Results for “manipulation” 91 found
Currency manipulation doesn’t actually work so well
That is the topic of my latest Bloomberg column, here is one excerpt on the empirical side:
It is also worth keeping in mind a number of empirical points. First, the ECB has not historically been all that expansionary. Rather, it is renowned for a fairly tight monetary policy. Eurozone rates of price inflation are usually below 2%, and that does not seem about to change.
Second, up through the late 1990s, Chinese currency manipulation consisted of keeping the value of the currency “too high” rather than “too low.” Yet in those earlier times, Chinese exporters still were gaining ground. And since 2005, the Chinese currency has risen considerably — arguably, it has attained the levels one would expect in a normally adjusting market. More recently, the Chinese central bank may be propping up the currency, to limit capital flight from China. That is currency manipulation, but in a manner that will damage Chinese exports, not help them, and indeed China probably is headed toward having permanent trade deficits with the rest of the world.
Finally, countries with low household savings rates tend to run trade deficits, and of course that is the U.S., with savings rates usually below 10% and often as low as 4%. Obviously, if you spend most of your money, some of those expenditures will go abroad, and that will hurt your trade balance. Whether or not you think that is a problem, America’s savings shortfall has little to do with Chinese currency manipulation.
Of course President Trump and yes also Elizabeth Warren are the main offenders here. Read Warren’s Medium essay on these topics, it is shocking in its crude nationalism: “A Plan for Economic Patriotism.“
Intrade Manipulation Fail
Brad Plumer at the Wonkblog discusses a recent attempt to manipulate Intrade.
On Monday night, after the debate, Barack Obama was leading Romney on Intrade by around 60 percent to 40 percent. But at around 10:00 a.m. on Tuesday morning, Romney surged to 48 percent. Was this evidence that the conventional wisdom was wrong? Had Romney actually won the debate handily? Or, alternatively, was the nosedive in the stock markets putting a dent in Obama’s re-election chances?
Neither. As economist Justin Wolfers pointed out on Twitter, the huge swing toward Romney appears to have been driven by a single trader who spent about $17,800 buying up Romney shares and pushing the Republican candidate’s chances on Intrade up to 48 percent. But the surge only lasted a few minutes before other traders whittled the price back down to what they saw as a more accurate valuation. Romney’s odds of winning are currently back at around 41 percent.
…As Wolfers pointed out, this mysterious trader ended up overpaying by about $1,250 for shares that quickly collapsed in value. Was this just someone who made a bad trade? Or was somebody trying to influence Intrade odds in order to sway perceptions of the race? And if so, was it worth $1,250 to jolt the markets for less than 10 minutes?
Plumer quotes me from 2008 discussing an earlier attempted manipulation:
This supports Robin Hanson’s and Ryan Oprea’s finding that manipulation can improve (!) prediction markets – the reason is that manipulation offers informed investors a free lunch. In a stock market, for example, when you buy (thinking the price will rise) someone else is selling (presumably thinking the price will fall) so if you do not have inside information you should not expect an above normal profit from your trade. But a manipulator sells and buys based on reasons other than expectations and so offers other investors a greater than normal return. The more manipulation, therefore, the greater the expected profit from betting according to rational expectations.
Addendum: Justin Wolfers offers more comment.
Manipulation of Prediction Markets
As many people suspected someone was manipulating Intrade to boost John McCain’s stock price:
An internal investigation by the popular online market Intrade has revealed that an investor’s purchases prompted “unusual” price swings that boosted the prediction that Sen. John McCain will become president.
Over the past several weeks, the investor has pushed hundreds of thousands of dollars into one of Intrade’s predictive markets for the presidential election, the company said.
This is big news but not for the reasons that most people think. Although some manipulation is clearly possible in the short run, the manipulation was already suspected due to differences between Intrade and other prediction markets. As a result,
According to Intrade bulletin boards and market histories, smaller investors swept in to take advantage of what they saw as price discrepancies caused by the market shifts – quickly returning the Obama and McCain futures prices to their previous value.
This resulted in losses for the investor and profits for the small investors who followed the patterns to take maximum advantage.
This supports Robin Hanson’s and Ryan Oprea’s finding that manipulation can improve (!) prediction markets – the reason is that manipulation offers informed investors a free lunch. In a stock market, for example, when you buy (thinking the price will rise) someone else is selling (presumably thinking the price will fall) so if you do not have inside information you should not expect an above normal profit from your trade. But a manipulator sells and buys based on reasons other than expectations and so offers other investors a greater than normal return. The more manipulation, therefore, the greater the expected profit from betting according to rational expectations.
