960s to 371,000 by 2010.

Working notes

Confidence. The framework, the paper list, and the robots coefficients are solid and checkable. The 50–70% figure is from the Econometrica abstract. The 40–50% new-tasks figure is Acemoglu’s spoken number in this interview, attributed partly to Autor’s The Work of the Future — I have not verified it against a printed source and it should not be quoted as precise.

⚠ Unverified: the Barany, Patel and Siegel paper. Cowen cites CEPR DP21700 on France through 2019 finding no labor-share harm. CEPR blocks automated fetching (HTTP 403) and the paper does not surface in search. There is a related January 2026 paper by Aseem Patel, “Labour Market Power and Automation”, using French administrative data — not the same paper, and I should not pretend it is. Someone with CEPR access should pull DP21700 and check what it actually claims. Acemoglu’s on-air response was “I don’t know this new paper,” so the exchange is unresolved on both sides.

The most useful unasked question — now answered. Draft 1 said the Japanese/Korean/German carmaker example was doing enormous work in his argument, was unverified, and would be the strongest empirical support for the “choice” claim if it held up. It holds up, in Chapter 6, with three corrections: no Korea, German manufacturing rather than carmakers, and retraining under Industry 4.0 rather than job redesign around robots. Written up above. It is still worth its own page in draft 2 — more so now, because the within-Germany establishment-level variation is a cleaner identification than the cross-country comparison the interview offers.

The same claim appears on Working-Class Liberalism and Community Elbow Room with the same corrections applied.

Still unverified after the book check: the 40–50% new-tasks figure (not in the book) and the Barany–Patel–Siegel paper above (not in the book either — it postdates it, and Acemoglu says on air he hasn’t read it).

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech. Numbers transcribed by machine deserve particular suspicion — the robot-density and employment figures above should be confirmed against print before republication.

Related: Automation and the Labor Share · AI Growth Forecasts - Whose Timeline · Induced Innovation and the Habakkuk Thesis · King and Plosser - Real Business Cycles

What the sources actually measure

A level gain is not an annual growth rate. A 1% GDP gain after ten years means output in year ten is 1% above the counterfactual level. It does not mean an extra percentage point of growth every year, or the sum of additional output across the decade. Total-factor productivity (TFP) concerns output relative to combined inputs; labor productivity concerns output per worker or hour. Investment and labor supply affect how those gains translate into GDP.

Sources: Acemoglu’s published paper, pp. 34, 38–39 and 42; Cowen’s conversation; Goldman’s April account and June clarification.

The correction: the earlier table used a roughly annualized TFP estimate as an annual GDP-growth increment. It also treated productivity benchmarks as GDP forecasts. That does not support either the $180 billion headline or the claim that Cowen’s estimate is ten times Acemoglu’s. Both are withdrawn.

Versions matter. The figures above use the published 2024 paper, rather than mixing its estimates with the April working draft. In the interview Acemoglu says he would raise the numerical estimates, but supplies no replacement number. The table records the 2024 exercise, not a quantified 2026 forecast.

The scenarios, compounded

To see what sustained growth would mean, separate the arithmetic from the attribution. Suppose an imaginary economy starts at $30 trillion, in fixed purchasing-power units, and grows at 2% annually for ten years. This is a round input, not a measured 2026 US GDP value.

Formula: 30 × (1 + annual rate)¹⁰. Differences are from the 2% baseline, in fixed purchasing-power units; they are endpoint differences, not accumulated gains.

These increments are chosen examples, not forecasts attributed to Acemoglu, Cowen, Goldman, or an AI company. In particular, +1.5 percentage points of productivity growth is not automatically +1.5 percentage points of total GDP growth.

For a cumulative GDP gain, the appropriate illustration is different: multiply the year-ten baseline by that gain. For example, a hypothetical 1% endpoint uplift on $36.57 trillion is about $0.37 trillion extra output in year ten. It is not $0.37 trillion accumulated over all ten years. The distribution of gains within the decade would be needed to calculate that total.

What counts as a game-changer?

There is no agreed numerical threshold in the exchange. Three questions help expose the disagreement:

• Size: how much more output, and measured in which units?
• Timing: a temporary acceleration, a delayed level shift, or persistently faster growth?
• Distribution: higher aggregate output for whom, with what changes to working lives?

The IT analogy can support both patience about adoption and caution about assuming a permanent boom. It does not by itself establish how fast AI will diffuse.

Reading

• Acemoglu, The Simple Macroeconomics of AI, published 2024 version — separate TFP and GDP calculations
• Goldman Sachs, April 2023 — the global GDP headline
• Goldman Sachs, June 2023 — annual productivity wording
• The original interview — Cowen’s benchmark and Acemoglu’s qualitative update

From the book

What Happened to Liberal Democracy? takes the same position and prints no number for it. Chapter 9 states it in words:

“My reading of the evidence is that the benefits from AI-based automation in the medium term, though non-trivial, are not huge. The rollout of AI will be slow, as is the rollout of any technology that requires major organizational changes.”

The quantitative estimates come from the paper. The book’s qualitative account does not validate a particular GDP conversion or a tenfold comparison.

The distribution argument is the one the book actually cares about, and it is where he spends his words:

“Even if such benefits were realized, how would they be distributed? The most natural path to broad-based benefits from productivity improvements would be via shared prosperity, working through higher wages for all sorts of work, and skills. This is what the Industrial Compact achieved until its unravelling in the early 1980s. But if all the gains are from automation, without the pro-worker AI possibilities I discuss below, then wage increases will not follow, and in fact, mass-scale automation could lead to significant joblessness.”

He then closes the obvious escape hatch — broad AI ownership, or universal basic income — on two grounds. Political economy: “Large tech companies would not be enthusiastic about redistributing all of their profits to the broader population.” And, more interestingly, a social objection that has nothing to do with money: “A society in which a large fraction of the population does not contribute to production or other beneficial activities would create huge social status gaps between the makers and the takers… Communities are most vibrant when their members feel they are contributing to society at large.”

This adds a separate question: who gains? A larger GDP estimate alone does not settle the distribution of wages, ownership, bargaining power, or opportunities to contribute. The growth arithmetic cannot rank these social outcomes.

On the AGI question underneath the timelines, the book is more explicit than the interview: “A quick transition to AGI seems unlikely, in part because existing models still do not show evidence of true comprehension or deep understanding, even in simple contexts” — the illustration being that a model can explain how to repair a garage door while having “no recognition of the social context of this problem,” and that “as far as they are concerned, there is no difference between repairing a garage door, summarizing an ancient text, and diagnosing cancer.” Whether that is a claim about current systems or about the architecture is left open, which is the same ambiguity the interview leaves.

Working notes

September 15 correction. Removed the unsupported GDP conversions and tenfold comparison. The illustrative table is generated from committed inputs in data/research/growth_scenarios.json; the short version also feeds the transcript popover. Reading packs are built from the same page.

Remaining limits. These source estimates are conditional exercises, not a common forecast. The paper’s introduction rounds one adjusted GDP result differently from its section 3.4 calculation; the table identifies the section used. No numeric replacement is inferred from the interview’s qualitative update. A forecast of explosive growth would require its own source, definition, and assumptions; the old arbitrary +10-point row is removed.

Source and its limits. The book material above retains the previous check against the audiobook edition, machine-transcribed. Chapter references and transcribed quotations still need confirmation against print before publication. This pass verifies the paper-based comparisons, not every book quotation.

Related: Acemoglu and Restrepo - The Task Framework · Automation and the Labor Share · Pro-Worker AI

What “first-order impact” means

Economists decompose an effect into orders. The first order is the direct arithmetic consequence; higher orders are the adjustments the system makes in response.

The labor share is compensation to workers divided by total output. Suppose a factory produces $100 of output, pays $60 in wages and $40 to capital. Labor share is 60%.

Now automate a task that had been done by a worker earning $10.

• First order: that $10 moves from the wage bill to the capital bill. $50 of $100 → labor share 50%. Nothing else has happened yet. This is what Acemoglu means by “That has a first-order impact on the labor share because fewer things are done by labor.” It is not a prediction, it is bookkeeping.
• Second order: production is cheaper, so the firm expands. Demand rises for the tasks still done by people. Some of the $10 comes back.
• Third order: new products and new jobs appear that did not exist. More comes back.

His claim, in one sentence: “there could be enough demand from nonautomated tasks for labor, but that would never come back to increase the wage enough to restore the labor share to where it is.” The second-order effects are real; they are just not complete.

Your reading, checked: do displaced workers earn less?

Yes — this is one of the better-established findings in labor economics, and it predates the automation debate.

The mechanism is specificity: a welder’s wage reflected skill in welding. Remove welding and the general labor market does not value that history. The worker is re-employed — often quickly — at a wage set by whatever they can do next.

Acemoglu’s precise version in this conversation: “the wages of people who used to be in especially blue-collar heavy manual tasks that were the ones that robots of the 1990s and 2000s went after, like welding and painting… went down quite a bit” — both in aggregate and in local labor markets.

The chart you asked for — and why it is a fight, not a line

Here is the US labor share, nonfarm business sector. Read the caveat below before using these numbers.

Source: BLS nonfarm business sector labor share, as reported in BLS TED and the FRBSF working paper.

The caveat is the substance. Cowen says “62 to 60”; the BLS headline series says roughly 66 to 57. Both are defensible, because the labor share is one of the most measurement-sensitive statistics in economics:

• Proprietors’ income. A sole proprietor’s income is part wage, part profit, and the split is a modeling assumption. Elsby, Hobijn & Şahin (2013) find one-third of the measured BLS decline comes from how this is handled.
• Equity compensation. Cowen’s adjustment. Stock grants to employees are labor income economically but land oddly in the accounts. Including them raises the recent labor share materially.
• Housing and depreciation. Imputed rent on owner-occupied housing is all “capital.” Rognlie (2015) showed much of the apparent capital-share rise is housing, not robots.
• Sector vs. aggregate. Acemoglu’s claim is explicitly at the firm and sector level, where he says evidence and theory “are very well aligned.” He concedes the macro series is contaminated: “those are macro things”, and there are “composition effects between like Walmart effects.”

So the honest framing is: Cowen and Acemoglu are not disagreeing about a number. They are disagreeing about which number is the right question. Cowen points at an aggregate that barely moved and says the worry is overblown. Acemoglu says the aggregate bundles automation with new-task creation and therefore cannot test his claim, which is about automation held separately.

From the book — which series Acemoglu himself uses

The measurement fight above has an obvious tie-breaker that draft 1 did not reach for: what number does Acemoglu print when he is not being interviewed? Chapter 6 of What Happened to Liberal Democracy? answers it, and the answer is neither the BLS headline nor Cowen’s.

“In 1980, the share of labour in US national income was 58%, with the rest going to capital. Since then, the national share of labour has fallen to 52%, and the share of capital has [risen].”

58 → 52, on national income, from 1980. Six points, not the nine the BLS nonfarm series implies over its longer window, and not the two Cowen offers. Three things follow:

• Cowen’s “62 to 60” is not a straw man of Acemoglu’s position, but it is not Acemoglu’s number either. The gap between them is 6 points against 2 — a real disagreement about magnitude, on top of the disagreement about which question the aggregate answers.
• The start date is doing work. Acemoglu begins at 1980, not 1947, which excludes the flat postwar stretch and starts the clock at the moment his own account says the industrial compact broke. That is a defensible choice and it is also an argumentative one — the same kind of endpoint choice flagged on The Printing Press - How Long Was the Adjustment.
• National income, not nonfarm business. A different denominator again, which is exactly why the table above needs its caveat.

The number he actually leans on is sectoral, and it is much larger:

“The labour share in value added in the manufacturing sector declined from 74% in 1981 to 46% in the mid-2010s, much larger than the decline in the aggregate economy that I mentioned previously, which was from 58% to 52%.”

74 → 46 in manufacturing. He then decomposes it in the direction his framework predicts: “While some manufacturing industries, such as apparel, had stable labour shares, the industries that were at the forefront of new robot installations, such as motor vehicles, chemical products, electrical equipment, and primary and fabricated metals, had sharper drops in their labour share and also cut down employment.”

That is the whole Cowen–Acemoglu exchange in miniature, and it clarifies who is arguing what. Acemoglu is not defending the aggregate series. He publishes it, calls it small, and then goes immediately to the sector where the effect is four times larger and where robot adoption and labor-share decline line up industry by industry. Cowen’s move — point at the aggregate, note it barely moved — lands on a number Acemoglu has already conceded is not the one carrying his argument.

For the chart. This resolves the “which series” decision the working notes below hand back to Adam, at least for panel 1: if the page is about Acemoglu’s claim, the honest series is the manufacturing labor share in value added, with the aggregate plotted beneath it for scale. That is two lines from published sources, it shows exactly why the two men are talking past each other, and it does not require anyone’s replication files.

