Category: Law
My excellent Conversation with Luis Garicano
Here is the audio, video, and transcript. Here is part of the episode summary:
Tyler and Luis start their conversation with Spain — housing, NIMBYism, and the productivity crisis; Spanish literature and why the Civil War still looms so large; and what Chicago taught Luis about party discipline in European politics. Then to the EU’s unanimity problem, capital markets, and Denmark’s flexicurity model; a round of overrated-versus-underrated on Rosalía, Penélope Cruz, and Sgt. Pepper’s; and finally into Messy Jobs — why retraining programs fail, whether AI’s leisure dividend has arrived, and whether Spain could ever achieve AI sovereignty.
Excerpt:
COWEN: Why is there a current productivity crisis in Spain?
GARICANO: Productivity indeed hasn’t grown for three decades, more or less. We’re currently having extensive growth. We’re having immigration, we’re having tourism, but we don’t really have productivity. A lot of it has to do with the political economy. I think if you want to explain the West, not just Spain, you have to understand who is voting and who is this being governed for. Spain is a particularly low-fertility, high life-expectancy country. We have the fifth-lowest fertility and the fifth-highest life expectancy in the world grosso modo, and very high pensions.
Essentially, all the GDP growth we’ve had has gone, 100 percent of the GDP growth we’ve had since 2008 has gone to pensions, to the pensioners. In terms of investment, there is very little in terms of productive investment. We have this fantastic highway network and this high-speed rail network. It’s not really getting the maintenance it needs. Just as one example, the country is basically being governed by and for the older retired people.
COWEN: What does the optimistic scenario look like? You don’t have to predict it’s going to happen, but lay out for me how it could all go well. You would get productivity growth of 1.5 percent a year, and economic growth a bit higher than that.
GARICANO: Spain has amazing fundamentals for the current situation, meaning we could easily be very electricity-energy rich due to solar, wind, and nuclear. We have all this empty space where you could put nuclear plants without much resistance. In fact, we are closing them for political reasons. The energy could be a big advantage. It’s a really amazing place to live. There is no more diverse geography and climate and nature anywhere in Europe, I think, or close to, and beautiful. You could easily see a situation where Spain becomes Florida, or Austin, Texas. Think of Texas. It attracts technology, attracts talent who wants to live there, attracts energy, builds the data centers, et cetera. That scenario is not impossible. The political economy is the tricky part.
COWEN: Doesn’t that mean you actually don’t have good fundamentals? You said you don’t even have a YIMBY movement. Life there really is quite good. I’ve been many times. It’s one of my favorite countries to visit. Isn’t that like a resource curse where there’s no sense of crisis? Old people live for a long time. It’s very comfortable. The weather’s great. The food is amazing. Aren’t those, in fact, liabilities in a time of very rapid change?
And another, on a very different topic:
COWEN: On Messy Jobs, your new and excellent book, you argue very persuasively, “In my view, AI will not lead to anything like mass unemployment, maybe not even to a rise in unemployment, because jobs will become messier and the AIs won’t be able to do them.” Is that a fair description of part of your argument?
GARICANO: I think that’s completely right. Yes.
COWEN: Now, I agree with you, but I think my worry is the opposite, that I know a lot of people, they want simple jobs, they’re okay with some measure of tedium, and that if most jobs become messy jobs, for them, that’s quite stressful, and they’re upset and disoriented. Do you worry about that?
GARICANO: I think that we will have to have a tolerance for human relations. Is that what your friends don’t want? We have a tolerance for relational work that is complex, that has politics in it, that has a big human component, and that’s the part that is going to stay. If people just want to be in their desk typing away, I think two good ways to think about it is work from home and outsourcing. If you think of which jobs were offshored, let me say offshored, in the big offshoring wave to India, those are jobs that are not messy. They’re clean. The company specifies the jobs. They say, “Okay, we can specify this job perfectly. Let’s ship you up.”
Those jobs are the same exact ones that are under complete threat of disruption. A lot of the work from home, when it doesn’t involve a lot of submittings, I guess, has the same feature. Those jobs are clean, single-task, and very often verifiable. You can just see how the performance is going and have your RL, your reinforcement learning loop work on those. I think those are gone.