An even more important lesson is that prediction markets have truly arrived when people think they are worth manipulating. Notice that the manipulator probably doesn’t care about changing the market prediction per se. Instead, a manipulator willing to bet hundreds of thousands to change the prediction of a McCain win must think that the prediction will actually affect the outcome. And if people think prediction markets are this important then can decision markets be far behind?
Hat tip to Paul Krugman.
Privatizing Law Enforcement: The Economics of Whistleblowing
The False Claims Act lets whistleblowers sue private firms on behalf of the federal government. In exchange for uncovering fraud and bringing the case, whistleblowers can receive up to 30% of any recovered funds. My work on bounty hunters made me appreciate the idea of private incentives in the service of public goals but a recent paper by Jetson Leder-Luis quantifies the value of the False Claims Act.
Leder-Luis looks at Medicare fraud. Because the government depends heavily on medical providers to accurately report the services they deliver, Medicare is vulnerable to misbilling. It helps, therefore, to have an insider willing to spill the beans. Moreover, the amounts involved are very large giving whistleblowers strong incentives. One notable case, for example, involved manipulating cost reports in order to receive extra payments for “outliers,” unusually expensive patients.
On November 4, 2002, Tenet Healthcare, a large investor-owned hospital company, was sued under the False Claims Act for manipulating its cost reports in order to illicitly receive additional outlier payments. This lawsuit was settled in June 2006, with Tenet paying $788 million to resolve these allegations without admission of guilt.
The savings from the defendants alone were significant but Leder-Luis looks for the deterrent effect—the reduction in fraud beyond the firms directly penalized. He finds that after the Tenet case, outlier payments fell sharply relative to comparable categories, even at hospitals that were never sued.
Tenet settled the outlier case for $788 million, but outlier payments were around $500 million per month at the time of the lawsuit and declined by more than half following litigation. This indicates that outlier payment manipulation was widespread… for controls, I consider the other broad types of payments made by Medicare that are of comparable scale, including durable medical equipment, home health care, hospice care, nursing care, and disproportionate share payments for hospitals that serve many low-income patients.
…the five-year discounted deterrence measurement for the outlier payments computed is $17.46 billion, which is roughly nineten times the total settlement value of the outlier whistleblowing lawsuits of $923 million.
[Overall]…I analyze four case studies for which whistleblowers recovered $1.9 billion in federal funds. I estimate that these lawsuits generated $18.9 billion in specific deterrence effects. In contrast, public costs for all lawsuits filed in 2018 amounted to less than $108.5 million, and total whistleblower payouts for all cases since 1986 have totaled $4.29 billion. Just the few large whistleblowing cases I analyze have more than paid for the public costs of the entire whistleblowing program over its life span, indicating a very high return on investment to the FCA.
As an aside, Leder-Luis uses synthetic control but allows the controls to come from different time periods. I’m less enthused by the method because it introduces another free parameter but given the large gains at small cost from the False Claims Act, I don’t doubt the conclusion:
The results of this analysis suggest that privatization is a highly effective way to combat fraud. Whistleblowing and private enforcement have strong deterrence effects and relatively low costs, overcoming the limited incentives for government-conducted antifraud enforcement. A major benefit of the False Claims Act is not just the information provided by the whistleblower but also the profit motive it provides for whistleblowers to root out fraud.
I was curious, so I asked
My prompt:
I read that ChatGPT will be starting connections with Shopify, Stripe, and perhaps other companies for LLM commerce, as you might call it. That is, you could search for something in the app and then buy it through GPT. OpenAI would receive some sort of commission, as Google does now if you buy something through a Google search. Using standard theories of tax incidence, provide an analysis of a) optimal pricing strategy for OpenAI and its partners, and b) where will the incidence of those new fees fall? On consumers? Retailers? Taken away from Google? Other?
The Claude 4.5 answer is here under the fold, sadly the extensive footnotes and citations do not transfer:
# LLM Commerce Economics: Market Structure, Pricing Strategy, and Tax Incidence Analysis
**OpenAI’s entry into transactional commerce on September 29, 2025 through partnerships with Shopify, Stripe, and Etsy marks a fundamental shift in digital commerce intermediation.** With an estimated 2% commission rate—dramatically lower than Google Shopping’s historical 12% or Amazon’s 15%—OpenAI’s model raises critical questions about optimal pricing strategy, competitive positioning, and where the economic incidence of new commerce fees will ultimately fall. The answer depends on elasticities, market power dynamics, and the structural differences between LLM-mediated and search-based commerce.
