Baumol's Sawdust

On the limits of competition for deep wants.

At scale, markets sometimes sell us thin proxies for our deeper wants, because what we actually want cannot easily be put into contracts (the incomplete-contracts problem from the previous piece). This essay aims to explain why that equilibrium doesn’t seem to correct itself through competition, as one might expect.

Edward Hopper, New York Movie, 1939

Introduction

In the early 1800s, bakers began cutting bread with chalk, coffee merchants mixed in sawdust, and milk was thinned with water and whitened with plaster. The chemist Friedrich Accum documented in 1820 that as the distance between producer and consumer grew, sellers discovered they could replace the “real thing” with fillers and buyers couldn’t tell.¹ It took a generation of public outrage and legislation to reverse course.

Something similar seems to be happening to us now, for our deeper wants: things like companionship, love, adventure. What we find in the market (Grok companions, Tinder, Airbnb experiences) come with a faint aftertaste of sawdust or chalk. But if markets only deliver an approximation of what people want, that’s unmet demand. Why doesn’t competition close the gap? And why can’t we just avoid using markets for these things altogether?

There are many reasons for this, but I’ll focus on three: the price gap between what people want and a cheap proxy keeps widening (cost disease), the places people would otherwise go are disappearing (eroding infrastructure), and people raised on the cheap version can learn to want it (preference adaptation).

In the next essay, I’ll explore what institutions we could build to counter these effects at scale, and make markets deliver on the kinds of outcomes we actually want in a world of powerful AI.²

I. Cost disease

In 1966, the economist William Baumol observed that a string quartet requires the same input it did a hundred years ago. Over that period, manufacturing output per worker had climbed, driving down the cost of manufactured goods while pulling wages up. The musicians still needed wages competitive with sectors where productivity had risen, so the quartet had gotten much more expensive.

Baumol called this the “cost disease”: when some sectors see large productivity gains, the things that can’t ride those gains become harder to afford. Televisions, clothing, and computation get drastically cheaper year after year, while the real cost of a college education, a therapy session, and a hospital stay rises.³

In Coasean Compression, I explained how deep wants like a certain kind of companionship can be sold to us by being “compressed” into some proxy (for example, a want to be “met and seen” in dating gets compressed into “access” to a dating app). The cost disease means the price gap between the service delivering on the real thing and the service delivering on the proxy can widen over time: software scales with almost zero marginal cost while human labor gets relatively more expensive. Providers of the real deal come under pressure to either dilute what they offer or accept a shrinking market for high-end customers.

For example, a therapist is charging $200/session. A $10/month therapy app enters the market. People who try the app first get a fleeting sense of progress but might never learn they are missing some elements of “good” therapy (for example, staying with uncomfortable emotions for long stretches of time). As the app catches market share, our therapist could adapt by bundling her work into an app herself. But the thing her clients come for (a particular quality of attention that she brings, let’s say) translates poorly into app-form. So she sees her practice narrow to a smaller set of premium clients, those willing to pay $300, then $350, while the chatbot keeps getting cheaper.

This is not meant to say AI therapy has no role to fill. It could help millions of people get access to the many aspects of therapy where human presence isn’t always needed. But even still, an AI therapy app on a subscription model has an incentive to give users a fleeting sense of progress, just enough to keep them subscribed, but not enough to resolve the underlying problem, and without the uncomfortable friction (sitting with hard emotions, say) that might drive cancellations. Well-meaning developers face pressure to remove this friction, since competitors who do can undercut them on price.⁴ Public backlash is a check on this, but a slow one, especially when negative outcomes are diffuse (as was the case with social media).⁵

II. Eroding infrastructure

In the second century, an ordinary Roman farmer in Britain lived in a house with mortared stone walls and a tiled roof. The stone came from regional quarries, the tiles came from specialist kilns and the mortar required knowledge transmitted through complex apprenticeship networks. Five centuries later, after Rome’s fall, a wealthy Anglo-Saxon lord from the same place lived in a timber hall with a thatched roof, even though he had more money and power than the farmer ever did. He could, in principle, commission mortared walls and a tiled roof, but it would be extraordinarily expensive. Without Rome’s trade networks in place, he would have to personally fund a quarrying expedition, source materials, and import hard-to-find craftsmen.⁶

A similar dynamic has hollowed out much of our social infrastructure, which makes the price gap bite harder, and makes “exit” less viable, for social goods.