Reading

• Acemoglu & Restrepo, “Automation and New Tasks”, JEP 2019 — the accessible statement
• Acemoglu & Restrepo, “Robots and Jobs”, JPE 2020 — the empirics
• Elsby, Hobijn & Şahin, “The Decline of the U.S. Labor Share”, BPEA 2013 — why the measurement fight is real
• Autor, “The Work of the Future” — the new-tasks accounting Acemoglu defers to
• Jacobson, LaLonde & Sullivan (1993), “Earnings Losses of Displaced Workers,” AER — the origin of the scarring literature

Working notes

On the chart. What Adam actually wants is a two-panel figure: labor share over time on top, and below it the wage path of workers in heavily-automated occupations versus everyone else. Panel 1 I can source. Panel 2 does not exist as a clean public series — it has to be constructed from the Econometrica paper’s occupation-group decomposition, which means pulling their replication files. That is a real task, not a lookup. Flagging it rather than faking it.

The table above is the honest first draft: numbers with provenance, and the reason a single line would be misleading. If you want a rendered chart in draft 2, the decision to make is which series — and that decision is the argument, so it should be yours, not mine.

Numbers I’d want double-checked before publication: the 1947 Q1 65.8% and 2000 Q4 62.8% figures come through secondary sources quoting BLS, not from BLS directly. The Elsby/Hobijn/Şahin one-third result is from their abstract.

What I could not resolve: Cowen’s “62 to 60, adjusting for equity compensation.” I could not find the specific series he’s using. It is plausibly Barkai (2020) or a Mercatus-adjacent calculation. Until that’s identified, treat the 62→60 as a claim in an argument rather than a fact. Still unresolved — the book check settles what Acemoglu’s number is, not what Cowen’s is.

What the book did settle: which series Acemoglu himself publishes (58→52 on national income since 1980, and 74→46 in manufacturing value added since 1981), and therefore which line panel 1 of the chart should be. Written up above. It also means the “62 to 60” exchange is not two people disputing a measurement — it is Cowen testing an aggregate that Acemoglu’s own book already sets aside as too small to carry the argument.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and the figures above are machine-transcribed numbers. Confirm 74/46/58/52 against print before publication — a transcript is exactly the wrong place to source a number to the percentage point, and these four are now load-bearing on this page.

Related: Acemoglu and Restrepo - The Task Framework · AI Growth Forecasts - Whose Timeline

The finding to know first

Eric Hanushek and Ludger Woessmann, The Knowledge Capital of Nations (MIT Press, 2015):

• Substitute cognitive skills measured by international tests (PISA, TIMSS) for years of schooling, and the fit to long-run growth improves dramatically.
• Years of schooling turns out to be a poor proxy. Two countries can each average 11 years and differ enormously in what students learn — Armenia is a case in point, see below.
• Their headline claim: the “Latin American growth puzzle” and the “East Asian miracle” both dissolve once you condition on skills. Neither was mysterious; both regions grew about as fast as their knowledge capital predicted.

That last point reframes Adam’s question. The interesting outliers are not “who over- and under-performs” in general — most countries are near the line. The interesting outliers are the ones that stay off it after conditioning.

The two kinds of outlier

Over-performers — income above what skills predict

Under-performers — income below what skills predict

Armenia, which is where the question came from

Cowen asks whether Armenia can work, given Russia. Acemoglu — who is Armenian — gives a blunt answer: “It has disappointed. Armenia has really disappointed. It has underperformed economically.” His explanation is almost entirely geographic and geopolitical, not educational:

“It’s a landlocked country. It cannot export anything because, again, transport is not a possibility. It has to find a way of using its talent.”

The data complicate the story in an interesting way:

Sources: World Bank Armenia learning analysis · TIMSS 2023 Armenia · World Bank on Armenian education and growth

This is the nuance worth carrying. Both Cowen and Acemoglu treat Armenian human capital as a given — “the people are so smart,” “the people are well educated.” That is the Soviet inheritance talking: Armenia was the USSR’s engineering and technical hub, and the reputation is real. But the measured learning of children in school today is weak, and the 11.3 → 8.0 adjustment says most of the gap is quality, not access.

So Armenia may be a smaller outlier than it looks. Part of the underperformance is stranded geography, as Acemoglu says. Part of it is that the knowledge capital is a legacy stock that is not being replenished — which is exactly what the Hanushek–Woessmann framework would predict from those test scores. Neither man raises this, and it slightly undercuts the “smart people, bad geography” story they agree on.

Reading

• Hanushek & Woessmann, The Knowledge Capital of Nations (2015) — the core work
• Hanushek & Woessmann, “Education and Economic Growth” (2021) — a compact survey version, free
• Pritchett, “Where Has All the Education Gone?” (2001) — the paper that broke the years-of-schooling proxy
• Acemoglu, Gallego & Robinson, “Institutions, Human Capital and Development” (2014) — Acemoglu’s own position: institutions dominate, and schooling’s apparent effect is partly institutions in disguise

Working notes

The interesting tension, and I’d want your read on it. Hanushek and Woessmann say cognitive skills drive growth. Acemoglu’s own published work says institutions drive growth and human capital is substantially downstream of institutions. These are rival explanations of the same correlation, and Acemoglu is a principal in the fight. In this interview he answers the Armenia question in pure institutions-and-geography terms without mentioning schooling quality at all — which is exactly what you’d predict from his research program, and it’s a nice illustration of how a scholar’s framework shows up in an off-the-cuff answer. That observation is mine and is worth a page of its own in draft 2.

⚠ What is not properly sourced here. The over/under-performer tables are constructed from the general shape of this literature, not lifted from a specific published residual analysis. They are directionally right and individually defensible, but there is no single table in Hanushek–Woessmann that looks like mine. If you want this to be rigorous rather than orienting, the move is to reproduce their growth regression and rank the residuals — which is genuinely doable with public PISA and Penn World Table data, and would give real answers instead of archetypes.

Also unverified: the specific Gulf-state PISA-versus-income claim, and the Argentina characterization, are conventional wisdom in this literature rather than things I checked today.

What draft 2 needs: a scatter plot — test score on x, GDP per capita or growth on y, with the residual outliers labeled. That is the picture the question is really asking for, and it’s buildable from public data. Armenia should be plotted on it.

Checked against the book — and the tension above is confirmed by absence. What Happened to Liberal Democracy? contains no PISA, no test-score data, no growth regression on cognitive skills, and no engagement with Hanushek and Woessmann. Armenia appears once, in an unrelated aside about inoculation. So the observation this page makes — that Acemoglu answers a schooling-quality question in pure institutions-and-geography terms — is not a quirk of an off-the-cuff interview answer. It is what his book does too: education enters Chapter 1 as a public service failing to convert spending into outcomes (see NYC vs London - School Spending) and never as a driver of growth. The rival-explanations framing stands, and the page can now say so about his stated position rather than inferring it from a research programme.

Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed; references are by chapter. A negative finding from a transcript is weaker than one from an index, though a literature engaged nowhere in twelve hours is not one being quietly relied on.

Related: NYC vs London - School Spending · Acemoglu and Restrepo - The Task Framework

The Hayek argument in one page

“The Use of Knowledge in Society” (AER, 1945) — one of the most cited papers in economics, and short.

Hayek’s target was the socialist calculation debate. Opponents had conceded that a planner would need vast information but argued that with enough computing power the problem was solvable. Hayek’s response was that this misdescribes the problem:

• The relevant knowledge is “knowledge of the particular circumstances of time and place” — this machine’s quirks, this customer’s reliability, this shortage next week.
• Much of it is tacit: known in practice, not statable in a form that could be transmitted to a center.
• It is created and destroyed continuously, so a snapshot is worthless.
• Therefore the planner’s problem is not computational scale. The information does not exist in transmittable form at all.

The price system solves this without solving it: a rise in the price of tin tells everyone who uses tin to economize, and none of them needs to know why. Prices are a compression of knowledge nobody possesses in full.

How Acemoglu uses it — and why that’s unusual

From the transcript:

“this relates to both Hayek and to some of the themes that I try to develop in What Happened to Liberal Democracy? … I believe that liberalism and the kinds of society that we want to build that supports freedom does need decentralization. Excessive centralization of many sorts is inimical to that agenda.”

This is a left-liberal Nobel laureate deploying the canonical anti-planning argument against a technology, not against the state. His claim:

“The large language model approach is centralizing. Both the philosophy and the practice of it is that you take all of the data that you can, and one model rules it all.”

And his alternative is explicitly Hayekian in shape: “true decentralization would be much better served with more domain-specific models of AI and other approaches that really use high-quality data for specific things.” Many models holding local knowledge, rather than one model holding everything — which is close to saying that the tacit, particular knowledge Hayek described should stay where it lives.

Note he is undecided about the empirics. Cowen’s challenge is strong — open-weight models, cheap distillation, an extraordinarily competitive market — and Acemoglu gives it “probably 30 percent or so probability, just making it up.” He is honest that this is judgment, not a result. His counter-evidence is behavioral: if there were no winner-take-all dynamics, the US and China would be collaborating and OpenAI and Anthropic would not be competing so hard. Valuations imply that participants believe otherwise.

The tension worth holding

Hayek’s argument was about knowledge that cannot be centralized in principle. Acemoglu’s worry is about power centralizing in practice even if knowledge diffuses. Those are different claims, and Cowen puts his finger on exactly that gap:

“Maybe to the knowledge, but not the power, right? The gains will go to the users, just as they have for electricity, antibiotics, fire.”

If LLMs actually do capture dispersed tacit knowledge and hand it back to everyone cheaply, they are Hayekian instruments, not Hayekian violations — a better price system, not a planner. That is the strongest form of Cowen’s position and Acemoglu doesn’t dispose of it. He falls back on market structure and valuations rather than on the epistemics.

From the book

The appropriation is not rhetorical. Hayek is one of the most-cited figures in What Happened to Liberal Democracy?, worked through three separate times, and Acemoglu is careful to mark where he takes from him and where he parts company.

Chapter 2 takes the epistemics. Hayek appears alongside Mill and Peirce as a source for the fallibility premise — “it was highly uncertain and, for all practical purposes, unknowable in advance how new knowledge would be created” — and again on social learning and on knowledge as the engine of civilization. The Cromwell line Hayek used as an epigraph in The Constitution of Liberty, “Man never mounts higher than when he knows not where he is going,” is quoted as the shared core: “Liberalism is about exploration, experimentation, and freedom to do new unconventional things.”

Chapter 3 states the divergence explicitly, which is what settles the question this page was holding open:

“despite some parallels between Hayek’s philosophy and the premises I use to justify liberalism in chapter two, Hayek and I end up in quite different places regarding several key features of liberalism.”

Four reasons follow, and the first is the structural one: “Hayek did not link his ideas to the notion of non-domination and thus to issues of self-government. Rather, he saw them as an affirmation of the classical liberal tradition’s emphasis on non-interference, and he remained suspicious of democratic processes.” The consequence Acemoglu draws is that Hayek could not see that “significant inequalities and inability to engage in self-government could be major impediments to true freedom.” The turn of the knife is the third reason: Hayek was right that an administrative state’s rules metamorphose into strictures and crowd out experimentation, “yet he did not note the possibility that a very unequal distribution of economic power and opportunities could also lead to similarly adverse consequences for the accumulation of both scientific and social knowledge.” Hayek’s own argument, pointed at private concentration. That is the move the interview is compressing, and it is developed at length rather than borrowed for the occasion.

Chapter 9 is where it becomes the AI argument. The chapter opens the section titled Information Centralization with exactly the framing Cowen challenges:

“Information centralization is foundational to generative AI, which aims to collect and use most of the information available to humankind. Is this desirable or even feasible?”

Then the naive reading, stated fairly before it is refused — “once we have powerful enough computers to encode and process the vast amount of dispersed information, central planning could perform better… perhaps this can be done by AI models, or the companies that control AI models” — and the Hayekian reply, which is the sharpest form of his position anywhere:

“What collective knowledge building needs isn’t computation or aggregation based on existing data, but to produce new data. This is what human experimentation is all about… Hence, without human experimentation, there may not be the right kind of useful information for AI to aggregate.”

He then quotes Hayek twice more, including the passage that does the most work for him: “if we can agree that the economic problem of society is mainly one of rapid adaptation to changes, then the ultimate decisions must be left to the people who are familiar with these circumstances. We must solve it by some form of decentralization.”

On the tension this page identifies, the book is better than the interview but does not close it. Cowen’s objection — that LLMs might be Hayekian instruments rather than violations, diffusing knowledge to everyone — gets a partial answer in print that never surfaces in conversation: the claim is not just that centralization fails to aggregate, but that it degrades the production of what there is to aggregate. “There is evidence that large language models, such as ChatGPT, are already discouraging certain types of human experimentation and knowledge creation. People become less engaged in posting questions and answers on discussion forums… Views and new content on Wikipedia also appear to have declined.” That is an empirical claim with a testable shape, and it is a real reply rather than the appeal to market structure and valuations he falls back on with Cowen. It also concedes the ground Cowen is standing on — “None of this is to belittle the potential contributions of AI to science” — and the “30 percent or so probability, just making it up” from the interview has no counterpart in the book, which argues the case without pricing it.