The messy component that I want to emphasize, many people will say, “Oh, AI will tend to the messiness as well.” I think that there is some messiness that is contingent and that AI can streamline, but there is a lot of messiness that is both relational and has to do with a deeper aspect of the knowledge problem that doesn’t really go away. As AI advances, many aspects of Hayek, Polanyi, and all these knowledge problems are still there.
COWEN: How much retraining will be required and how frequent will that retraining have to be? If I think of me working with agents, I have to retrain myself every month or two. That’s difficult for me. It’s not stressful given my position, but I can imagine it would be stressful. Can we really just put a big chunk of the labor force through that?
Definitely recommended.
What should I ask Moxie Marlinspike?
Yes I will be doing a Conversation with him, live at the Roots of Progress event next week. From Wikipedia:
Moxie Marlinspike is an American entrepreneur, cryptographer, and computer security researcher. Marlinspike is the creator of Signal, co-founder of the Signal Technology Foundation, and served as the first CEO of Signal Messenger LLC. He is also a co-author of the Signal Protocol encryption used by Signal, WhatsApp, Google Messages, Facebook Messenger and Skype.
There is much more at the link, for instance he is also an anarchist of some kind or another. So what should I ask him?
The polity that is Singapore
Police in Singapore have charged a man who is accused of posting an AI-generated image of a saltwater crocodile in a popular reservoir.
Ye Lin was charged with communicating a false message and obstructing the course of justice for allegedly deleting the picture and the application he used.
The fake image caused public concern, authorities allege. The national water agency suspended its work at the city-state’s largest reservoir for two days last month after receiving information that a crocodile had been spotted.
Here is the full story, via Kyle.
*Shade*
The author is Sam Bloch, and the subtitle is The Promise of a Forgotten Natural Resource. An interesting book on a neglected topic, here is one excerpt:
Shade is not part of L.A.’s modern identity. In the 1930s, the city was rezoned to Federal Housing Administration design standards and banned high-density developments like row houses. Although apartments were once common, city leaders bowed to a prevailing wisdom that L.A. should not resemble a dark and cramped East Coast city. Freestanding single family-homes that were touched by sun on every side became mandatory. In came the cars. L.A.’s curbside trees were removed to accommodate shrinking sidewalks and expanding roads, and new rules that require parking minimums dealt another below to the urban forest. Mediterranean-style courtyards became endangered species as the shaded commons were converted to outdoor car storage. For decades, no building could be taller than the twenty-seven-story city hall…
Since the 1970s, an individual right to sunshine has been practically enshrined in state law.
The book also serves as an alternative history of Los Angeles (though it covers much more than that) through this alternative lens.
The Federal Lands: An Economic Property Rights Perspective
The US federal government owns and administers 472,892,659 acres or 21% of the land area of the lower 48 states, the country’s largest landowner. The resource is held and managed as a collective resource, the Federal Lands, through political and bureaucratic interpretation of the Multiple Use principle and generally, the biological aim of maximum sustained-yield. By contrast, access, exchange, and investment for most other US natural resources are through private property rights and markets. Despite the magnitude of the resource, economists have devoted relatively limited attention to the economic and welfare impact. The objective is to suggest economic implications and to encourage additional economic analyses. The discussion summarizes federal lands privatization through 1891, when withholding of federal lands began. The literature reveals no demonstratable market failure or increased resource scarcity from private exploitation between 1870 and 1957 when most lands were withheld. Because land was nonmobile and observable private property rights could have been assigned and any externalities addressed via Pigouvian restrictions or Coasean exchange. Federal ownership was not obviously required. Progressive Era reformers, driven by concerns of impending resource depletion, called for scientific, sustained-yield management by government officials. The institutional change is economically important. As outlined by Dixit and others, private rights holders have high powered incentives for efficient resource use that are lacking in decision making by agency officials who do not hold exchangeable property rights and do not directly bear the economic costs and benefits of their actions. Consequential public goods delivery could be an offset, but these are not measured for tradeoff calculations. Following Krueger, a rent-seeking framework is presented for comparing outcomes with economic property rights and political management. The analysis suggests that a.) federal lands will have lower production value than comparable private, all else equal; (b). federal lands management will be less responsive to shifts in economic costs and benefits. Public goods may be provided for high amenity, recreation, and ecological areas, but the dominant Multiple Use management principle provides no objective criteria for allocation or for periodic outcome assessment and adjustment. A literature review and data for contemporary federal forests, range, and oil and gas lands are provided.