## OpenAI’s announced partnerships reveal aggressive market entry pricing
OpenAI officially launched “Instant Checkout” on September 29, 2025, enabling direct purchases within ChatGPT conversations through confirmed partnerships with three major platforms. **Stripe serves as the foundational payment infrastructure partner** (since March 2023), co-developing the open-source “Agentic Commerce Protocol” that enables encrypted payment tokens across different processors. **Etsy went live immediately** with single-item purchases from U.S. sellers available to ChatGPT Plus, Pro, and Free users. **Shopify integration is rolling out to over 1 million merchants** including major brands like Glossier, SKIMS, Spanx, and Stanley 1913.
The business model operates through commission-based revenue, though specific rates remain confidential. **Sam Altman publicly hinted at approximately 2% affiliate fees** in a Stratechery interview, positioning this as an alternative to traditional advertising models where OpenAI would “never take money to change placement.” This represents market-entry pricing 83% lower than Google Shopping’s former 12% commission rate and 87% lower than Amazon’s typical 15% referral fees.
**OpenAI emphasizes that merchants remain the merchant of record**, retain customer relationships, and control fulfillment—a significant departure from Amazon’s model. The company explicitly states that fees do not influence product ranking or placement, with products ranked by relevance to user intent, availability, price, quality indicators, and whether Instant Checkout is enabled. This positions OpenAI’s model closer to affiliate marketing than traditional marketplace economics, but with substantially greater influence over the consideration set through AI curation.
The technical architecture matters for competitive dynamics. The Agentic Commerce Protocol is open-sourced and platform-agnostic, working across payment processors beyond Stripe. This suggests OpenAI aims to establish infrastructure standards rather than create a closed ecosystem—potentially enabling rapid scaling while reducing merchant lock-in concerns that might otherwise dampen adoption.
## Current commerce market structure shows high but declining commission rates
The online commerce landscape OpenAI enters is characterized by **dominant players charging substantial take rates** and **emerging competitive pressure** from social commerce platforms. Understanding existing commission structures and market shares is essential for predicting where OpenAI can extract value and how competitors will respond.
**Amazon maintains 37-40% of U.S. e-commerce** with a complex fee structure that effectively extracts 30-35% of transaction value when including referral fees (8-45% by category, typically 15%), FBA fees, and advertising costs. The company generated nearly $600 billion in GMV in 2024 and over $50 billion in advertising revenue, demonstrating how marketplace power enables layered monetization beyond transaction fees.
**Google fundamentally restructured its commerce model** by eliminating commissions entirely for “Buy on Google” (previously 5-15%, averaging 12%) to compete with Amazon and social platforms. Google now relies exclusively on advertising revenue from Shopping Ads using cost-per-click pricing. This strategic pivot acknowledges that Google’s advantage lies in discovery rather than transaction completion, and that charging commissions reduced merchant adoption.
**Social commerce represents the fastest-growing segment** at 30-36% CAGR, reaching $1.16-1.63 trillion globally in 2025. TikTok Shop charges 6-8% commissions (most categories 6%), while Instagram and Facebook charge 5% per shipment plus 2.9% processing fees. These platforms are capturing impulse purchases and discovery-driven shopping, particularly among younger demographics.
**Traditional affiliate networks** operate on variable commission structures ranging from 5-30% depending on product category, with Amazon Associates paying 1-10% (averaging 3-5%). YouTube Shopping charges approximately 15% median commissions. These models demonstrate that recommendation-based commerce has historically commanded premium rates compared to search-based discovery.
The competitive landscape shows **declining commission rates due to platform competition and merchant bargaining power**. Google’s elimination of transaction fees, TikTok’s aggressive 6% pricing, and the proliferation of direct-to-consumer channels via Shopify have created downward pressure on take rates. OpenAI’s ~2% estimated commission must be understood in this context—it’s not just below-cost pricing for market entry, but recognition that merchants now have credible alternatives.
## LLM commerce fundamentally differs from search in ways that affect elasticities
The economic analysis of commission incidence requires understanding how LLM-mediated commerce differs structurally from traditional search, as these differences directly impact price elasticities, switching costs, and market power—the key determinants of tax incidence.