Reading

• Hayek, “The Use of Knowledge in Society” (1945) — ~15 pages, read it
• Hayek, The Road to Serfdom (1944) — the popular argument
• Lange & Lerner, the market-socialist position Hayek was answering
• Acemoglu & Robinson, The Narrow Corridor (2019) — his own account of why concentrated power threatens liberty, from the previous Conversations with Tyler appearance

Working notes

The interesting essay here is not “what did Hayek say” but the appropriation itself: Hayek is normally a right-liberal totem, and Acemoglu is using him to argue for regulating and reshaping AI development. Draft 1 said that whether this is intellectual honesty or rhetorical borrowing “depends on whether the book develops it seriously,” and left it open.

Checked. It develops it seriously — Hayek is worked through in Chapters 2, 3 and 9, and Chapter 3 spends several pages saying precisely where Acemoglu departs from him and why. So the borrowing charge doesn’t hold. The essay that is left is a better one: Acemoglu accepts Hayek’s epistemics in full and rejects his politics on the grounds that Hayek applied the knowledge-problem to states and not to concentrations of private economic power. Whether that asymmetry is Hayek’s blind spot or Acemoglu’s convenience is the real argument, and it is now an argument between two stated positions rather than a suspicion about one of them.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech — including the Hayek quotations, which are quotations inside a transcription and should be verified against Hayek’s own text before republication.

Related: AI Growth Forecasts - Whose Timeline · Pro-Worker AI · Social Contract Theory and Its Critics

Who. Henry Sidgwick, Cambridge. His The Methods of Ethics (1874) is regarded by many philosophers as the finest work of moral philosophy in English. Rawls called it the first genuinely academic treatise on ethics. John Maynard Keynes’s circle grew out of his. He also co-founded Newnham College to get women into Cambridge, and spent years investigating psychic phenomena, which tells you something about his commitment to following evidence wherever it led.

What he argued. Sidgwick examined three “methods” of ethics — egoism, intuitionism, and utilitarianism — and tried to show they could be reconciled. He concluded that utilitarianism was correct but could not be fully squared with rational egoism. He called this the “dualism of practical reason” and, remarkably, admitted he could not solve it. That honesty is why philosophers respect the book.

Why Cowen names him. In the transcript Cowen’s move is:

“If we’re Sidgwick, we might say, when values clash, we look to utility, or maybe it’d be cost-benefit analysis to decide which gets priority. You’re not willing to say that, so why not?”

This is the decisive question of the segment. Acemoglu has said nondomination and noninterference can conflict and that he has no metric to resolve it. Sidgwick’s utilitarianism offers one: a single currency, welfare, into which all values convert. Cost-benefit analysis is that idea in working clothes, and it is how most policy actually gets made.

Acemoglu’s answer is that the conversion itself needs a rulebook — “Who decides how to compare, how to turn things into utils?” — so utilitarianism doesn’t escape the problem, it relocates it. See The Book Problem.

Worth knowing: Sidgwick is also the origin of the “government house utilitarianism” charge — that a utilitarian elite might rationally conclude the public should be taught a simpler morality than the true one. That connects directly to The Straussian View, which Cowen raises a few minutes earlier. The same worry appears twice in this conversation under two names.

Reading

• Sidgwick, The Methods of Ethics (1874) — Book IV is the utilitarian core
• SEP: Henry Sidgwick
• Parfit, On What Matters — a modern continuation of Sidgwick’s project

Working notes

Short by design — Adam asked “who is.” The genuinely interesting thread is the Sidgwick→Strauss link on esoteric morality, which I’ve flagged above and would develop if it interests you.

Checked against the book: Sidgwick is not in it. Neither is the dualism of practical reason, nor any sustained treatment of utilitarianism — utilitarian does not occur anywhere in What Happened to Liberal Democracy?, and Bentham and Mill appear only as classical liberals criticized for defining liberty as non-interference. So this page and the utils exchange in The Book Problem rest entirely on the conversation. Cowen supplied Sidgwick; Acemoglu answered live. Worth knowing before the Sidgwick→Strauss thread gets developed — it would be building on the interview, not on the book.

Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed; a proper noun a machine transcriber could mangle is the weakest kind of negative finding, but no passage discusses his argument under any spelling.

Related: John Rawls · The Book Problem · Rorty, Objectivism, and Moral Progress

Habakkuk’s argument

Sir John Habakkuk, American and British Technology in the Nineteenth Century (1962) — Acemoglu links this exact edition in the transcript.

The puzzle: why did the United States, poorer and less industrially mature than Britain, adopt more mechanized production methods?

Habakkuk’s answer: abundant land made labor scarce. A worker could always go west and farm, which put a floor under wages. Expensive labor made machinery relatively cheap, so American manufacturers mechanized aggressively — interchangeable parts, the “American system of manufactures” — while British firms, with cheap abundant labor, had less reason to.

The general principle: the direction of technical change responds to relative factor prices. Technology isn’t an exogenous rain of ideas; it points wherever the expensive input is.

The lineage

Acemoglu’s own line — “The models that I’ve been working on for the last 25 years on this topic always said that’s a possibility” — refers to his directed technical change work of the late 1990s and early 2000s, which supplied the market-size and price effects that make induced innovation work formally rather than as an intuition.

Why it matters for the fertility argument

If technology is exogenous, fewer young workers simply means less output and a shrinking economy — the demographic panic.

If technology is directed, scarcity of young workers raises their relative price, which pushes innovation toward labor-saving methods, which raises output per worker. The demographic shock partly answers itself.

That is precisely what Baby Busts and Growth Booms claims to find empirically: regions and countries with fewer young workers show more labor-saving patents, more high-tech activity, and higher TFP growth.

The caution Acemoglu himself puts on it

Twice, and both are important:

“it’s not a certainty, but it’s only after seeing the evidence that I’m more in that camp.”

and, on choice:

“In both papers, you also see the element of choice. You have to do the technology, and not every society does that technology in the same way. We could get that wrong.”

This is the load-bearing connection to the rest of his work. Induced innovation is a tendency, not a law. Automation that responds to genuine scarcity is good; the same automation deployed where labor is abundant is the displacement story of Automation and the Labor Share. Same mechanism, opposite welfare consequence, depending on conditions.

It is also why the Habakkuk framing is a better fit for him than the RBC framing Cowen offered. RBC says the economy self-corrects. Habakkuk-via-Acemoglu says the economy responds to incentives, and whether the response is good depends on what the incentives are — which leaves plenty of room for policy and for getting it wrong.

Reading

• Habakkuk, American and British Technology in the Nineteenth Century (1962)
• Hicks, The Theory of Wages (1932), ch. 6 — the original statement
• Acemoglu, “Directed Technical Change,” ReStud 2002 — the modern foundation
• Acemoglu & Restrepo, “Demographics and Automation”

Working notes

Habakkuk’s thesis is contested and I should say so plainly: economic historians have pushed back for sixty years, notably on whether American labor really was that much scarcer once you account for skill mix, and on whether British firms were as unmechanized as the story requires. Peter Temin and others have argued the wage-gap evidence is weaker than Habakkuk claimed. Acemoglu calls it “intriguing,” which is doing some work. Draft 2 should give the counter-literature a proper paragraph rather than the sentence it has here.

Checked against the book: Habakkuk is not in it, and neither is the nineteenth-century labor-scarcity literature. The book’s version of “technology responds to conditions” is contemporary and institutional rather than historical — Chapter 6 on German versus American robot adoption, where the difference is union bargaining and apprenticeship rather than relative wages, and Chapter 9 on why business models and AGI ideology push AI toward automation. That is induced innovation with the inducement coming from institutions and beliefs instead of factor prices, which is a meaningfully different claim and arguably a more defensible one. Worth noting in draft 2 alongside the counter-literature: Acemoglu calls Habakkuk “intriguing” on air and builds on something else in print.

Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed; references are by chapter.

Related: King and Plosser - Real Business Cycles · The Lucas Critique · Acemoglu and Restrepo - The Task Framework · Automation and the Labor Share

The famous machinery. Rawls’s original position, behind a veil of ignorance, asks what principles people would choose without knowing their place in society. His two principles give basic liberties priority over the second principle’s requirements for opportunity and economic distribution:

1. The liberty principle — the most extensive basic liberties compatible with the same for everyone.
2. Fair equality of opportunity — positions genuinely open to all.
3. The difference principle — inequalities are permitted only if they benefit the least advantaged.

The priority rules exclude trading basic liberties for greater income and put fair opportunity ahead of the difference principle. This is background to Rawls, not the target of Acemoglu’s objection. See the account of the two principles.

Why Acemoglu objects. In the transcript:

“Ultimately, even though he’s a true liberal, John Rawls also goes into that direction.”

Chapter 3 explains the objection: a philosopher’s account of what people would agree to cannot stand in for their actual participation and consent. The earlier draft guessed that Acemoglu meant Rawls’s ordering of principles. The book review corrected that guess; the hypothetical agreement is what is at issue.

Rawls vs. Nozick, if you want the other half: Robert Nozick’s Anarchy, State, and Utopia (1974) accepts the contractarian starting point and derives a minimal state, arguing that redistribution violates rights that people hold prior to any distribution. The two books are usually taught together.

From the book

Chapter 3, Liberalism Without Social Contract, is where Rawls is argued with rather than named in passing, and it settles what “goes into that direction” means. It is not lexical priority. The objection is to the hypothetical agreement itself:

“John Rawls, whose 1971 book, A Theory of Justice, is foundational to twentieth-century liberalism, was much more committed to liberal ideas than Rousseau was, but I will argue that the same sort of presumed-to-have-agreed reasoning creates tensions for Rawls’s philosophy too.”

The tension is the same one he presses on Rousseau, transposed:

“In practice, nobody is ever asked their views about justice when a social contract is being designed or behind a veil of ignorance… For those living in a real society, not behind a curtain, it would be little comfort to hear that their condition has been justified by a philosopher or statesperson claiming they would have assented to such a social contract behind the veil. Given the chance, they would readily tear up that social contract and the veil.”

And then the general form of the complaint, which is about who does the deciding, not about what got ranked first:

“The root problem is that social contract theories, interpreted as the basis for general ethical rules, go against self-government. They don’t seek the actual consent of individuals and communities to public choices by a process of deliberation, communication, and debate. Rather, they attempt to circumvent any opposition by invoking high-level precepts without broad buy-in from the population.”

Three things this changes.

1. The ordering is not his target. Chapter 3 discusses the difference principle and describes basic liberties as a guardrail on it. Presenting them in that order does not reverse their priority: a guardrail on redistribution is compatible with liberty taking precedence. The objection developed in the chapter is instead to presumed agreement replacing self-government.

2. He does mention the later Rawls. In the same passage: “In later work, Rawls also grapples with the issue of shared moral principles, or public reason, which I discuss below.” So Political Liberalism is not the unnoticed refutation it looks like from the interview — it is acknowledged, and folded into his own account of deliberation later in the chapter.

3. He grants Rawls a legitimate role, which the interview never gets to. The chapter names an “alternative interpretation that Rawls himself offers for his contractarian philosophy, as an input into a debate on shared moral principles” — and accepts it, with a limit:

“However, in this case, such reasoning should be used mainly in the context of deliberation and cannot justify actual policies or impositions of values on people, for example, becoming the basis of laws and how policymakers judge different options.”

By the end of the chapter this has become his positive proposal: political philosophy, “especially those based on liberal principles like John Rawls’s,” as input into deliberation, never as its conclusion. The quarrel is not with what Rawls concluded. It is with anyone treating a philosopher’s conclusion as consent already given.

Reading

• Rawls, A Theory of Justice (1971) — §§1–4, 11–17 give you the argument without the whole book
• Rawls, Political Liberalism (1993) — the overlapping-consensus move
• SEP: John Rawls
• Nozick, Anarchy, State, and Utopia (1974) — the standard counterweight

Working notes

Resolved against the book, and my reading was wrong. Draft 1 guessed that “goes into that direction” meant Rawls’s lexical priority rule, and called checking it the single highest-value verification on the whole rabbit-hole list. It was — and it came back negative. The words lexical and lexicographic do not occur anywhere in the book, the priority ordering is never discussed as an ordering, and the objection Chapter 3 actually makes is the looser one I listed as the alternative: that contractarian reasoning derives principles philosophically instead of democratically, so nobody is ever actually asked. The From the book section above replaces the guess.

That resolution propagates. Social Contract Theory and Its Critics and The Book Problem were both waiting on this answer, and both were leaning on the same wrong reading.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter rather than page. Quotations are transcribed speech and may differ from the printed text in punctuation; they should be re-checked against print before being quoted as exact. A negative result — “the book never says lexical” — is weaker evidence from a transcript than from a printed index, though a rule that is never discussed in twelve hours of text is not a rule he is arguing against.