That is from a new paper by Gary D. Libecap.
*Fear of Data*
The author is Omri Ben-Shahar, and the subtitle is How Privacy Panic Led Tech Regulation Astray — and How to Fix It. I would describe this book as bracing, and full of substantive engagement. Basically the author wishes to give privacy considerations less weight in social decisions. Excerpt:
What is the concrete evidence for the benefits of facial recognition technology in investigation of human-trafficking crimes? I would love to have found global estimates of the magnitude — of the trafficking victims rescued through the most advanced facial recognition methods — but all I have is a collage of reports [reports are then described].
One chapter is entitled “The Futility of Personal Rights.” Agree or not, this book is full of actual arguments, so I approve.
Optimal liability for offensive and defensive AI
How much liability should AI providers bear when their services enable both attack and defence? Liability can improve welfare while increasing harm. Providers sell a common input to productive users, attackers and defenders. Within a defended contest, a higher common price reduces effort without changing attack success or attacker profits, saving resources and improving the target’s security payoff. Compensation weakens defence and raises attacker profits. Optimal liability balances these effects against productive exclusion. Greater competition can lower optimal liability; every such decline must end at an outcome retaining defence. With cybersecurity access fixed, monopoly can warrant partial liability but never full liability when provision is worthwhile. When guardrails preserving productive uses are available, strong competition favours universal guarding socially but encourages unilateral removal at insufficient liability. At a fixed provider count, sufficiently many productive users ensure a pure equilibrium with universal guarding under high liability. A universal-guarding requirement makes liability redundant. Under monopoly, adoption follows a unique liability threshold, while zero liability remains uniquely optimal for a range of parameters with sufficiently many productive users.
That is from a new paper by Joshua Gans.
Banning self-recursive improvement in AI models?
Some people are suggesting this, including Ezra Klein.
I do not understand how it is supposed to work. Put aside the issue of foregone innovations, let us say I seek to access an AI from overseas. Am I allowed to visit their website? So many sites have AI behind them, including Chinese open source AI. Do all of those sites get banned? And how? Firewall imposed on Americans? And we cannot put the good open source models on our hard drives? How enforced? Does the government have to read the creation logs of the model (how can they?), and then decide whether I can use that model or not?
What if Anthropic licenses its IP to an independent subsidiary in the Cayman Islands and they keep on using code to improve the AI? Or maybe only foreign nations are allowed to have the best models and the really fast rates of improvement? How are multinationals supposed to operate across borders? (Presumably you cannot just use the RSI of your foreign affiliate, but then the whole MNC is crippled and eventually rendered uncompetitive?) How are we to keep “AI sovereignty” for Americans?
If the top U.S. labs have to give up on what is supposed to be really important, they will look very “catchable.” The policy creates a big incentive for some other parts of the world to invest a lot more in compute. How safe is that? And in the short run, do we also have to stop American firms from selling their compute abroad?
Does this whole thing mean that regular companies can use AI to write code, but the main labs cannot? Solve for the equilibrium there. Or the whole code writing function goes away for everybody?
What is the political mechanism through which RSI progress gets turned back on again, and when? Is it as rational as our current debate over data centers? If the top companies are no longer building RSI models, how are they supposed to know how safe or unsafe they might be?
At current margins, are humans even capable of writing the next steps of code that are required for further progress? At whatever speed?
I do not find this to be a workable idea. I think it is “people wanting to do something that sounds reasonable and less scary,” without thinking through the practical details. I have made a related point before, but one does best with AI policy when you start with the stuff that otherwise you might put into your last two or three paragraphs of an essay (“It remains to be seen how…but we are all human beings and surely we should not give up hope…”). Even if you believe in a deal with China, and I do not, we (and others) now know competitors can catch up more quickly than we had thought, at least if they think they have a chance of winning. They might even buy some space compute from Elon.
C’mon people, I know for many of you this is a difficult take to swallow, because it feels like “we don’t care enough” if we do not take some dramatic steps. Nonetheless the reality is that, a long time ago, we made a whole series of decisions that imply…we have no other choice than to just see this one through and get it more or less right. The longer we deny that truth, the harder it will be for all of us.
Obama on agentic AI
Obama recently said:
“If we are thinking about AI just in terms of how do we cure cancer or get better energy, you can do that without having agentic AI and having it just roaming free in the internet.”