**User intent and conversion dynamics diverge significantly**. ChatGPT shopping users demonstrate 8% higher engagement, 12% more pages per visit, and 23% lower bounce rates compared to traditional search traffic. However, Adobe Analytics data across 1 trillion+ visits shows generative AI traffic is initially 9% less likely to convert than other sources, despite these engagement advantages. The trajectory matters: shopping queries on ChatGPT doubled from January to June 2025, and click-through rates tripled from 2.2% to 5.7% in the same period, suggesting rapid improvement as users develop comfort with the channel.
This creates a **high-consideration, lower-immediate-conversion funnel** fundamentally different from search’s high-intent, direct-conversion model. The economic implication: LLM commerce currently excels at the discovery and consideration stages where it reduces search costs for consumers, but underperforms at transaction completion due to trust and habit factors. This bifurcated purchase journey means LLM platforms may capture influence value without capturing transaction volume, creating misalignment between contribution to sale and commission collection opportunity.
**Trust factors introduce significant friction with economic consequences**. State-of-the-art LLMs achieve approximately 80% accuracy on benchmarks, meaning one in five recommendations could contain errors. Unlike traditional search where users see multiple sources and can verify independently, LLM commerce presents synthesized, curated results where verification asymmetry creates information problems. This trust gap appears in behavior: 46% of shoppers report being unlikely to let AI manage their entire shopping trip according to Walmart surveys, and 72% of commerce leaders cite privacy and security as moderate to major roadblocks for AI implementation.
The trust deficit has direct implications for elasticities. When consumers don’t fully trust recommendations, they’re more likely to conduct additional research or comparison shopping before purchasing, reducing the price insensitivity that platforms need to extract surplus. Trust must be built through accuracy and transparency, creating a time dimension to elasticity estimates—current elasticities may not reflect long-run behavior once trust develops.
**Integration depth creates new switching costs absent in search**. Traditional search imposed low switching costs: users could easily try different search engines, and no accumulated personalization locked them in. LLM commerce creates multiple lock-in mechanisms through personalization-driven data accumulation. The system learns user preferences through conversation history, product interactions, and stated requirements, creating a bilateral learning dynamic where both user and system lose accumulated knowledge when switching platforms.
Quantified impacts demonstrate these effects: AI-powered sales convert at double the rate of traditional search traffic, and 64% of AI-powered sales come from first-time shoppers according to Rep AI data, suggesting strong conversion advantages for platforms with established user relationships. However, the paradox exists that AI can simultaneously reduce technical switching costs (through instant data migration and format conversion) while increasing psychological switching costs through preference attachment.
**Market power dynamics shift from indexing to intelligence**. Search-based commerce derived competitive advantage from index comprehensiveness, ranking algorithm quality, and distribution deals (default placement). LLM commerce shifts market power sources to model sophistication, proprietary training data, user interaction history accumulation, and computational resources for model training and inference. This creates fundamentally different barriers to entry—crawling infrastructure versus GPU clusters, SEO optimization versus Generative Engine Optimization (GEO), and link-based ads versus integrated recommendations.
The network effects differ critically. Traditional search showed limited network effects (better search doesn’t necessarily require more users), enabling long-term market stability with Google maintaining 90%+ share for 25+ years. LLM commerce demonstrates strong network effects where more users generate more query data, enabling better recommendations, attracting more users in a virtuous cycle. However, the competitive dynamics appear more contestable: OpenAI’s market share among AI chatbots declined from 50% in 2023 to 25% in 2025 as Anthropic captured 32%, demonstrating that even with network effects, technological leapfrogging remains possible.
## Two-sided platform pricing theory provides the analytical framework
Standard economic theory of two-sided markets, developed by Rochet and Tirole and extended by Weyl, provides the essential framework for analyzing optimal pricing strategy for OpenAI and predicting commission incidence.
**The fundamental insight of two-sided platform theory** is that platforms must “get both sides of the market on board” by balancing price allocation between users through internalization of indirect network effects. The structure of prices—how total platform fees are divided between buyers and sellers—matters as much as the level of fees. Platforms rationally subsidize one side of the market (even below-cost) to attract participation that creates value for the other side.
In LLM commerce, this manifests as **zero fees to consumers while charging merchants approximately 2%**. This allocation reflects several economic factors. First, consumers have substantially lower search costs in the AI era—they can easily try ChatGPT, Perplexity, Google Gemini, or traditional search without switching costs. Second, consumers don’t internalize the value they create for merchants through their attention and purchase intent. Third, merchants have fewer credible alternatives for reaching 700 million weekly ChatGPT users at high-intent shopping moments. Fourth, network externalities run primarily from consumers to merchants (more shoppers attract merchants) rather than vice versa.