Related: Social Contract Theory and Its Critics · The Book Problem · Henry Sidgwick

Who they are

Cowen says “Long and Plosser, King and Plosser” — he is naming the pair of collaborations, not confusing himself.

What real business cycle theory claims

The provocative core: recessions are not failures. They are the efficient response of a well-functioning economy to real shocks — mostly shocks to productivity.

• The driver is technology, not money or demand. Money is roughly neutral.
• Employment falls in a downturn because workers choose less work when the real wage is temporarily low (intertemporal substitution of labor).
• The outcome is Pareto-efficient. Stabilization policy has nothing to fix, and attempts to fix it make things worse.
• The residual — the unexplained part of growth, “total factor productivity” — is treated as the fundamental driving force.

It was, and is, deeply contested. The standard objection is that it requires believing the Great Depression was a large voluntary vacation. But its method won even where its conclusions didn’t: RBC established the calibrated microfounded model as the standard tool, which is why every modern DSGE model is a descendant.

What Cowen is actually saying

Follow the logic of his question:

“Does that mean you’re not really such a Keynesian? Because the fertility crisis, aggregate demand, aggregate supply, they’re shrinking forever. A Keynesian or a supply-sider could get quite worried. Paul Romer, the market size is shrinking, he should get quite worried. In a funny way, you end up at a real business cycle point of view where as long as the induced innovation is positive, the macro will go fine. You’re like Long and Plosser, King and Plosser in your macro.”

The argument step by step:

1. Falling birth rates shrink both demand and supply, permanently.
2. A Keynesian should fear the demand side — fewer people, less spending, chronic shortfall.
3. A Romer-style growth theorist should fear the supply side — ideas depend on the number of people looking for them, and a smaller market means weaker incentives to innovate. (Romer’s own model makes population a driver of growth.)
4. Acemoglu fears neither, because his paper says technology responds: scarce labor induces labor-saving innovation, and growth per working-age adult rises.
5. That is an RBC-shaped answer. The real side — technology, the residual — determines outcomes; demand doesn’t matter much; the economy self-corrects through the productivity channel without policy.

Cowen’s next line makes the target explicit: “It’s about the residual, and if the residual is—” The residual is TFP, the thing RBC theorists put at the center.

So the jab is: the market-failure economist has ended up at a self-correcting-economy conclusion.

Acemoglu’s answer

Three moves, and they’re worth separating:

1. He doesn’t recognize the framing. “I never thought of King and Plosser about induced innovation.” Honest — he came to this through the growth-theory literature, not the macro one.
2. He supplies his own ancestor instead: “Habakkuk was the first one who was about this.” See Induced Innovation and the Habakkuk Thesis. This is him saying: my lineage is 19th-century economic history, not 1980s Minnesota macro.
3. He half-accepts the charge: “I never saw myself as a traditional Keynesian… and I never understood the new Keynesian models very well.” But he keeps a demand channel — “there are episodes in which aggregate demand shortages create problems” — and argues the fertility decline is too slow-acting to produce one.

The honest bit: “I’m also finding these results surprising. We’ve checked them so many times, and that’s why we’ve done them so many different ways, but it seems to be there in the data.” He is reporting a result he didn’t want.

Reading

• Long & Plosser, “Real Business Cycles,” JPE 1983 — the founding paper
• King & Plosser, “Money, Credit and Prices in a Real Business Cycle,” AER 1984
• Summers, “Some Skeptical Observations on Real Business Cycle Theory” (1986) — the classic demolition, four pages, entertaining
• Romer, “Endogenous Technological Change” (1990) — the position Cowen says should worry

Working notes

Cowen’s “you’re like Long and Plosser” is a compliment and a needle at the same time, which is characteristic. The substantive question underneath — is induced innovation a self-correcting mechanism strong enough to make demographic decline harmless? — is genuinely open, and The Lucas Critique is the strongest reason to doubt it.

Checked against the book: not there. No real-business-cycle literature, no Long and Plosser, and no treatment of demographic decline as a growth question. The Baby Busts result Cowen presses him on is a working paper, not book material. What the book does supply, in Chapter 6, is the empirical half of the induced-innovation argument underneath this exchange — German firms adopting robots faster than American ones and reallocating rather than shedding workers, which is automation responding to institutions rather than to scarcity. See Acemoglu and Restrepo - The Task Framework. The open question below — whether induced innovation is self-correcting enough to make demographic decline harmless — gets no help from the book either way.

Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed; references are by chapter.

Related: Induced Innovation and the Habakkuk Thesis · The Lucas Critique · Acemoglu and Restrepo - The Task Framework

New York City

New York State as a whole spends about $30,723 per student — also first in the nation. The state’s spending gap over the national average has more than quadrupled since 2000.

The enrolment fact is the one that makes the number strange: NYC has roughly 100,000 fewer students than in 2019–20, and spending per student rose anyway. Fixed costs — pensions, debt service, buildings, transportation — do not fall when a child leaves.

Sources: Citizens Budget Commission · Empire Center · Reason Foundation

London / England

London attracts an area cost adjustment above the national formula — inner London is historically the highest-funded region in England. I could not source a clean 2025–26 inner-London per-pupil average; a rough estimate of £8,000–9,000 is plausible but unverified.

At roughly $1.27/£: £8,500 ≈ $10,800.

Sources: DfE school funding statistics 2025–26 · ASCL information paper · IFS annual report on education spending

The comparison, honestly

Headline: NYC ≈ $42,000 vs London ≈ $11,000 — about 3.8×.

Why that overstates the real gap:

1. Different things are being counted. The US per-pupil figure is total current expenditure: teacher pensions and retiree health, debt service, student transportation, food service, special education, central administration. England’s NFF number is the mainstream schools block — it excludes high-needs funding, most capital, and portions of pension costs funded through separate grants.
2. Pensions and retiree health are enormous in NYC and are a legacy-cost transfer, not instruction.
3. Health care is in the number. US teacher compensation carries employer health insurance; the UK’s does not, because the NHS does.
4. Cost of living and salary structure differ; so does the definition of a “pupil.”

A fairer frame — national totals. OECD comparisons of total primary-and-secondary spending per student, PPP-adjusted, put the US around $15,000–16,000 and the UK around $12,000–14,000 — a gap of maybe 15–25%, not 400%. NYC is an outlier within the US, not a representative American figure. ⚠ These OECD figures are from memory and should be re-sourced before use.

What survives the caveats: NYC really does spend far more per student than almost anywhere in the world, the gap has widened sharply since 2000, and spending has continued rising as enrolment falls. Cowen’s point stands. The 3.8× number does not survive scrutiny; a 2–2.5× real-resources gap probably does.

Why this matters in the conversation

This is the exchange where Acemoglu is at his most exposed and he knows it. His claim is that America has “given up on the education of about a third to a half of the youth.” Cowen’s counter is that NYC is the richest-funded school system on earth and the results are middling — so money is plainly not what was given up on.

Acemoglu concedes immediately — “Oh, yes, I know… We’re throwing money at it, but we’re not really dealing with the root causes” — and pivots to a claim about dysfunction and ideology, then agrees teachers’ unions “need to be reconstituted.” That is a substantial retreat from “we have given up,” and it happens in three lines.

The interesting thing for you: the CBC’s framing is literally “Highest Costs, Middling Marks.” Both men accept the premise that the money isn’t reaching the outcome. Neither offers a mechanism.

From the book

The interview’s NYC-versus-London comparison is Cowen’s, and it is not in What Happened to Liberal Democracy?. What the book has is the same argument at national scale, in Chapter 1, as part of the case that liberal democracy stopped delivering:

“Building schools is expensive, but it’s much easier than increasing the quality of education once the buildings are there… Many indicators point to worsening educational performance over the last decades, even as inflation-adjusted spending per pupil rose from around $6,500 in the 1970s to almost $17,000 in 2019, both in 2022–23 dollars. By the late 2010s, only a third of fourth-grade students could read at the proficient level expected for their grade.”

A 2.6× real increase in per-pupil spending, against a third of fourth-graders reading at grade level. That is the shape of the claim this page is testing, sourced to a figure Acemoglu himself publishes, and it does the job the “3.8×” ratio was going to do — without the London estimate the working notes below say not to publish.

Two things worth noting about how he uses it:

• It is deliberately not a story about schools. The passage sits inside a longer list — health spending three times France’s or Japan’s against lower life expectancy, one in four Americans without a primary care physician — and the conclusion drawn is about public services in general failing to convert money into outcomes. Education is one instance of “the crisis of drains and dole” he names in Chapter 10, not the subject.
• He does not blame teachers’ unions, in the book or, notably, anywhere the transcript reaches. The causes offered are that the low-hanging fruit was picked, that American schools serve a harder population, and that the pandemic made it worse. Cowen’s framing of the exchange invites the union answer; Acemoglu doesn’t take it.

The book supplies no international comparison of school spending, so nothing above needs re-sourcing to it — and the NAEP-style outcome data the working notes ask for is exactly what his “only a third of fourth-grade students” figure is.

Working notes

Confidence. The NYC figures are well sourced and mutually corroborating across three independent organizations with different political priors — CBC, Empire Center, Reason. Treat $42,168 (2025–26 projected) and $35,914 (2021–22 actual) as solid.

⚠ The weakest numbers on this page are the London ones. The England-wide NFF average is solid; the inner-London figure is my estimate and is marked as such. The dollar conversion compounds that with an exchange-rate assumption. Do not publish the “3.8×” figure. If this page graduates to something reader-facing, someone needs to pull the DfE’s local-authority-level schools block allocations for the London boroughs and do the arithmetic properly. That is a one-hour job with the right spreadsheet and I’d rather flag it than fake it.

The inflation-indexed history you asked for is only half here. I have “+21% real over ten years to 2024” for NYC and the “quadrupled gap since 2000” for the state. What’s missing is a proper CPI-deflated series 2000→2026 for NYC specifically. The IBO publishes the underlying data; it needs assembling, not searching.

Draft 2 should add:

• A real deflated time series, NYC per-pupil in constant 2026 dollars, 2000–2026
• The same for enrolment on a second axis — the divergence is the whole story
• Outcome data alongside it (NAEP scores) so “middling marks” is shown rather than asserted — and the book gives the national version of exactly this pairing, so the page can lead with Acemoglu’s own numbers and use NYC as the local case rather than the other way round

Checked against the book. No NYC–London comparison exists in it; the national spending-versus-outcomes claim does, in Chapter 1, and is written up above. That does not rescue the “3.8×” figure — still do not publish it — but it does mean the page no longer depends on it to make its point.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter. The $6,500 / $17,000 figures are machine-transcribed numbers and the transcript garbles the base-year phrase; confirm both against print before publication.

Related: Education Quality and Economic Performance - The Outliers

The definition

Two categories, from the same underlying framework as Acemoglu and Restrepo - The Task Framework:

His own examples: “nurses now can do much more, journalists can do much more, electricians can do much more.” Note that all three are non-elite skilled occupations. That is not incidental — it’s the working-class-liberalism argument in technological form. See Working-Class Liberalism and Community Elbow Room.

And the exclusion is explicit: “Automation, again, narrowly construed, isn’t pro-worker AI.”

The part everyone gets wrong

Cowen sets a trap — a good one. If you want pro-worker AI, does a regulator inspect what Stripe and Mercatus are doing with their models and approve it? Acemoglu’s answer is unusually emphatic:

“I do not think that we know today or we desire today a regulator deciding what is pro-worker AI and what’s not, and then saying only pro-worker AI should be enabled. I’ve never argued that, and I do not believe that.”

And again: “No, I do not believe that we need to have a regulator that says, oh, only pro-worker AI is allowed. Absolutely not.”

What he wants instead is narrative and aspiration change among technologists, with public policy in a supporting role:

“Cultivating it has to start by changing the narrative, changing the aspirations of tech entrepreneurs and tech companies, and public policy has a role but has a supporting role in that.”

This is a much weaker lever than his critics usually assume, and arguably weaker than his own diagnosis warrants. If the automation/new-task mix is determined by firms’ choices under competitive pressure, exhortation is a strange instrument to reach for.

The obvious problem, unaddressed

The distinction may not survive contact with an actual product. A tool that lets one nurse do the work of two is simultaneously (a) expanding that nurse’s capability and (b) reducing headcount. Which is it? Acemoglu’s framework says the answer depends on whether total tasks expand — but that is only knowable after the fact and in aggregate, not at the moment of design.

Neither man raises this, and it is the question a working engineer would ask first.

From the book

Pro-worker AI is not a passing phrase in the conversation — it is the centre of Chapter 9, Liberalism in the Age of Algorithms, and returns as the fourth challenge of Chapter 10. The definition is the same one he gives Cowen: “Pro-worker AI explicitly aims to expand worker productivity and capabilities.”