As Roon noted: “this sounds completely incoherent to me”
From Rob Saker:
An agent is a system that can plan, use tools, write code, query data, run experiments, check its own work, and keep going. That is how the work gets done. Treating “agentic” as optional decoration, as if the serious version of AI is a polite chatbot locked in a box while the unserious version “roams free on the internet”, is the kind of sentence you write when you’ve heard the buzzwords and never watched a lab actually use the technology. You do not cure cancer with a model that only answers questions. You cure cancer with systems that can read the literature, propose hypotheses, design assays, analyze results, rewrite the next experiment, and do that loop a thousand times faster than a human postdoc. That loop is agentic. Strip the agency out and you are left with a very expensive autocomplete.
I agree with those points, but my main concern is different. How would we enforce such a prohibition on commercial agents? Set up a Chinese-like firewall that bans Americans from accessing sites with foreign agents? Monitor all those sites over time, so we know which suppliers to ban? Ban VPN as well? Give our government the power to inspect hard drives, in case agentic functions might be embedded there? Set up FBI “phishes,” luring Americans in with the prospect of agent access from abroad, and then arresting them, as we do with child ****?
Something else? How about insisting that all American (and foreign?) web sites set up tough captcha problems, so that agents may not be used (ha ha)?
You might also ask how an AI “agent” is to be defined, after all even O3 had some “agentic” abilities, such as opening up museum web sites to see which exhibits are on. Is it typing things into boxes and filling out forms that is to be prohibited? How much regulation of software would that require, how would that regulation actually occur, and what else would end up being restricted?
I do not think Obama intends to be supporting massive restrictions on freedom of speech and civil liberties. Rather this is a good example of how some political factions will simply have “ideas about what might be good,” without having freedom concerns — or for that matter practicality and civil liberties concerns — center of mind in the first place.
And this would cause the immediate bankruptcy of both Anthropic and OpenAI, right?
I can readily imagine that, upon the advent of agentic AI, we need some significant changes in our cyber laws. Now would be a good time to both commission and also cite some peer-reviewed academic literature on this matter!? That is not an impossible thing to do, yet our public discourse seems oddly resistant to the notion.
Addendum: And do not forget the classic Gwern piece, remarkably early in its prescience.
My excellent Conversation with Annie Lowrey
Here is the audio, video, and transcript. Here is the episode summary:
Annie Lowrey’s new book, The Time Tax, argues that the most regressive levy in American life is the hours each of us must spend on hold, in waiting rooms, and filling out forms a competent state would have pre-filled — roughly fifty a year on average, and far more if you’re poor. Tyler spends much of this conversation pressing the opposite case, asking whether that bureaucracy is quietly serving a useful purpose by rationing knee surgeries that don’t work, deterring fraud, or screening for immigrants with gumption.
So Tyler and Annie hash it out over an hour, covering why TANF and other welfare programs are the worst offenders, the political economy of submerged benefits like the child tax credit, whether the postal service should be zeroed out over ten years the way Denmark did, whether the government should be nudging people out of rural areas, lessons from the COVID unemployment fraud, areas where the time tax is flat instead of regressive, which private sector time taxes are most onerous, whether AI agents will pay these taxes for us within five years, the logic of easy benefits plus randomized harsh audits, the case for less federalism, whether voting should be easier or harder, how the UBI studies changed her priors, where you can find the best moss in the world, her favorite example of Brutalist architecture, tattoos as permanent ephemera, Christianity’s role in liberalism and the post-Christian right, how reporting beats tourism, the time tax of friendship, what she’ll learn next, and more.
Excerpt:
LOWREY: The postal service and Social Security’s retirement programs. Those are both pretty good. Social Security also has other programs, SSI, which is mostly used by very, very low-income seniors with disabilities or who are blind. Disability insurance is run by the Social Security Administration. The retirement program, it’s less than 1 percent overhead. It’s great. It does all the work for you.
COWEN: Isn’t our whole postal service a time tax in the aggregate? Denmark, just this year, abolished its postal service. They don’t have any. Now, I understand we’re not in a position to do exactly the same, but shouldn’t we be moving away from postal service, basically zeroing it out over the next 10 years?
LOWREY: Did they make it so that it’s all digital? Did they digitize things?
COWEN: Well, I don’t know who the “they” is. Direct mail there has gone away.