The pubs, churches, and neighborhood spots where loose ties form, so-called “third spaces”, work when enough regulars show up often enough to make showing up worthwhile. In America, that’s all been declining for decades.⁷ People increasingly work from home, stay in watching TV, order in food, and so on.

Many of the things people get from these spaces now have cheap digital alternatives: Facebook for connection, Tinder for dating, delivery apps for a restaurant visit, Peloton for the sports league. These aren’t always substitutes and sometimes complement, but they offer a low-friction option for people who were on the fence. Shared human settings depend on those people showing up: as the marginal bar goer chooses to stay home, order in and watch Netflix, the risk of a wasted evening rises for those who remain. They might drive twenty minutes to a half-empty bar, talk to no one they know, and go home thinking they should have stayed in too. The low-friction option now competes with an even emptier bar. The bar owner has to raise prices, which makes the calculus worse for whoever remains, which drives more people toward the alternatives.

In short, once certain attendance habits are gone, leaving the digital alternatives becomes increasingly less appealing. Someone who wants connection outside the feed may find the alternatives socially inert. New bars and clubs have to create a crowd from scratch, which takes capital, marketing, programming, driving up the price, and can also make the resulting place feel somewhat manufactured.

III. Preference adaptation

Sometimes, people who cannot get what they want adapt their preferences to what’s available, as the psychological cost of being unable to get it is too high. The philosopher Jon Elster called this sour grapes, building on the fox from Aesop’s fable who can’t reach the grapes and decides they were probably sour anyway.⁸

Elster’s work is theoretical, but I think we might be seeing something like this happen in certain markets, like dating. A majority of Americans believe relationships that begin on apps are just as successful as those that begin in person, even as almost 90% of recent users report frequent disappointment with the people they encounter, and couples who meet online report lower average marital satisfaction than those who meet through friends, church, or school.⁹

Someone who grew up swiping might not experience this as a loss. They might think of apps as how dating works, and find reasons to prefer it.

Meanwhile, the companies serving the compressed version see confirmation everywhere they look. When your business handles millions of strangers, it sees each customer only briefly and it might adopt a degraded view of what it means to serve them. Reed Hastings, Netflix’s co-founder, once described displacing social habits as a competitive victory: “Think about if you didn’t watch Netflix last night. What did you do? There’s such a broad range of things that you did to relax and unwind, hang out and connect — and we compete with all of that.” He added: “We actually compete with sleep. And we’re winning!”¹⁰ Mark Zuckerberg strikes a similar note, saying they’re simply “giving people what they want” when asked whether their recommender systems deliver real value or reward-hack their users.¹¹

The logic underneath this is called revealed preference: if people keep clicking, scrolling, and subscribing, they must be getting what they want. Modern economists are well aware of the limits of this reasoning. For example, the Nobel laureate Amartya Sen argued that people often choose out of habit, social pressure, or from a lack of alternatives. When the alternatives that would deliver on deeper wants get priced out or diluted in order to stay competitive at scale, or when the “third spaces” we’d ideally exit into don’t exist anymore, increasingly less of what we actually want remains on the shelf for us to choose from.

Countercurrents

There are forces pushing against this dynamic.

The most obvious is reputation. A festival with a track record for curation or good vibes can charge for it, and a competitor making false claims can be called out. Reputation works best for tight-knit communities (like niche music scenes), but at larger scales there are sometimes credible third-party reviewers (music critics for festivals, Michelin for restaurants) or aggregate rating systems like Google Maps or Amazon reviews (which have their problems, but do real work).