The worked examples the interview only gestures at. He builds three, and is careful that two of them do not yet exist:

• The electrician tool. An AI trained on electrical-engineering knowledge, equipment specifications, and “past use cases where expert electricians dealt with different problems”, which diagnoses faults from sensors and photographs and then guides an iterative human search for the cause. He notes “this exact version… does not exist,” but that a global company is building related products for engineers and field technicians. The framing matters: “pro-worker AI can be pro-manual worker too.”
• The classroom tool. Trained on curricular material and past classroom experience, fed by short quizzes, recommending in real time that the class split into smaller groups and proposing strategies the teacher can modify. “This tool doesn’t exist at the moment either, but a prototype is being developed.” The test he applies is the sharp one: “While automated tests and other current instructional offerings attempt to reduce the need for teachers, a successful rollout of an AI tool of this sort would increase the demand for teachers.”
• Healthcare, more briefly — nurses “more actively engaged in diagnosis and care decisions,” checking drug interactions, with real-time guidance on when to involve a physician.

There is also an argument the interview never reaches: pro-worker AI as the answer to Baumol’s cost disease in exactly the sectors where public services are most expensive, which is what ties Chapter 9 to the public-services challenge in Chapter 10.

Why we don’t see more of it. Two reasons, given in order: an economic one and an ideological one. The economic reason is that “the business models for generating revenues from digital ads… and for automating work by selling software services to companies are well-established,” while “there is less of a track record in digital technology of monetizing new tasks” — and that whoever builds the first version usually isn’t the one who monetizes it. Then the tax code, which “subsidize[s] capital expenditures relative to labor outlays.” The ideological reason is stronger, and is the most quotable thing in the chapter:

“This approach produces excessive focus on AGI and automation, and may even put blinders on technologists, making them miss out on feasible and profitable pro-worker AI opportunities. Belief in AGI can also be self-fulfilling.”

Self-fulfilling twice over: within firms, because “more and more sophisticated-looking models are marketed and adopted for automation, even when they are not ready for prime time, with alternative technological directions remaining unexplored”; and between states, because if AGI is imminent and decisive then “US-China relations inevitably become more conflictual.”

And here the book and the interview do not say the same thing. Cowen’s regulator trap draws an emphatic denial — “I’ve never argued that, and I do not believe that” — and a retreat to narrative change with “public policy… in a supporting role.” Chapter 9 ends like this:

“Creating new aspirations for technology and enacting appropriate policies and regulations could ensure that the market process still functions, but prioritizes technological developments that differ from our current path. This calls for new regulations as well. … the aim should not be to create a completely deregulated economy… The point is to focus on the more important regulations, and in the age of algorithms, the regulation of social media and AI should be at the top of the list.”

Chapter 10 goes further and names the instrument: after arguing for cutting ex-ante rules generally in favour of ex-post accountability on the Federal Reserve model, it makes an explicit carve-out — “though for AI, social media, and certain issues relating to health and safety, ex-ante regulations would also be necessary.” Plus antitrust: “reducing the dominance of technology monopolies” on the grounds that “monopoly in the tech sector discourages new directions of innovation.” Plus the tax-code fix, plus union support.

Strictly, it isn’t a contradiction. What he denies to Cowen is narrow and specific — a regulator that inspects models, rules on which are pro-worker, and permits only those. Nothing in the book proposes that. But “narrative change, with public policy in a supporting role” is a poor summary of a programme containing ex-ante AI regulation, antitrust action, a rewritten capital-versus-labour tax treatment, and a reconstituted labour movement bargaining over technology choice. The interview makes his lever sound weaker than his book does. The criticism on this page — that exhortation is a strange instrument for a problem he diagnoses as structural — turns out to be a criticism of the interview rather than of the position.

Reading

• Acemoglu & Johnson, Power and Progress (2023) — the book-length version of the “technology is a choice” argument
• Acemoglu & Restrepo, “Automation and New Tasks” — the mechanism
• Autor, “The Work of the Future” — the empirical case that new tasks are most of the story

Working notes

This page exists because AI Growth Forecasts - Whose Timeline and Working-Class Liberalism and Community Elbow Room both needed to point at it — you highlighted the passage but didn’t ask a question about it. If it isn’t useful, say so and I’ll fold it into the task-framework page.

The strongest unexamined thread: Acemoglu says he has been talking about pro-worker AI “like a broken record… for 10 years.” If the narrative-change strategy is the plan and it has had a decade, it would be worth asking what evidence there is that it works. Nobody asks. The book check weakens this question without dissolving it — narrative change is not the whole plan in print, so a decade of it failing is less damning than it looks. But narrative change is still where Chapter 9 starts, and the book does not say what would count as it working either.

The “obvious problem, unaddressed” is still unaddressed. The book gets closer than the interview — the classroom example is explicitly designed so that success increases demand for teachers, which is a real attempt at a criterion — but that is a claim about one hypothetical tool’s aggregate effect, not a way to tell at design time which side of the line a product falls on. The nurse case, where the same tool plausibly does both, is not taken.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech. Re-check wording against print before quoting as exact — particularly the regulation passages, which are the load-bearing ones here.

Related: Acemoglu and Restrepo - The Task Framework · Hayek and Decentralization · AI Growth Forecasts - Whose Timeline

Three positions, cleanly separated

Note on the word. “Objectivism” here means moral realism. It is not Ayn Rand’s Objectivism, which is a specific ethical egoism plus a metaphysics and a politics. Cowen means the first. Acemoglu answers the first.

What Rorty actually argued

The article Cowen links is “The Priority of Democracy to Philosophy” (1988). Its claim is deliberately deflationary:

• Liberal democracy does not need a philosophical foundation. It does not need a theory of human nature, of rights, or of moral truth.
• Attempts to provide one are not just unnecessary but harmful — they make the polity hostage to a philosophical argument that can be lost.
• What we have instead is a historically contingent set of practices and a widening circle of who counts as “us.” Rorty calls this solidarity, not objectivity.
• Moral progress is real but it is not progress toward anything. It is getting better at including people and reducing cruelty.

The famous formulation from Contingency, Irony, and Solidarity (1989): a liberal is someone who thinks cruelty is the worst thing we do — and the ironist knows that this commitment has no grounding deeper than her own contingent formation.

Where Acemoglu sits, and why Cowen pokes

His actual words: “in the same way that many people would agree science makes gradual and sometimes uncertain progress toward some better understanding, I think morally, philosophically, we also have to have gradual uncertain evolution toward what we agree on.”

Read that carefully. He offers a science analogy — and the science analogy is exactly what a realist uses. Science converges because there is a world constraining it. If moral inquiry converges the same way, something is doing the constraining, and that something is an external standard.

Then Cowen asks directly what the standard is and he says: “I don’t have one.”

So there are two readings:

• Rorty-compatible reading. “Progress” just means “increasing agreement,” full stop. No target. Consistent, but then “progress” is doing no work — you cannot say a society that agreed more about something terrible had progressed.
• Quietly realist reading. He does think some moral views are better, and the evolution is toward them, and he simply declines to name the standard. Honest about his uncertainty, but not a stable position — it is realism with the metaphysics unpaid for.

The tell is his next answer: he does name something after all — meaningful individual freedom, as the “only starting point that we need to have.” That is a foundation, offered one exchange after denying he has one. It is thin, and he presents it instrumentally (freedom is required for human improvement), but it is doing foundational work.

“Objectivism, in some sense, is too strict” is his own gloss: he is not denying moral facts exist. He is denying we can ever have a clear map of them at any given time. That is an epistemic claim, not a metaphysical one — and epistemic humility about moral facts is fully compatible with believing they exist. Which is why the Rorty comparison doesn’t quite land, and why Cowen keeps pushing.

Is objectivism live? — the honest answer

Yes, and more so now than when Rorty was writing. Worth knowing:

• Derek Parfit, On What Matters (2011) — the most ambitious modern defense, arguing Kantian, contractualist and consequentialist ethics “climb the same mountain from different sides.” If he’s right, convergence is evidence for realism.
• Companions in guilt arguments — if you accept mathematical or logical facts as mind-independent, the standard arguments against moral facts apply to those too.
• The evolutionary debunking challenge (Sharon Street) — the strongest live attack: our moral intuitions were selected for fitness, not truth, so why think they track moral facts?
• Quasi-realism (Simon Blackburn) — you can earn the right to talk like a realist without the metaphysics. Arguably where Acemoglu actually is.

From the book

Rorty is not in it. Neither is Sidgwick, neither is Parfit, and there is no chapter, section or passage on moral realism, moral progress, or the standing of ethical claims. The comparison is Cowen’s, supplied live, and Acemoglu accepted the frame without having worked in it. That absence is the finding, and it is more interesting than a quotation would have been: the exchange Cowen forces is one Acemoglu has not had in print.

What the book has instead, in the place where a foundation would go, is an epistemic premise:

“Premise 1. Fallibility. Every belief, proposition, or piece of knowledge we hold could be wrong, even spectacularly wrong. This humility should be the starting point for any discussion of knowledge, given the tendency of human communities to hold beliefs and follow practices that later turn out to be only partially true, incorrect, or even downright misguided.”

The lineage he claims for it is Mill on the fallibility of ages, and Peirce: “we cannot in any way reach perfect certitude or exactitude. We can never be absolutely sure of anything.” Peirce is the founder of the tradition Rorty belongs to — so Cowen’s instinct that there is a pragmatist here was right, and he named the wrong pragmatist. Peirce is also, unlike Rorty, a convergence theorist: the whole point of Peircean fallibilism is that inquiry gets somewhere in the long run, which is exactly the science analogy Acemoglu reaches for and exactly the thing Rorty denied.

And the book has a third premise that is the real answer to Cowen’s question, though it is about beliefs rather than about values:

“Premise 3. Gentle Epistemic Improvement. Incorrect beliefs can spread and may sometimes become dominant. Nevertheless, there is a sometimes weak tendency for incorrect beliefs to fade over time, often under the influence of what David Hume called a gentle force. This is especially the case when there is room for debate and open exchange of information, and when the cage of norms is not too suffocating… Incorrect beliefs either have logical flaws or contradict observed empirical patterns, and these contradictions open the door for sceptics to challenge the system.”

That is a substantive claim about why the drift is toward better rather than merely elsewhere: wrong beliefs are eventually embarrassed by logic or by the world. It is stated as a premise, not defended — he grants immediately that it is “a slow, laborious, and hazardous process” — and it is doing exactly the work Cowen is asking him to account for. It also does not obviously extend to values, which is where the argument on this page actually lives, and the book never asks whether it does.

The instrumental move, confirmed. Draft 1 read Acemoglu as offering individual freedom as a thin, instrumentally-justified foundation rather than as a moral fact. The book is unambiguous on this, and states it most clearly about property rights:

“This supports property rights not because of a notion of natural rights or some principle that one’s holdings of goods and assets are sacred. Rather, they are instrumental, a vehicle for effective experimentation and knowledge building.”

Whether that dissolves Cowen’s question or relocates it is the live issue. Nothing in the book explains why “effective experimentation and knowledge building” is the thing worth having, and Premise 4 simply asserts it: “Knowledge is not an end in itself. It is an instrument for achieving society’s diverse objectives.” Whose objectives, chosen how, is where the regress lands — one step further back than Cowen pushed him, and unaddressed on either side.

Reading

Working notes

The strongest thing on this page is the observation that Acemoglu uses a science-convergence analogy and then denies having a standard, and that these do not sit together. Draft 1 flagged this as unestablished and wanted it checked against the book, on the grounds that he may well address it in print.

Checked. He does not address it in print — there is no treatment of moral realism, moral progress or the grounding of values anywhere in the book. So it is not a compression artifact of a spoken interview; the gap is real. What the book supplies instead is a Peircean fallibilism plus a stated premise that wrong beliefs tend to fade, both of which are about knowledge rather than about values and neither of which is defended. That makes the observation stronger than draft 1 assumed, and it also makes it less of a gotcha: he is not dodging a question he has answered elsewhere, he is working without a net in a place where his book never needed one.

Rorty on “why believe what you believe.” Cowen’s follow-up (“Then why believe in what you believe?”) is the classic anti-pragmatist challenge, and Rorty had a well-known answer Acemoglu doesn’t reach for: because I was raised in this tradition and I am not ashamed of it. Rorty thought demanding a deeper reason was the mistake. Acemoglu instead reaches for individual freedom as a starting point, which is a more conventional liberal move.

Answered: the book contains no engagement with Rorty — no mention at all — so the second branch is the one that obtains. Cowen supplied the comparison and Acemoglu accepted the frame without having worked in it. Written up above.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech. This page leans harder than the others on an absence — no Rorty, no Sidgwick, no section on moral realism — which a machine transcript establishes less firmly than a printed index would, particularly for a proper noun a transcriber could mangle. What is not in doubt is the positive finding: where a grounding argument would sit, the book puts fallibilism and an instrumental justification of rights.