**Weyl’s (2010) framework on monopoly pricing for networks** illuminates why OpenAI prices at approximately 2% despite market power. Profit maximization requires internalizing network externalities to marginal users, not average users. The marginal merchant considering integration weighs OpenAI’s 2% commission against alternatives: Amazon’s 15% plus advertising costs, Google Shopping’s 0% commission but higher customer acquisition cost through ads, TikTok Shop’s 6%, or direct channels. OpenAI must price to attract the marginal merchant who has credible alternatives, even if inframarginal merchants (those without good alternatives) would pay more.
The empirical content of platform pricing models depends critically on user heterogeneity, not just aggregate demand elasticities. Merchants differ substantially in their valuation of ChatGPT distribution: a DTC brand with weak Amazon presence values new discovery channels highly, while established Amazon sellers with optimized FBA operations may view OpenAI as supplementary. OpenAI’s commission structure cannot easily discriminate among merchant types (beyond category-based variation), limiting its ability to extract differential willingness to pay.
**Recent empirical work demonstrates that commission fee regulation often backfires** through unintended platform responses. Li and Wang (2021) studied commission fee caps in 14 U.S. cities for food delivery platforms and found that independent restaurants paying reduced fees experienced declining orders and revenue, while chain restaurants paying original fees saw increased orders. The mechanism: platforms discriminated against independent restaurants post-regulation by reducing recommendations and increasing consumer delivery fees in regulated cities.
This finding has direct implications for LLM commerce. Platforms possess multiple levers beyond commission rates: product ranking, recommendation frequency, prominence in results, consumer-side pricing, and service quality allocation. If OpenAI faces pressure to reduce or cap commission rates, economic theory and empirical evidence suggest it would adjust these other dimensions, potentially harming precisely the merchants that fee caps aimed to help.
**Tremblay (2023) shows that greater fee discrimination allows platforms to serve more markets but increases aggregate double marginalization**, creating a welfare tradeoff where platform profits increase with fee differentiation but social welfare decreases. Yu (2024) demonstrates that commission fee changes affect not just prices but also quality and innovation—welfare analysis ignoring these quality adjustments underestimates policy impacts by 0.91-2.06 percentage points.
The implication for OpenAI’s optimal pricing strategy: uniform low commissions (approximately 2%) may be preferable to complex discriminatory pricing, both for merchant adoption (simplicity, fairness perception) and regulatory risk mitigation. The company’s explicit stance against paid placement reinforces this strategy, positioning commissions as a neutral tax on transactions rather than pay-for-performance positioning.
## Tax incidence theory predicts burden distribution across market participants
Standard tax incidence theory, adapted to two-sided platform markets, predicts where the economic burden of OpenAI’s commission fees will fall. The critical insight is that **statutory incidence (who legally pays) diverges from economic incidence (who bears the burden)**, with distribution determined by relative elasticities, market structure, and pass-through mechanisms.
In a two-sided platform context, a commission rate τ imposed on merchants creates three potential incidence pathways: merchants absorb through reduced margins, consumers bear through higher prices, or the burden falls on displaced competitors (Google, Amazon) through lost market share and pricing power.
**The elasticity principle determines primary incidence**. The side of the market with more inelastic demand or supply bears greater burden. Consumers face relatively elastic demand for most products due to extensive substitutability in e-commerce—if prices rise, they can switch products, brands, or platforms. However, LLM commerce may reduce price elasticity through several mechanisms: curated results show fewer options; AI recommendations frame price within context of fit and quality rather than isolation; trust in recommendations reduces comparison shopping; convenience value raises willingness to pay.
Merchants face potentially inelastic demand for platform access if they lack good alternatives. For merchants currently underserved by Amazon or Google Shopping (artisan creators, niche products, emerging DTC brands), ChatGPT represents non-marginal new distribution opportunity. These merchants can’t easily substitute away from OpenAI without losing access to 700 million weekly users. However, established merchants with optimized Amazon/Google presence face more elastic platform demand—they can reduce ChatGPT focus if commissions or effective costs rise.
**Pass-through rates vary systematically by merchant characteristics and product category**. Economic theory and empirical evidence show that pass-through from cost increases to consumer prices depends on demand curvature (not just elasticity), market structure, product substitutability, and merchant cost structures.