A second is counter-cultural demand. Once the compressed proxy saturates a market, opting for the alternative becomes a marker of taste. Gen Z are throwing phone-free parties in response to their generation being raised on algorithmic feeds, for example.¹² However, counter-culture starts on the margins by definition, and as it becomes more mainstream, it can get co-opted by the market forces it was supposed to rebel against; specialty coffee becomes Starbucks, and the organic food movement becomes Whole Foods. These capture the aesthetic of the original movement, but increasingly less of the substance.

Some movements have resisted this better than others. Open source software, for example, protected itself through licensing, and the ability to fork projects means it’s easy to defect when at risk of co-option. Burning Man’s codified principles and a structure where newcomers learn norms through camps are active defenses against the market equilibrium (a luxury festival where most people pay to consume an experience). It has not entirely escaped that pressure (there are plug-and-play camps and influencers that don’t contribute), but the culture has survived more than it should as the “festival” scaled to tens of thousands of participants. In both cases, the defenses against co-option are built intentionally into the structure of the movement, not left to culture alone.

A note on labor displacement

Baumol is sometimes used to tell an optimistic story about AI and labor. Alex Imas, for example, argues that as AI cheapens commodity production, rising incomes will shift spending toward sectors where human involvement is constitutive of the value: care, education, hospitality, therapy, theatre, craft.¹³ I think this is mostly right, but even if human labor is intrinsically valuable in these sectors, it doesn't necessarily mean the market makes that value legible.

A therapist with an established reputation can charge because clients are buying her in particular, and know what outcomes to expect from her. But in many cases, the buyer pays for a ticket or a session without knowing exactly what to expect, and can’t easily tell what they’d be missing in a cheaper version without human labor. A competitor can strip out most of the costly human work, sell the same nominal good at a lower price, and still claim to have delivered. If outcomes that require human involvement can’t be made legible, human labor competes only on the contractable dimensions. On those dimensions, AI labor and capital probably wins.

After all, people might already want to spend more on relational goods if real community and real belonging were actually on the shelves.

Baumol’s sawdust

I call the pattern I’ve described above Baumol’s sawdust. The price gap between the proxy and the real thing widens over time, so providers of the real thing face pressure to dilute. The places people would otherwise go thin out. And many of the people who grow up on the proxy may stop noticing it’s a proxy, or find reasons to prefer it.

AI could accelerate this trend towards sawdust. Millions of young people already spend hours a day talking to AI companions.¹⁴ Unlike a real friendship, which weaves you into a social fabric, a relationship with an AI enriches no one else. This thins the market for spaces where serendipitous friendships might form. Similarly, a good tutor challenges you to sit with hard problems, but AI tutors that gamify learning and reward quick answers might capture more of the market, and many students who never learned to struggle may stop seeking out the places that encourage it.¹⁵

Reputation and culture are throttles to all this, but they are slow and imperfect. At the Meaning Alignment Institute, we think there are institutional upgrades to markets that could more fundamentally reverse these dynamics. The next essay asks how to do that.

Thanks to Joe Edelman for generative discussions, and Maximilian Kroner Dale, Merlin Stein, Jan Kulveit, Alex Chalmers and Scott Moore for comments on earlier drafts.

1. Accum, F. (1820). A Treatise on Adulterations of Food, and Culinary Poisons. Accum documented widespread food fraud in industrializing London (bread bulked with alum, coffee with chicory and sawdust, milk with chalk) and named specific vendors. The book sold out rapidly and helped spark the public pressure that eventually led to the UK’s Adulteration of Food Act (1860) and Sale of Food and Drugs Act (1875).

2. For a broader treatment of how AI systems could gradually reduce human agency and capability, see gradual-disempowerment.ai.

3. Nordhaus (2008), “Baumol’s Diseases: A Macroeconomic Perspective,” The B.E. Journal of Macroeconomics. Using U.S. industry data from 1948–2001, Nordhaus confirms the pattern at macro scale: sectors with stagnating productivity show systematically faster price growth, while those with rising productivity get cheaper.