Related: Social Contract Theory and Its Critics · The Book Problem · Henry Sidgwick

The idea and its lineage

Every social contract theory makes the same three moves:

1. Imagine people outside political authority. Hobbes’s state of nature, Locke’s, Rawls’s “original position.”
2. Ask what they would agree to. The answer is engineered by how you describe step 1.
3. Treat that agreement as the source of legitimacy. The state may do what people would have consented to, and no more.

The step-1 description does all the work. Hobbes describes the state of nature as terrifying and gets an absolute sovereign. Locke describes it as inconvenient but tolerable and gets a limited government. Rawls puts you behind a veil of ignorance and gets redistribution. Nobody in any of these stories ever signed anything.

Why Rousseau is the sharp end

Acemoglu names “the Rousseau end of it,” and that is the precise place to press.

Rousseau distinguishes the general will (volonté générale) — what the people collectively will when aiming at the common good — from the will of all (volonté de tous), the mere sum of private preferences. The general will is not a vote count. It is the right answer, and a majority can be wrong about it.

That produces the notorious line in Book I, Chapter 7: whoever refuses to obey the general will “shall be compelled to do so by the whole body,” which “means nothing less than that he will be forced to be free.”

Acemoglu’s structural complaint is exactly this: the general will is defined to outrank any actual agreement a real community reaches. So when a community’s shared values conflict with what the theory says the general will requires, the community loses by definition, not by argument. The hard question — how do we handle people who disagree with us, including intolerantly? — never gets asked.

What he is proposing instead

From the transcript, his positive view has two layers:

• A floor: meaningful individual freedom, which he treats as the one non-negotiable, because “any kind of improvement in the human condition requires individual initiative.”
• Above the floor: genuine consensus-building, with “enough elbow room to people to form their own community-level agreements.”

He is explicit that this is not grounded in an external standard. Cowen presses him on it three times and he answers “I don’t have one.” That is a real philosophical cost and he pays it deliberately — see Rorty, Objectivism, and Moral Progress and The Book Problem.

The strongest objections to his objection

Worth holding, because he does not answer them in this conversation:

1. Rawls already did this. Political Liberalism (1993) introduces “overlapping consensus” precisely to handle deep disagreement — the idea that people with incompatible worldviews can converge on shared political principles for their own different reasons. That is a consensus mechanism, not a book of absolutes. Acemoglu says Rawls “goes into that direction” anyway; the transcript doesn’t show why. The book does — see below, and the answer is that he does not deny the later Rawls so much as absorb him.
2. The contract is the consensus device. A contractarian can say the hypothetical agreement is not a substitute for consensus but a model of what consensus under fair conditions would look like.
3. Without an external standard, what stops a community consensus that is monstrous? His answer is the individual-freedom floor. Whether that floor is thick enough to do the work is the whole argument.

From the book

The whole of Chapter 3 is titled Liberalism Without Social Contract, and it is the chapter this page was guessing at.

What he rejects, precisely. Not the conclusions of any particular contract theory — the standing of a hypothetical agreement to settle anything:

“The root problem is that social contract theories, interpreted as the basis for general ethical rules, go against self-government. They don’t seek the actual consent of individuals and communities to public choices by a process of deliberation, communication, and debate. Rather, they attempt to circumvent any opposition by invoking high-level precepts without broad buy-in from the population.”

Rousseau gets the sharper version, and it is structural rather than about any specific general will:

“In every application of the social contract lies a contradiction. If it is the general will that supports true freedom, then true freedom must entail violations of individual freedoms conceived or misconceived differently from those enshrined in the general will… Under the general will, however, there is no room for deliberation or debate, and people cannot have any recourse against the general will.”

What he keeps. The chapter’s positive move is that contractarian arguments are perfectly good arguments — they are just not consent. Rawls’s own alternative reading of his project is quoted approvingly and then bounded:

“such reasoning should be used mainly in the context of deliberation and cannot justify actual policies or impositions of values on people, for example, becoming the basis of laws and how policymakers judge different options.”

By the chapter’s end, philosophy is demoted to an input and the deliberation is the thing:

“One can conceive of the role of political philosophy, including ideas based on treating others as we would want them to treat us, or contractarian arguments, especially those based on liberal principles like John Rawls’s, as providing input into building such principles via debate and deliberation.”

And he shows his work. The chapter’s two worked examples of consensus actually happening are the 2018 Irish abortion referendum, preceded by the Citizens’ Assembly, and the US path to the Respect for Marriage Act (2022) via the 2012 Maine, Maryland and Washington referenda. Both are cases where a genuinely contested value question was settled by argument among real people over time rather than by a principle derived in advance. That is what “real consensus among real people” means in his framework, and the interview never supplies it.

The floor is in Chapter 10, not here. The interview’s “elbow room” phrasing makes the floor sound undefined. It isn’t — Chapter 10 names it: freedom of expression, freedom to acquire education, equality before the law, and the right to exit the community. See Working-Class Liberalism and Community Elbow Room.

Reading

Working notes

What I’m confident about: the lineage, the general-will/will-of-all distinction, the “forced to be free” passage, and that step-1 framing determines the output. Standard material.

Checked against the book, and the guess was wrong. Draft 1 said “someone should read What Happened to Liberal Democracy? and check whether he names lexical priority.” Someone has. He never does — the word does not appear, and the priority ordering is never discussed as an ordering. His target is contractarianism generally, exactly as the fallback reading had it, and specifically the presumed-consent move rather than any principle derived from it. The From the book section replaces this. John Rawls carries the detail.

The consequence for this page is that it and The Book Problem do not merge into one argument. They were only going to merge if lexical priority turned out to be the target; it isn’t.

Where I think he’s weakest — still stands, and the book sharpens it. He objects to a rulebook, then names a floor that outranks everything else. Chapter 10 makes that floor explicit and four items long (expression, education, equality before the law, exit), which is more than the one line the interview implies, and it is asserted rather than deliberated. A four-item book is still a book. What the book adds in his defence is that the floor is justified instrumentally throughout — Chapter 2 argues rights are warranted because they enable experimentation and collective knowledge, “not because of a notion of natural rights.”

Draft 2 should add: what “community-level agreements” concretely means in his framework. This is now answered from Chapter 10 rather than constructed — see Working-Class Liberalism and Community Elbow Room.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech. Re-check wording against print before quoting as exact.

The metaphor, traced

It appears four times in about ninety seconds, and each appearance does different work:

1. The general will is a book. “we have a book in our hand which says when values clash… here is a higher value and we’re going to follow that.”
2. Rawls has one too. “even though he’s a true liberal, John Rawls also goes into that direction.”
3. Utilitarianism claims not to have one — "‘No, we don’t have a book. All we do is we turn everything to utils and then we compare those utils.’"
4. But it does. “Who decides how to compare, how to turn things into utils? When there are externalities that are important? There again needs to be a book.”

Why step 4 is the good move

This is the strongest argument Acemoglu makes in the philosophical half of the interview, and it is worth stating carefully because it generalizes.

Utilitarianism advertises itself as procedural rather than dogmatic: no list of sacred values, just a common currency and arithmetic. Acemoglu’s reply is that building the currency requires every contested judgment the currency was supposed to avoid:

• How do you weigh one person’s intense preference against many people’s mild ones?
• Do you count preferences people would not endorse on reflection?
• How do you price a life, a language, a landscape?
• Whose welfare counts — future people, foreigners, animals?
• How do you handle effects on third parties?

Each answer is a value judgment made before the arithmetic starts. If those choices are concealed, the result can look more neutral than it is. That is an objection to hidden assumptions; it does not establish that welfare comparisons must hide them or that an open but badly justified rule is better.

Anyone who has watched a cost-benefit analysis get used in a policy fight recognizes this immediately. The number looks objective. The assumptions that produced it were where the politics went.

The strongest reply from a defender of explicit welfare comparisons: publish the assumptions, show whose interests count, and test how the answer changes when the weights change. A comparison can inform a public decision without replacing it. Refusing numbers does not remove trade-offs; it can leave them implicit. On this view, acknowledging value judgments is a reason to make the analysis more inspectable.

Acemoglu’s reply would still have force: transparency does not give an analyst authority to bind the people affected. Even a documented calculation may miss domination, exclude a minority, or assume consent that was never given. But deliberation also needs scrutiny: who can participate, who sets the agenda, and what happens to dissenters? Neither method settles its own legitimacy merely by naming it. These are reconstructed arguments, not additional quotations from either speaker.

What he proposes instead

“once we have the building blocks of individual liberty, then I think we have to go back to society and seek consensus on what our temporary book should be.”

Three properties:

• Thin foundation. One thing is not up for consensus — meaningful individual freedom.
• Consensus-written. Everything above the floor is settled by real agreement among real people.
• Explicitly temporary. The book is revisable. That is the whole point.

The obvious objection

He still needs to justify the floor. If meaningful individual freedom limits what a community may decide, why is that restriction legitimate when some members reject it? A defender can argue that expression, education, equal legal standing, and exit make deliberation possible in the first place. These are conditions for meaningful participation, not simply four extra policies. That argument needs a defence, but counting the rules does not answer it.

The remaining disagreement is whether that justification succeeds, how the floor can be revised, and what counts as actual consent in a society where unanimity is rare. Explicit welfare analysis and deliberation could work together; the conflict arises when either claims authority to close the other off.

From the book

“The book” is not in the book. The metaphor is a spoken coinage — it appears four times in ninety seconds of conversation and nowhere in twelve hours of text. What the book has instead is the same commitment stated flatly, twice, as the closing move of two different chapters:

“Liberalism is a framework, rather than a set of cast-in-stone policies and rules that can help us navigate these trade-offs. But ultimate decisions will have to be made by self-governing societies.” — Chapter 2, closing the fourth pillar

“How these elements can be fused together and the trade-offs navigated should not be viewed as being cast in stone. Liberalism, as I have already emphasized, should be humble and understood as a framework for supporting freedoms, building collective knowledge, and navigating various trade-offs, rather than as a precise set of rules that can be easily applied in all cases.” — Chapter 3

So “temporary book” is a fair gloss on his position. “Written by consensus” is the part the book does more work on than the interview: Chapter 3 supplies two cases of a contested value question being settled by argument among real people rather than by a rule — the Irish abortion referendum and the American path to same-sex marriage — which is what the interview asserts but never demonstrates.

Where the temporariness comes from. Not from modesty. Chapter 2 builds the whole framework from four premises, and the first is the load-bearing one:

“Premise 1. Fallibility. Every belief, proposition, or piece of knowledge we hold could be wrong, even spectacularly wrong. This humility should be the starting point for any discussion of knowledge.”

He runs it back to Mill on the fallibility of ages and to Peirce’s fallibilism — “we cannot in any way reach perfect certitude or exactitude” — and forward through three more premises (social learning in community; gentle epistemic improvement; the desirability of collective knowledge) to four pillars: the vitality of experimentation, non-domination and self-government, economic growth, and soft communitarianism. The book is revisable because every belief is revisable. That is an epistemic argument, not a political preference, and it is the strongest available answer to “why should your book get to be temporary when Rousseau’s wasn’t.”

The utilitarianism move is interview-only — and that is worth knowing. Step 4 above is the sharpest thing Acemoglu says in the philosophical half of the conversation. It has no counterpart in the book. Utilitarian and utilitarianism appear nowhere in it; Bentham and Mill show up only as classical liberals who defined liberty as non-interference, and are criticized on that basis, not on the arithmetic. He is answering Cowen’s frame in real time, well, and off the page. Which cuts both ways: it is genuine live thinking rather than a rehearsed line, and it is also unbacked — the objection he presses on utils in ninety seconds is not one he has committed to print and had to defend.

The one-line book, re-measured. The interview makes the floor sound like a single item (“meaningful individual freedom”). Chapter 10 names it, and it is four:

“Liberal democracy must strive for maximum tolerance but should never accept community traditions and practices that suppress basic individual freedoms, including freedom of expression, freedom to acquire education, equality before the law, and the right to exit the community.”

Four named freedoms do not make his position weaker merely because four is more than one. The relevant questions concern their scope, justification, and revisability. Chapter 2 justifies property rights through their contribution to experimentation and knowledge, and explicitly allows limits on them. That is evidence for an instrumental defence of property rights; it does not establish that each freedom named in Chapter 10 can be overridden in the same way. A revisable rulebook could also contain such a justification. The live question is how Acemoglu’s framework resolves conflicts within the rights floor, and who has standing to revise it.

Working notes

The utilitarianism critique here is genuinely good and not original to Acemoglu — it’s close to Sen’s argument about the informational basis of welfarism, and to Amartya Sen and Bernard Williams’s Utilitarianism and Beyond (1982). Worth adding that lineage in draft 2 so it’s clear he’s joining an argument rather than inventing one. The book check makes this more interesting, not less: he isn’t joining that argument in print at all, so the interview is where he does it.