In LLM commerce, pass-through mechanisms operate through multiple stages: Platform (OpenAI) charges commission τ → Merchant faces increased cost per transaction → Merchant decides whether to raise prices, absorb cost, or reduce quality → Consumer responds to any price increase by adjusting quantity demanded or switching to alternatives.
The Congressional Research Service analysis of Digital Services Taxes provides relevant guidance. DSTs function as excise taxes on intermediate services, with burden typically falling on purchasers of digital services and downstream consumers through price increases. The Tax Foundation’s empirical evidence across Austria, France, and UK shows that platforms commonly pass costs to advertisers and merchants, who then pass to consumers—DSTs function as consumption taxes in practice.
Critical insight: **when platforms have market power and merchants face competition, platforms can successfully pass through cost increases**. OpenAI possesses market power from model quality, user base, and network effects. Merchants face intense competition within product categories on the platform. This asymmetry enables OpenAI to maintain commission rates while merchants absorb costs or attempt partial pass-through to consumers.
However, the current low 2% commission rate significantly limits pass-through potential. For a $100 product, the $2 commission represents 2% of transaction value. For a merchant operating on 20% gross margins, this represents 10% of margin. If the merchant attempts to pass through the full commission, the consumer would see a 2% price increase. Given typical e-commerce price elasticities (often greater than 1 for discretionary goods), a 2% price increase would reduce quantity demanded by more than 2%, making pass-through profit-reducing for most merchants.
**The more likely incidence outcome**: merchants absorb most of the 2% commission through reduced margins, with minimal pass-through to consumers. This burden distribution would differ markedly from the counterfactual of higher commission rates (10-15% like traditional platforms), which would force substantial pass-through or merchant exit.
**Incidence on displaced competitors represents the third pathway**. If LLM commerce captures market share from Google and Amazon, the burden manifests as reduced revenues, lower advertising pricing power, and loss of data network effects for these incumbents. This is not traditional tax incidence but rather incidence through competitive displacement.
The magnitude of this effect depends on substitutability between channels. Current evidence suggests complementary rather than purely substitutional relationships: users employ LLM commerce for discovery and consideration while still completing transactions on familiar platforms. Adobe data showing 9% lower conversion rates for AI traffic but 8% higher engagement suggests LLM commerce currently directs traffic to traditional platforms rather than capturing the full transaction.
However, as Instant Checkout adoption increases and user trust develops, the substitution effect will strengthen. Google faces particular vulnerability: its elimination of shopping commissions to compete with Amazon demonstrates declining pricing power. If ChatGPT captures product discovery at 2% commissions while Google can’t monetize through commissions at all, the incidence manifests as Google’s lost advertising revenue and market power erosion.
Amazon faces different dynamics. Its high commissions (15% referral fees plus FBA plus advertising) create margin for competitive entry, but its conversion rates, fulfillment infrastructure, and Prime ecosystem provide substantial moat. The incidence question becomes: do Amazon merchants view ChatGPT as additive distribution (expanding total sales) or substitutional (shifting sales between channels)? If additive, Amazon bears minimal burden. If substitutional, Amazon sees commission revenue per merchant decline as sales shift to lower-rate channels.
**Network effects amplify incidence impacts through dynamic adjustments**. When consumers shift attention to ChatGPT for shopping, this reduces the value of Amazon and Google to merchants (weakening indirect network effects), potentially forcing these platforms to reduce fees to retain participation. When merchants integrate with ChatGPT, this increases platform value to consumers, strengthening OpenAI’s network effects and pricing power.
Birge et al. (2019) demonstrate that platforms should charge different commissions to different trader types based on network positions, even when traders have identical value distributions. Revenue loss can be unbounded when charging uniform fees. This suggests OpenAI’s current uniform 2% commission may be suboptimal in the long run—differentiated pricing by merchant or category could extract more surplus as the platform matures.
Even-Tov et al. (2022) show that retail investors are highly responsive to commission fee changes, with fee elimination leading to 30% increases in trading frequency. The retail trading context provides analogy: when commissions represent small absolute amounts but meaningful percentages of transaction sizes, users demonstrate high elasticity. This supports the prediction that merchants and consumers will be highly responsive to LLM commerce commission rates, justifying OpenAI’s low-rate market entry strategy.
## Market power considerations determine long-run pricing and incidence
The long-run incidence of LLM commerce fees depends critically on whether OpenAI and competitors maintain pricing discipline or engage in fee competition, which in turn depends on market structure, barriers to entry, and competitive dynamics.