LOWREY: Oh, yes, that’s interesting. I don’t know. I think in a high inequality country such as ours, where you still have a lot of people who don’t have smartphones and don’t have computers, that’s why I would be very concerned about the postal service going away.
COWEN: That’s why we need 10 years. You want to tax people, tax in the broad sense of the term, living in rural areas, if only because it’s much harder to have government help them, right?
LOWREY: Yes.
COWEN: If rural areas don’t get mail anymore, maybe that’s a good thing.
LOWREY: I don’t think that that’s a good thing for the people in the rural areas, though. There’s always going to be people in rural areas. As a general point, do you want the government to be encouraging people to not live in rural areas?
COWEN: Absolutely, that’s me. Welcome to the show.
LOWREY: That’s feels very nanny state-ish of you, Tyler.
COWEN: I don’t want to subsidize rural areas, and we do it in so many ways, including with our farm programs.
LOWREY: We do. We really, really absolutely do.
COWEN: It gets harder to help people with government. There’s a lot more market power in rural areas, like how many supermarkets are there, how many hospitals, probably just one, if that. Let’s get everyone into the suburbs, or heaven forbid, the cities.
LOWREY: I can’t imagine a less popular policy than that.
COWEN: Absolutely.
Interesting throughout, and full of good spirit. Again, here is Annie’s very good book The Time Tax: How Government Wastes Our Time — And How to Fix It.
What Regulatory Capture Actually Looks Like
It’s amazing how a theory can take over a brain. Consider the idea that people believe what serves their interests. As heuristics go, it’s a good one. I use it all the time. Yet when Dario Amodei says AI is dangerous, perhaps even an extinction risk, some people conclude he must be running a marketing campaign. That is stupid. Which is more likely, that a useful heuristic sometimes misfires or that “our product might kill you” is a clever way to sell it? Death threats are a poor marketing strategy.
We also have plenty of evidence that the fears of AI experts are sincere. Amodei, Altman and Musk were all publicly warning about AI risk long before they had AI companies to promote. The worry runs well beyond the executive suite; rank and file researchers share it. And it extends outside the industry altogether, to computer scientists with no product to sell, among them Nobel laureate Geoffrey Hinton. Hinton left a high-paying job at Google precisely so he could speak out and Hinton is not in a Berkeley polycule with Eliezer Yudkowsky, at least as far as I know. Whatever else you may say about the belief that AI presents a serious risk, plenty of AI researchers believe it sincerely.
Similarly, when Amodei recently proposed to slow the pace and install independent safety teams at AI companies many people jumped to the conclusion that this was regulatory capture. Sorry, but no, that theory doesn’t make sense. To see why, we should review the theory of regulatory capture.
Regulatory capture came out of the political science literature especially Marver Bernstein’s 1955 classic, Regulating Business by Independent Commission. Bernstein argues for a regulatory life cycle: Gestation, Youth, Maturity, Old Age. A scandal brings a bureaucracy into existence—or gives an existing one new powers. The public’s attention, like Sauron’s eye, fixes on the issue of the day: something must be done. The thalidomide scandal, for example, helped establish the modern FDA.
Gestation gives way to a youthful burst of reform and the do-gooders come to Washington ready to battle the industry. Inevitably, however, the public’s eye looks elsewhere. But the industry never looks away. It lobbies Congress, hires former regulators, trains future ones, and supplies much of the information the agency needs. As the agency matures, accommodation replaces confrontation. By old age, the regulator has become the industry’s protector. The classic example is the ICC, created to regulate the railroads but it eventually came to shield them from competition from the trucking industry.
Notice that classic regulatory capture takes time, it’s a process of erosion rather than a battle, it happens in the shadows, in the backrooms, away from the public’s eye. As Culpepper argues in Quiet Politics and Business Power, business power goes down as political salience goes up. Regulatory capture and lobbying does a good job explaining why roasting coffee beans was defined as “domestic manufacturing”, thereby lowering Starbuck’s tax rate by 2%. It does less well at explaining big cross-industry issues the public cares about such as environmental regulation or race and gender discrimination regulation. Finally, don’t confuse capture with firms making the best of a bad situation. Philip Morris supported the 2009 Tobacco Control Act not because FDA regulation was Philip Morris’s unconstrained ideal but because it knew regulation was coming and it wanted a seat at the table to nudge the rules in its favor. That’s ordinary political bargaining—or rent-seeking—not evidence that the regulator has been captured.