Unresolved and worth your judgment: when should a publicly justified principle constrain an actual community decision? The case for explicit comparisons is accountability about trade-offs. The case for deliberation is that affected people should help determine the terms. Each can be transparent and revisable; each can also exclude people. Property rights’ qualified status does not automatically settle the status of the four basic freedoms.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech — re-check against print before quoting as exact. The claim that utilitarian never occurs rests on that transcript, which is weaker than a printed index; what is solid either way is that no passage anywhere in the book argues against utilitarianism.

Related: Social Contract Theory and Its Critics · John Rawls · Henry Sidgwick · Rorty, Objectivism, and Moral Progress

The critique itself

Robert Lucas Jr., “Econometric Policy Evaluation: A Critique” (1976). Nobel 1995, substantially for this.

The setting: 1960s macroeconomics had large statistical models fitted to historical data, used to answer questions like “what happens to unemployment if we accept 2% more inflation?” The Phillips curve seemed to give a stable menu.

Lucas’s argument:

1. Those estimated relationships summarize how people behaved given the policy regime they expected.
2. People form expectations about policy and act on them.
3. Change the regime and expectations change, so behavior changes, so the estimated relationship changes.
4. Therefore using the old model to evaluate the new policy is invalid — the model’s parameters are not policy-invariant.

The 1970s obliged by demonstrating it: exploit the Phillips curve deliberately, workers build inflation into wage demands, and you get stagflation — high inflation and high unemployment, which the old curve said was impossible.

The methodological consequence was enormous. It pushed macro toward “microfounded” models built on preferences and technology — things assumed to be deeper than policy — which is where DSGE modelling came from.

The objection Cowen is making

The paper under discussion is Acemoglu, Autor, Beirne & Scott, Baby Busts and Growth Booms (NBER w35401, July 2026), which finds that falling birth rates are associated with higher GDP growth per working-age adult and higher wage growth, because technology responds to scarce young labor by becoming labor-saving. Identification comes partly from World War II deaths — historical variation in the size of younger cohorts.

Cowen’s move:

“there’s a lot of young labor missing after World War II, but right before us is this period of great population plenty, and to some extent, it’s expected. No one thinks there’ll be a 1.3 total fertility rate, whereas today, you have all these countries, some of them quite poor—Brazil, I think, is at 1.36 or so—and that just seems to be ongoing… Isn’t there a Lucas critique objection to your lack of fear of the fertility crisis?”

Unpacked, three distinct differences between the historical episodes and now:

Each breaks the extrapolation in its own way. If the shortage is global there is no elsewhere to import labor or capital from, so general-equilibrium effects differ. If it is permanent and anticipated, investment and family behavior adjust in advance in ways the historical data never contained. That is a textbook Lucas critique: the estimated response to labor scarcity was conditional on a regime that is ending.

Acemoglu’s answer — he concedes it

“Absolutely, 100 percent. Yes, that’s a very good point. The last paragraph of that paper says that.”

He then lists the cases himself — China at a scale never seen, South Korea, Brazil, Turkey, Mexico, India — and grants that “when the whole world goes through it… perhaps the general equilibrium, general world equilibrium effects are going to be different.”

His one counter is that another thing is changing simultaneously: people are living longer and healthier, so they can “accumulate more and more relevant human capital throughout their lives.” Note that this is not a rebuttal — it is a second unmodelled change, which is arguably a second Lucas problem rather than an answer to the first.

This is the most intellectually honest exchange in the interview. Cowen finds the weak point of a paper Acemoglu just co-authored, and Acemoglu agrees without hedging and points at where he’d already written the caveat down.

Reading

• Lucas, “Econometric Policy Evaluation: A Critique” (1976)
• Sargent & Wallace on policy ineffectiveness — the follow-on
• Acemoglu, Autor, Beirne & Scott, Baby Busts and Growth Booms — read the last paragraph, per his own pointer

Working notes

The framing of the three differences (local/global, temporary/permanent, unanticipated/anticipated) is my unpacking of a compressed question — Cowen gestures at all three but only states them loosely. I think it’s a faithful expansion, but it’s an expansion.

Checked against the book: not there. No Lucas critique, no rational-expectations discussion, no macroeconomic methodology at all. This exchange is two economists talking shop, and the book is not a macro book — its argument about technology runs through institutions and choices rather than through model stability. Nothing on this page needs revising against it, and nobody needs to search it again.

Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed; references are by chapter.

Related: King and Plosser - Real Business Cycles · Induced Innovation and the Habakkuk Thesis · Acemoglu and Restrepo - The Task Framework

The timeline

So how long?

There are three defensible endpoints, and the choice of endpoint is the argument:

Even the friendliest number is over a century. Acemoglu’s own framing — “think of American society or European society or even India or Latin America in the 1980s… human community was very well adjusted” — puts the settled state at roughly five hundred years after Gutenberg.

Was it worse than social media? Almost certainly, yes — and that’s Cowen’s point

The body count is not close. The Thirty Years’ War killed somewhere between 4.5 and 8 million people and depopulated parts of Germany by a fifth to a third. The French Wars of Religion killed millions more. No serious estimate puts social media’s direct harms in that range.

Cowen presses the case harder than the religious wars, and this is the strongest thing he says in the exchange:

“Say book markets and the printing press, they weaponized Marx and Lenin for many decades. That’s an algorithm in the very broad sense that led to terrible consequences, much worse than whatever you might think current social media have done, and yet we believe in free speech for the printing press.”

That is a serious argument. The 20th century’s totalitarian catastrophes were print-borne. If your standard for restricting a communication technology is “it has been used to cause mass death,” print fails the test far worse than TikTok.

Acemoglu’s answer is not “print was fine.” He grants all of it — “I don’t know what my reaction to that would have been” if he’d lived through the religious wars. His answer is that algorithms are categorically different, and the difference is targeting:

“Once algorithms can see, ‘Tyler has a weakness for this specific thing, I can fan the flames,’ I think that’s a completely new world.”

A book is broadcast: it says the same thing to everyone, and you have to go get it. An algorithmic feed is personalized and adversarial: it selects, for you specifically, the content most likely to capture you. His proposed remedy follows exactly from that distinction — not banning speech, but requiring chronological, non-algorithmic feeds. “That’s like the Myspace kind of model.”

Whether that distinction holds is the crux. Two ways to attack it: propaganda has always been targeted at susceptible populations (print runs aimed at particular regions, languages, classes), so the difference may be one of degree; and conversely, personalization at individual scale with real-time feedback may be a genuine phase change that has no print analogue.

Faster or slower this time? — arguments both ways

Faster:

• We have institutions now — regulators, courts, a body of media law — that took centuries to invent the first time. The adjustment does not start from zero.
• Adoption itself was near-instant, so the reaction cycle compresses too.
• We can see what is happening. Nobody in 1520 had a research literature on the effects of pamphlets.

Slower, or never:

• The technology keeps changing. The printing press was largely static for centuries; recommender systems are re-engineered continuously, and each change resets the adaptation.
• The adaptation is adversarial. Print did not optimize against your resistance; a feed does.
• The print settlement required the state to give up control (1695). The current analogue would require a handful of firms to give up optimization, and nothing forces them to.

My honest read: the printing press analogy is better at establishing that adjustments are long and violent than at predicting this one. It mostly refutes the complacent reading of itself — “we adjusted to books, we’ll adjust to this” is true only if you are prepared to accept a 150-year timeline with wars in the middle.

The finding worth taking away

Reading the timeline as a whole, the adjustment to print was not primarily people developing better habits. It was, in order:

1. Censorship (Index 1559, licensing regimes) — crude, state-run, largely failed
2. Political settlement (Westphalia 1648) — the wars ended by treaty, not by adaptation
3. Property and liability law (Statute of Anne 1710, libel doctrine) — who owns and who is answerable
4. Professional norms (much later — editors, mastheads, the idea of journalism)
5. Constitutional protection (1791) — last, and only once the rest existed

Acemoglu is proposing step 3 and Cowen is defending step 5. They are arguing about different points on a sequence that historically took two centuries to traverse — and the historical order runs against Cowen’s implicit assumption that free speech came first and worked.

From the book

The analogy is Cowen’s, not Acemoglu’s. The press appears once in twelve hours of What Happened to Liberal Democracy?, and not as a technology-adjustment story at all — it is Chapter 2’s example of a community refusing a technology, the Ottoman Empire that gave Europe inoculation while it “steadfastly opposed the dissemination of other valuable knowledge and the adoption of useful technologies such as the printing press.” There is no adjustment period, no counting of years, and no comparison to social media anywhere in the book.

The Reformation is in there, and it is on the other side of the ledger. Chapter 8, Social Engineering, treats the century after Luther as a case of institutions tightening rather than of a society adapting:

“In the centuries after the Reformation, following Martin Luther’s influential 95 Theses that launched Protestantism, the Catholic Church doubled down on indoctrination and thought policing. Beginning with the Council of Trent, 1545 to 1563, the Church tried to stop the spread of Protestantism by more closely policing religious practice and thought. Recent scholarship suggests that these efforts did more than slow down the advance of the dangerous ideas of Protestantism, they stopped scientific inquiry and even the sharing of scientific findings.”

That is a striking convergence with the finding this page arrives at independently — that the adjustment was institution-building rather than habituation — with the sign reversed. The page treats the censorship regimes as part of the settlement. Chapter 8 treats them as the disaster: the Index and the Inquisition are its examples of social engineering doing lasting epistemic damage. Both readings are available from the same events, and which one you take decides what “we adjusted” is supposed to prove.

On social media, the book does not use an adjustment frame at all. Chapter 9’s diagnosis is structural and current: an “algorithmic cage of norms” produced by a business model — “Social media companies have an incentive to build an algorithmic cage” because “digital ads exploit strong emotions and proclivities toward intense norm enforcement and thought policing.” His counterfactual is not that we will get used to it but that it could have been built otherwise: “if social media platforms did not algorithmically boost the most sensationalist and emotion-triggering content and were organized more like neutral repositories, accurate information could be obtained.” And the prescription is immediate rather than generational — “in the age of algorithms, the regulation of social media and AI should be at the top of the list.”

So the honest placement of this page is that it answers Adam’s question and Cowen’s frame, not a thesis of the book. Acemoglu’s on-air concession that the press had a violent adjustment period which is now over is a concession made in conversation; his book never argues it, and the mechanism it does supply for social media — designed incentives, correctable now — is one on which waiting for an adjustment is the wrong response.

Reading

• Dittmar, “Information Technology and Economic Change: The Impact of the Printing Press”, QJE 2011 — cities adopting print grew ~21 pp faster 1500–1600
• Rubin, “Printing and Protestants” — towns with a press by 1500 were 52 pp more likely to adopt the Reformation by 1530
• Eisenstein, The Printing Press as an Agent of Change (1979) — the canonical account
• Pettegree, The Book in the Renaissance (2010) — the commercial reality, a useful corrective to Eisenstein
• Wikipedia, Propaganda during the Reformation — orientation

Working notes

Numbers. Death tolls for the Thirty Years’ War and the French Wars of Religion are genuinely contested — the ranges given are the conventional spans, not precise figures, and the German population loss figure varies enormously by region and by whether you count plague and famine as war deaths (most historians do). The Dittmar and Rubin coefficients are from their published abstracts and are solid.

A weakness in my own framing. I’ve treated “adjustment over” as a single event, which is too clean. Print censorship in France persisted past 1695; the Index was not abolished until 1966. If you use the Index as the endpoint, the adjustment took five centuries. I picked endpoints that seemed defensible, but the choice is contestable and I’d rather you saw the seam than have it hidden.

Confounding worth stating plainly: the printing press did not cause the religious wars in a simple sense. The Reformation had causes of its own, and print amplified and accelerated it. Any comparison to social media inherits the same problem — it is very hard to separate the technology from what it carried.

Draft 2 could add: a side-by-side adoption-curve comparison (share of European cities with a press, 1450–1600, vs. share of adults on social platforms, 2004–2026). The print side is available from the incunabula short-title catalogue. That would answer the “faster or slower” question with a picture instead of an argument.

Checked against the book, and the framing needed a caveat. The press-as-adjustment analogy is Cowen’s; the book contains one passing reference to the printing press and no adjustment argument at all. Chapter 8 does cover the post-Reformation century, and reads the same censorship regimes this page counts as settlement as the central case of social engineering doing damage. That is a genuine tension between the page and the book, and it is now stated above rather than smoothed over. The endpoint question below is unaffected — the book takes no position on it either.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech. The claim that the press appears only once rests on that transcript and is the weakest kind of finding it can support; the positive readings from Chapters 2, 8 and 9 are solid.

Related: Social Contract Theory and Its Critics — the free-speech-absolutism question runs on the same tension between a floor of rights and a community’s room to regulate.

Strauss’s actual argument

In Persecution and the Art of Writing (1952), Strauss argued that philosophers living under censorship — Maimonides, Al-Farabi, Spinoza — developed a technique of esoteric writing. The text carries an orthodox surface meaning; careful readers detect a contrary teaching in the contradictions, the odd omissions, the deliberate errors.