**OpenAI currently possesses substantial market power** from several sources: ChatGPT holds 80.1% of AI-driven search market share; 700 million weekly active users create reach advantages; proprietary GPT-4 and successor models demonstrate technological leadership; accumulated conversation data enables superior personalization; and brand recognition in conversational AI exceeds competitors.
However, this market power faces important constraints. **Barriers to entry are high but not prohibitive**. Anthropic captured 32% of enterprise LLM market share by 2025 (surpassing OpenAI’s 25%), demonstrating that technological leapfrogging remains possible. Open-source models like Meta’s Llama (9% enterprise share) prevent complete platform closure. Google possesses distribution advantages through Chrome, Android, and Search integration. Amazon has Rufus shopping assistant integrated with its marketplace.
The FTC analysis identifies control over AI’s essential building blocks—training data, specialized talent, computational resources, and platform effects—as sources of market power distinct from traditional search. Data network effects create barriers as platforms with more user interactions improve recommendations faster, attracting more users. OpenAI’s 700 million weekly users generate interaction data that competitors must overcome through superior model capabilities or alternative data sources.
**Switching costs cut both ways in determining market power sustainability**. Personalization creates psychological attachment and bilateral learning lock-in, supporting platform pricing power. But low monetary switching costs (AI platforms are free to users) and AI-assisted data migration reduce technical barriers to switching, maintaining contestability.
The Brookings Institution research on firm-level AI adoption (2010-2018) shows that AI investments increase industry concentration, with largest firms benefiting disproportionately through data advantages and computing resources. The effect creates winner-take-more dynamics, but the research also notes rapid technological evolution prevents permanent dominance.
**The landmark Google antitrust case provides crucial precedent** for understanding competitive dynamics. Judge Mehta’s September 2025 ruling noted that “the emergence of GenAI changed the course of this case” and declined to order Chrome divestiture because AI competition was deemed “highly competitive” with “numerous new market entrants.” ChatGPT was mentioned 28 times, OpenAI 30 times, demonstrating judicial recognition that LLM platforms created competitive constraints on Google’s search monopoly.
The American Action Forum analysis emphasizes that market power in AI is “transitory” because “ChatGPT hit 1 billion searches/day in less than 2 years; Google needed 11 years.” This suggests that even substantial market share doesn’t guarantee sustained pricing power when technological substitution occurs rapidly.
**Multi-homing behavior by merchants reduces platform market power**. If merchants easily integrate with ChatGPT, Google Shopping, Amazon, and TikTok Shop simultaneously, platforms must compete for merchant attention and consumer recommendations. The Agentic Commerce Protocol’s open architecture facilitates multi-homing by reducing integration costs. However, if consumers consolidate shopping queries on a single LLM platform due to personalization quality, this gives the dominant platform substantial leverage over merchants despite multi-homing capability.
Song (2021) demonstrates that platform mergers may benefit some agents by lowering prices or attracting more users on the other side through indirect network externalities, showing that market concentration doesn’t necessarily harm all participants. In LLM commerce, if OpenAI gains dominance, merchants might benefit from reduced integration complexity (single platform rather than many), but lose bargaining power over commission rates.
**The critical policy levers for maintaining competitive pricing discipline** are data access and interoperability. The Brookings, FTC, and NBER analyses converge on the insight that data market operation determines whether market power emerges and persists. If shopping interaction data, merchant product feeds, and user preference information can flow across platforms, competition remains viable. If proprietary data creates insurmountable moats, winner-take-all dynamics enable monopoly pricing.
OpenAI’s open-sourcing of the Agentic Commerce Protocol suggests recognition that standards-based interoperability supports rapid merchant adoption and reduces regulatory scrutiny, even if it constrains long-run platform lock-in. Stripe’s co-development ensures payment infrastructure remains open across platforms, preventing vertical foreclosure concerns.
**Recommendation bias emerges as the key market failure concern**. Unlike search advertising where paid placement is clearly marked, LLM recommendations blend algorithmic judgment with potential commercial influence. If OpenAI can bias recommendations toward higher-commission merchants or partners while maintaining plausible deniability (“our AI determined this was the best fit”), it gains ability to extract surplus through implicit rather than explicit pricing mechanisms.
The FTC’s emphasis on algorithmic transparency and prohibition of exclusionary conduct becomes critical. If platforms face credible regulatory constraints on recommendation bias, commission rates remain the primary revenue extraction mechanism. If platforms can subtly favor commercial partners in recommendations, explicit commission rates understate true merchant costs of platform access.