Now let’s evaluate Amodei’s call for regulation in light of regulatory capture theory. AI regulation is in gestation. Public attention is fixed on the industry, and much of that attention is hostile. The big profits in AI lie in automating work, and job loss is a much more salient fear than extinction. AI politics is now loud–precisely the environment in which Culpepper predicts business power will be weakest. A mature industry can bend regulation to its purposes through revolving doors, longstanding relationships and obscure rulemaking. An industry under Sauron’s eye has much less power and faces much greater risk that politics will bend regulation to its purposes. Political actors are eager for an excuse to redistribute AI rents away from capitalists and toward favored groups (ala Peltzman).
Regulation will reduce AI profits. That doesn’t prove that every rule Amodei favors is innocent of self-interest, an absurd proposition. I suspect that Amodei’s ideal may be something like a single regulated AI monopoly—safe and reasonable profitable, like the old AT&T. But that’s not the profit maximizing outcome. If transformative AI can capture even a fraction of the enormous labor market–the world’s biggest market–laissez-faire would mean vastly greater profits. Amodei may simply prefer a smaller fortune and a safer world. That is perfectly consistent with self-interest playing a role; it is not consistent with the bastardized theory that profit maximization is the only thing that matters or that “capture” is universal.
Go ahead: argue that Amodei and other AI experts are wrong about AI risk. Ask whether his proposals favor Anthropic. But calling “our product might kill you” a clever marketing and regulatory-capture strategy isn’t sophisticated analysis. The facts don’t fit regulatory capture theory and trying to make them fit requires epistemically painful Ptolemaic epicycles. Even a dull Ockham’s razor cuts through that story to the obvious alternative: Amodei actually believes what he’s saying.
Did the ACA reduce mortality?
Many of us brought up related points at the time, but basically we were booed off the reservation:
While recent research has provided evidence that the Medicaid expansions of the Affordable Care Act (ACA) reduced mortality, there is no evidence on the effect of the Affordable Care Act (ACA) net of the Medicaid expansions on mortality. This is an important gap in knowledge because the ACA significantly increased health insurance coverage in non-expansion states. In this article, we exploit the large increase in health insurance coverage brought forth by the ACA to examine the effect of the ACA and Medicaid expansions on mortality. Unlike prior studies that relied solely on geographic variation in Medicaid expansions to estimate the net effect of the expansion, we use a novel empirical approach that allows us to investigate the effect of the ACA net of Medicaid expansion on mortality, the incremental effect of the Medicaid expansion, and the overall effect of the ACA including Medicaid expansion. We use longitudinal data from the NHIS Linked Mortality Files (LMF) and a nationally representative sample of 40 to 58-year-olds combined with a difference-in-differences and a difference-in-differences-in-differences research design to obtain estimates of the effect of the ACA on mortality. We find no evidence that the Medicaid expansions had a beneficial effect on mortality but do find that the ACA net of Medicaid expansion reduced mortality.
That is from a new NBER working paper by
AI, Redistribution, and the Size of the Pie
Anthropic’s economic team, including Anton Korinek and Chad Jones, have a valuable new paper, Economic Scenarios for Transformative AI. They make their assumptions explicit and provide a scenario explorer that lets you change them. How capable will AI become? How quickly will firms adopt it? Will it augment workers or automate their tasks? You can see what different answers imply for growth, wages, and unemployment.
In their extreme scenario AI takes on a lot of tasks, GDP is 32.4% higher by 2030 than without AI and labor share declines from 60% to 45.2% but they make this striking point:
“Total labor income in 2030 in the extreme scenario is almost exactly what it would have been without AI: the labor share falls by a quarter while GDP rises by a third, and 0.45×1.32 ≈ 0.60.”
Exactly. That is the central point of my paper, How Much Redistribution Will AI Require. A falling labor share does not necessarily mean falling labor income. Workers can receive a smaller share of a much larger economy and still earn as much as they would have without AI.
Using the code behind their scenario explorer I updated their results to a 10 year horizon and plotted them on my redistribution graph. Only under the modest scenario is some net labor transfer required to make AI Pareto improving at the aggregate level.