4. An analogy is Meta’s congress hearings. Their mission is to “give people the power to build community and bring the world closer together.” Hearings in 2018 and 2021 could pursue the company on data privacy and child safety, dimensions that are less ambiguous, but had no framework for evaluating whether Facebook was delivering on “community” or “closeness.”

5. Concerns about teen mental health, attention fragmentation, and the effect on democratic discourse built for over a decade before they translated into meaningful market or regulatory pressure. The harm was distributed across millions of small interactions rather than concentrated in identifiable incidents, which made it slow to read as harm at all.

6. Ward-Perkins, B. (2005). The Fall of Rome and the End of Civilization. Oxford University Press. Ward-Perkins uses archaeological evidence (pottery, roof tiles, coinage) to argue that the collapse of Rome’s trade networks destroyed productive capabilities, not just access to goods. In Britain, material culture fell to prehistoric levels within a generation.

7. Putnam, R.D. (2000). Bowling Alone: The Collapse and Revival of American Community. Putnam estimated the post-1965 decline in civic engagement was roughly 50% generational turnover (the WWII civic generation aging out), 25% electronic entertainment (television), 10% suburbanization and commuting, and 10% time pressures from dual-income households, with 15% unexplained. The decline was well underway before the internet, let alone social media. Bruni and Stanca (2008), “Watching alone: Relational goods, television and happiness,” Journal of Economic Behavior & Organization, find individual-level econometric evidence consistent with Putnam’s crowding-out claim: television viewing is associated with significantly lower consumption of relational goods, with the effect robust to instrumental-variable estimation addressing reverse causality.

8. Elster, J. (1983). Sour Grapes: Studies in the Subversion of Rationality. Elster argues that preferences frequently adapt to feasibility — people learn to want what they can get because the psychological cost of wanting the unavailable is too high. This is distinct from genuine preference change and makes revealed-preference reasoning unreliable in constrained choice environments.

9. SSRS Opinion Panel Omnibus (2024) finds 61% of Americans believe app-started relationships are just as successful as in-person ones. Pew Research Center (2023) finds 88-90% of recent dating app users report frequent disappointment. On marital satisfaction, see the Institute for Family Studies finding that couples who meet online score lower on average marital satisfaction than those who meet through friends, church, school, or college.

10. Hastings, R. (2017). Remarks at Summit LA17, November 3. Reported in Raphael, R., “Netflix CEO Reed Hastings: Sleep Is Our Competition,” Fast Company, November 6, 2017.

Dwarkesh Podcast@dwarkeshpodcast

"Maybe in the future, my AI girlfriend is on the other side of the screen or something." Mark Zuckerberg responds to Dwarkesh's fears of getting reward-hacked by AR.

4:53 PM · May 1, 2025 · 183K Views

---

10 Replies · 10 Reposts · 132 Likes

11.

12. Phone-free bars and restaurants now operate in at least eleven U.S. states, and The Offline Club, founded in Amsterdam in 2021, has expanded to nineteen cities. See “Phone-Free Spaces Grow as Gen Z Leads Digital Detox Drive,” Axios, April 24, 2026.

13. Imas, A. (2026), “What Will Be Scarce?” Ghosts of Electricity, April 14. Imas draws on Comin, Lashkari, and Mestieri (2021) on nonhomothetic preferences and on his own experimental work with Madarasz and Mandel showing that exclusivity and human involvement command a robust premium in willingness-to-pay settings.

14. Character.ai had 20 million monthly active users as of early 2025, with the average user spending roughly two hours per day on the platform. Over half of users are between 18 and 24. See Business of Apps, “Character.AI Revenue and Usage Statistics” (2026).

15. Bastani et al. (2025) find that after roughly ten minutes of AI assistance on math and reading tasks, participants performed worse unassisted and gave up more quickly. The authors argue AI conditions people to expect immediate answers, denying them the experience of working through challenges. See “The Consequences of Artificial Intelligence on Human Effort,” arxiv.org/abs/2604.04721