Two claims are bundled here, and separating them matters:

1. Historical: philosophers did this, because they’d be killed otherwise. Defensible, and about specific texts.
2. Normative: they were right to, because some truths corrode the beliefs societies need to function. Far more contentious, and the source of the word’s modern political charge.

Strauss’s followers extended (2) into the idea of the noble lie — from Plato’s Republic, where the guardians propagate a founding myth to hold the city together. Whether Strauss himself endorsed this or merely described it is genuinely disputed. What is not disputed is that “Straussian” now means it.

What Cowen is actually proposing

Read his setup again, in order. First he lays out his own view plainly:

“There’s an educated elite in most countries. They decide who gets a Nobel Prize, who at MIT gets tenure… I think for the most part they know best. Other parts of the working class need to feel included enough not to bring the system down. But actual outcomes and policies, we do as much as possible want to be decided by the educated elite.”

Note the phrase “feel included enough not to bring the system down.” That is already the Straussian structure: the reality is elite decision-making; the management problem is public perception. Then he makes it explicit — “But we want to cover up the role of the elite to some extent, right? The Straussian view.”

Cowen is doing what he often does: stating the uncomfortable version of a widely-held position out loud to see whether his guest will own it.

Why Acemoglu won’t take it

His answer works on two levels, which is a nice irony given the topic.

The joke: “Okay, well, if you cover it up really well.” — i.e., the strategy only works if nobody notices, and everybody has noticed. The cover-up has already failed. Populist politics across the West is the evidence.

The substance: he immediately reframes to the problem he does care about — “how do you utilize and respect expertise without creating the wrong kind of technocracy that is insular and too self-confident and, frankly, looking down upon the rest?”

That distinction is his real position, and it is not anti-expertise. He is a Nobel laureate at MIT; he says so — “I am also very committed to expertise, obviously, given where I am.” His claim is that expertise and technocracy are different things, and that liberalism’s crisis came from conflating them: the educated elite became “both much more powerful and more numerous,” and a philosophical shift gave it more authority. His earlier line is the strongest statement of it:

“We cannot have true freedom and true flourishing of a population when there is a clear hierarchy, whether it is by education, it is by race, it is by income.”

The connection Cowen doesn’t make

The Straussian worry appears twice in this interview under different names. Here, and earlier as Henry Sidgwick — whose utilitarianism generated the charge of “government house utilitarianism,” the idea that an enlightened elite might rationally teach the public a simpler morality than the true one. Same structure, different century, different vocabulary. Acemoglu declines it both times.

Reading

• Strauss, Persecution and the Art of Writing (1952)
• SEP: Leo Strauss
• Plato, Republic III, 414b–415d — the noble lie itself
• Melzer, Philosophy Between the Lines (2014) — the best modern defense that esotericism was real and widespread

Working notes

The reading of Cowen here is mine and is somewhat uncharitable-adjacent — he may be probing rather than endorsing. But “they know best” and “feel included enough not to bring the system down” are his own words in his own voice, and he prefaces the whole thing with “that’s close to my view.” I don’t think I’m overreading it.

Draft 2 could add: Cowen’s public writing on elite competence and on “the great forgetting” would either confirm or complicate this. That’s a research task, not a lookup.

Checked against the book, which cannot settle this — it is a page about Cowen’s view, and Acemoglu’s book has no Strauss, no esoteric-writing discussion, and no treatment of elite competence in these terms. Two things in it do bear on the substance, though, and both cut against the Straussian reading: Chapter 2 identifies the abandonment of liberalism’s pillars with precisely this posture, “as the college-educated acquired more social power and started favouring a top-down, technocratic outlook”; and Chapter 10 makes non-denigration a principle rather than a courtesy — liberal society may argue for its values “without indoctrination, imposition, and denigration.” Whatever Acemoglu thinks of Cowen’s frame, his book has an answer to it. The draft-2 task above is unchanged: it needs Cowen’s own writing, not Acemoglu’s.

Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed; references are by chapter.

Related: Working-Class Liberalism and Community Elbow Room · Henry Sidgwick

First, who counts as working class

He redefines it, and the redefinition is doing serious work:

“If you’re getting your income from the labor market mostly, not from capital, not from your parents, then in my definition you’re part of the working class.”

This is enormous — it takes in nurses, electricians, software engineers, journalists, small-business owners, and (he says, laughing) tenured MIT professors, “the edge of the working class.” Cowen’s challenge is fair: if the category includes almost everyone, does it mean anything?

Acemoglu’s answer is that the unifying feature is not a shared job or a shared culture but a shared structural position — your income depends on your labor being wanted. And he accepts the cost: the class so defined is wildly heterogeneous, “a very diverse set of skills,” which is precisely why he needs communities to have their own practices and values. The elbow room is not a nice-to-have. It is required by the breadth of his own definition.

⚠ The book defines it much more narrowly, and this is the largest single discrepancy the book check turned up. Chapter 10, four sentences into the argument:

“By working-class… I mean people without college degrees in blue-collar or service sector jobs, a group that includes much of the native-born white population who have increasingly turned away from liberal policies.”

No nurses-to-MIT-professors continuum. No structural-position criterion. A conventional demographic category, and one defined against the college-educated — which is the axis the whole book runs on, from the “meritorious college educated and the rest” through post-industrial politics in Chapter 7. Cowen’s challenge in the interview is that the broad definition includes almost everyone and so means nothing. On the book’s definition the challenge doesn’t arise, because the definition is narrow and unremarkable. The expansive answer Acemoglu gives on air is improvised, and it makes his own project sound vaguer than it is. Everything below about heterogeneity requiring elbow room follows from the interview’s definition, not the book’s.

The structure, as I read it

The critical asymmetry: exit and voice are floor rights. A community can adopt practices you dislike; it cannot stop you leaving, speaking, or organizing against them. That is what makes local variation liberal rather than merely local.

My best guess at his example

He doesn’t give one, so this is construction, flagged as such. The domain where his framework has the most content is work:

This fits the automation half of the interview exactly. His example there is that German manufacturers redesigned jobs around robots to create new tasks, while American firms mostly did not — a difference not in the technology but in who had standing to shape its introduction. (The interview says “Japanese, Korean and German carmakers”; the book says German manufacturing, with Japan and the Nordics as comparable cases and no Korea — see below.) That is community-level agreement doing real economic work, and it is why he wants unions reconstituted rather than abolished, including the teachers’ unions he criticizes.

Applied elsewhere, the same shape gives you: national floor on what every child must be taught and every child’s right to leave; local discretion on how schooling is organized. Or a national floor on speech; local variation in norms about it.

The hard case — avoided in the interview, not in the book

What happens when a community’s agreement is illiberal — restricting women’s work, enforcing religious conformity, excluding outsiders? Cowen doesn’t press it, and the transcript leaves it hanging. The book does not. Chapter 10’s second challenge is titled Freedom versus Community and is about nothing else; see below.

From the book

Chapter 10, Working-Class Liberalism, is the chapter this page was reconstructing. It organizes the whole programme as five challenges: community values; freedom versus community; building voice and citizenship; shared prosperity in post-industrial society; and drains for the twenty-first century — his shorthand for public services and the safety net, from his own line that “the crisis of liberal democracy is also a crisis of drains and dole”.

The floor is named, and it is four items. Not the single “meaningful individual freedom” the interview implies:

“Liberal democracy must strive for maximum tolerance but should never accept community traditions and practices that suppress basic individual freedoms, including freedom of expression, freedom to acquire education, equality before the law, and the right to exit the community.”

Exit is singled out — “an essential recourse for escaping, if one wishes, from peer pressure and other norm-based impositions on individual choices” — and then, unusually, bounded in the other direction too: “this doesn’t imply that billionaires could invoke the right to move out of a country in order to avoid paying taxes. The right to exit should be there to protect individuals against the cage of norms, not for financial arbitrage.”

And he works the hard cases rather than gesturing at them. Sati and female genital mutilation as the clear violations. The Amish and the Hutterites curtailing children’s schooling, and Sharia law in certain British Muslim communities, as cases where “the American, British, and Canadian governments have acquiesced to such abridgement of individual liberties.” American states’ rights as the domestic parallel. Then, deliberately, a case that goes the other way — a community holding anti-gay views:

“If these remain in the realm of pure speech, without turning into incitement of violence or an abridgment of other people’s basic freedoms, the principle should be the same one that the ACLU applied in defending the neo-Nazis’ freedom of expression… It should be actions and real restrictions that matter, not intolerant words and thoughts.”

That is the dividing line the page above says he needs and does not draw: acts, not attitudes. It is a real answer, and it is a demanding one — it obliges a liberal society to tolerate community views it finds repellent while refusing community practices far short of repellent.

The elbow room is federalism, and the page guessed the wrong domain. The construction above built the example around work. The book builds it around governance:

“Fourth, public service provision has to be married with the community-level self-government agenda. Federalism is a useful tool in this regard… Federalism also allows different local units to pursue their own policy preferences which supports community-level experimentation.”

The reason is the Chapter 2 argument, restated in Chapter 10: “Experimentation is powered by variation. If all communities had the same values, true experimentation would not materialize.” Local variation is not a concession to communities that disagree; it is the mechanism by which a society learns anything.

But the workplace reading is not wrong — it is the fourth challenge. Chapter 10’s shared-prosperity section is squarely about who has standing to shape technology at work, and it does want unions reconstituted rather than abolished:

“Working-class liberalism should additionally give support to the labor movement so that unions can protect workers against abuses by powerful employers and rise to the challenge of promoting a pro-worker AI agenda. But the labor movement itself must change in the process as well… bargaining just for higher wages and hoping automation will be slow is not a realistic strategy. In fact, when there are plenty of automation technologies on offer, most companies would meet higher wage demands by choosing more automation.”

— with a specific and unusual ask: “union leaders themselves becoming sufficiently well-versed in AI and digital technologies to be effective in these negotiations,” leveraging the fact that “the workers themselves best understand their work, certainly better than tech companies do, and often better than their management does.”

So the honest correction is that elbow room and the workplace are two different challenges in his scheme, not one. Community elbow room is Challenges 1 and 2 and cashes out as federalism and a rights floor. Job design is Challenge 4 and cashes out as union bargaining over the direction of technology. The page above welded them together; the book keeps them apart, and the rights floor does no work at all in the workplace argument.

The carmaker claim, verified — with three corrections. Chapter 6 supplies it:

“German manufacturing firms have introduced industrial robots considerably faster than their American counterparts. The number of robots per industrial worker in Germany is about twice that in the United States. But German manufacturing companies must negotiate with worker representatives and unions when making these choices, and their blue-collar workers are often very well trained… while in the 1990s and 2000s American companies deployed digital technologies, mostly for automation, displacing and laying off blue-collar workers whenever they could, German manufacturing firms adopted a very different strategy. When introducing new technologies, they retrained their blue-collar workers for technical and supervisory occupations.”

The outcome claim holds: “robot adoption in Germany was not associated with large employment drops. Blue-collar work declined, but there were concurrent increases in technical occupations.” And the mechanism claim — that this was institutional, not technological — gets its sharpest support from within Germany: “in establishments where labor was better organized, there were less steep employment declines, and more pronounced efforts to reallocate blue-collar workers to technical tasks.” More robots, fewer layoffs, because of who was in the room.

Working notes

“My best guess” has now been checked, and it was half right. Draft 1 said the construction should be tested against What Happened to Liberal Democracy? before being treated as his view. It has been. The workplace framing is genuinely his — it is Chapter 10’s fourth challenge — but it is not what community elbow room means in his scheme, which is Chapters 1 and 2 of that list and cashes out as federalism plus a four-item rights floor. Draft 1 also flagged local government as the alternative a different reader might pick; that reader would have been right. The From the book section keeps the workplace material and re-labels it.

The carmaker claim is verified, in Chapter 6, with three corrections — Korea is not in the book’s list, it is German manufacturing rather than carmakers specifically, and the mechanism is retraining under Industry 4.0 rather than job redesign around robots. The table above has the detail. The correction needs to propagate to Acemoglu and Restrepo - The Task Framework, which carries the same claim.

The hard case is no longer “avoided.” That framing was true of the interview and false of the book, and the page said it about Acemoglu rather than about the transcript. Fixed above. What remains genuinely open is whether acts, not attitudes survives contact with the cases he does not take — a community whose norms are not enforced by anyone but are suffocating anyway, which is his own “cage of norms” from Chapter 2, and which the exit right is asked to carry alone.

Source and its limits. Checked against the audiobook edition (Penguin Random House Audio, narrated by John Lee), machine-transcribed, so references are by chapter and quotations are transcribed speech. Re-check wording against print before quoting as exact.

Related: Social Contract Theory and Its Critics · The Book Problem · The Straussian View · Pro-Worker AI