Aggregate labor income, of course, conceals differences among workers. Korinek et al. find that cognitive occupations lose income while other occupations gain. In their extreme scenario, restoring the cognitive occupations’ wage bill to its no-AI level would require about 9% of GDP. They argue that compensation on this scale in response to technological change has no precedent.
I think this makes the adjustment problem look too pessimistic.
First, adjustment happens through retirement and entry. A retiring accountant need not retrain as a nurse. A young person enters nursing rather than accounting. Neither becomes unemployed even as labor reallocates. Korinek et al. understand these channels but give them limited scope in their model. Admittedly, those margins don’t do much work by 2030 but they matter over ten years.
Second, compensation can take the form of shifting taxes from labor to consumption. In the long run labor’s share of consumption tends to equal its share of GDP, so the lower labor’s share becomes, the more relief a given tax shift provides. At a 45% labor share, each dollar shifted from labor taxation to consumption taxation reduces labor’s net burden by 55 cents. Shifting taxes worth 5% of GDP would thus provide net relief to labor of 2.75% of GDP, without increasing total tax revenue. Unemployed workers would still need payments, but compensation need not come entirely through additional government spending.
Third, we do have experience expanding income support rapidly. U.S. unemployment benefits reached approximately 2.5% of GDP in 2020, and that during a contraction. Britain’s compensation to slaveowners following abolition amounted to roughly 5% of GDP in a one-time settlement. These episodes show that governments can mobilize substantial resources for compensation. Moreover, the extreme AI scenario brings an enormous increase in output from which to finance compensation.
Preserving aggregate labor income does not protect every worker. But even the extreme Korinek scenario reinforces the point that a dramatic decline in labor’s share can coexist with stable or increasing aggregate labor income. To the extent labor income does decline, growth makes compensation affordable and attrition, entry, and tax shifting can make the intra-labor task smaller than it first appears.
Does AI assistance enhance or erode expertise?
From a new NBER working paper:
Whether AI assistance builds or erodes professional expertise is unsettled. In a pre-registered three-month randomized controlled trial, we gave 133 practicing patent lawyers at eleven U.S. intellectual property law firms access to a custom AI drafting assistant and measured both their performance while using AI and their professional judgment afterward without it. All work was scored by blinded expert patent attorneys. Paralleling findings from other white-collar domains, AI access raised the quality of work delivered on benchmark patent drafting tasks at 10 days (0.34 SD, p = 0.03) and 90 days (0.38 SD, p = 0.01), with larger gains among junior lawyers. After three months, all subjects redlined an existing patent application without AI, a core task of patent practice requiring expert judgment. Treated lawyers outperformed controls by 0.32 SD (p = 0.04), but this advantage was concentrated entirely among senior lawyers (0.45 SD, p = 0.02). Junior lawyers showed no average gain; their scores instead bifurcated, with sharply fewer mediocre scores offset by more poor and more good ones. The largest gains from AI thus accrued to the lawyers who retained the least. Foundational expertise may be a prerequisite for extracting durable skill from AI-assisted practice.
That is by David Autor, et.al. Do note that over time the allocation of humans to tasks will evolve so that more of the humans become more productive, not less. RCTs somehow have the odd disadvantage of requiring too many things to be held constant, and so they can miss the benefits of longer-term adjustments.
Diversity Is Our Strength?
Diversity is our strength is a common motto. Indeed it is one of GMU’s core values but what is the scientific evidence for this thesis? A new scoping review:
Recent years have witnessed many strong claims that ‘diversity’ leads to more original and impactful science, which is a science-focused form of what we call “The Diversity Hypothesis.” However, what evidence supports the claim that diversity enhances scientific output or impact? This pre-registered rapid scoping review seeks to collate and evaluate the scientific evidence for the Diversity Hypothesis…
…Based on over 100 scientific articles, we find that only between 15% and 28% of results reported in the literature are consistent with the hypothesis, with the balance of the results not being consistent with it.
…Overall, the results of the analysis section indicate that there is little empirical evidence that diversity improves scientific output and/or impact. In fact, with the possible exception of disciplinary diversity—a type of informational or viewpoint diversity operationalized at the team level—the majority of the evidence seems to point in the other direction. These findings are robust regardless of how the data is parsed and analyzed. The same conclusions can be drawn looking at the full data set, only the population-adjusted results, or only the results of high quality based on the MMAT analysis.
Hat tip: Colin Wright.