Lin, Pan, and Zheng (2020) show that platform pricing interacts with strategic buyer behavior and cost expectations in complex ways that can contradict conventional wisdom. Applying this to LLM commerce: merchants’ willingness to pay commissions depends on expectations of platform growth, consumer adoption trajectory, and competitor behavior. OpenAI’s initial low 2% rate may be strategically optimal even with market power if it accelerates merchant adoption and consumer trust, enabling higher rates later as merchants develop platform dependency.
## Optimal pricing strategy balances adoption, extraction, and competitive response
Synthesizing the theoretical frameworks, empirical evidence, and market structure analysis yields predictions for OpenAI’s optimal pricing strategy and likely commission incidence distribution.
**For OpenAI’s optimal commission rate**, the 2% level appears strategically sound for market entry and early growth phase, justified by:
**Adoption economics**: Low rates reduce merchant integration friction and eliminate price objections, enabling rapid scaling to critical mass of product coverage. Network effects mean platform value is convex in merchant participation—reaching 1 million Shopify merchants creates substantially more than 10x the value of 100,000 merchants. Sacrificing short-term commission revenue for faster network growth likely maximizes long-term platform value.
**Competitive positioning**: The rate undercuts all major competitors (Amazon 15%, Google Shopping’s historical 12%, TikTok Shop 6-8%), creating clear value proposition for merchants. It’s low enough to appear insignificant to merchants operating on 20-30% margins, reducing scrutiny and negotiation friction. The rate positions ChatGPT as additive distribution channel rather than Amazon alternative, reducing competitive retaliation risk.
**Regulatory strategy**: A “small fee” below typical platform rates reduces antitrust scrutiny. The explicit prohibition on paid placement and rank manipulation preempts pay-to-play concerns. Open-sourcing the Agentic Commerce Protocol signals non-proprietary intentions. Low rates make merchant complaints less likely, reducing political pressure for regulation.
**Elasticity calibration**: At 2%, commission represents small enough percentage of transaction value that merchant pass-through is profit-reducing for most products given typical demand elasticities. This means OpenAI captures revenue while minimizing consumer price distortion and deadweight loss.
However, the long-run optimal rate likely exceeds 2% if OpenAI achieves sustainable competitive advantages. The platform could increase commissions to 4-6% (matching TikTok Shop range) once merchant dependency develops through: (1) proprietary user data accumulated through shopping interactions; (2) GEO optimization by merchants specifically for ChatGPT; (3) consumer habit formation around ChatGPT shopping; (4) branded product discovery (“find it on ChatGPT” as consumer behavior).
The timing and magnitude of rate increases will depend on competitive dynamics. If Perplexity, Google Gemini, or other LLM platforms offer shopping at 0-1% commissions, OpenAI faces constraint. If competitors adopt similar 2% rates, industry norm develops enabling eventual increases.
**Alternative revenue models may dominate commission extraction**. Sam Altman’s comments suggest preference for subscription tiers with shopping benefits over advertising, and skepticism toward traditional ads. OpenAI could extract surplus through: Premium memberships with better shopping features (perhaps commission-free purchases for Pro subscribers); data licensing to merchants (aggregated shopping insights); SaaS fees for merchant tools and analytics; API fees for third-party shopping agents built on ChatGPT; advertising with strict quality constraints (“tasteful ads” that maintain user trust).
The optimal long-run model likely combines low commissions (maintaining participation) with premium subscription value (extracting consumer surplus) and merchant services (extracting merchant surplus through non-commission channels). This multi-sided revenue approach reduces reliance on any single mechanism and aligns incentives across participants.
**For commission incidence distribution**, the theoretical and empirical analysis predicts:
**Short-run (2% commission rate)**: Merchants absorb 70-90% of commission burden through reduced margins; consumers bear 5-15% through modest price increases on some products; displaced competitors (Google, Amazon) absorb 5-25% through lost market share and reduced pricing power. Minimal deadweight loss due to low rate and high elasticities.
**Long-run (4-6% commission rate, if achieved)**: Merchants absorb 40-60% of commission burden; consumers bear 20-40% through noticeable price increases; displaced competitors absorb 10-30% through sustained market share loss. Moderate deadweight loss as price increases reduce transaction volumes; quality adjustments as merchants optimize for commissions rather than pure fit.
**The distribution depends critically on elasticities that remain uncertain**: Consumer price elasticity in LLM commerce after trust develops; merchant supply elasticity to ChatGPT vs. alternative platforms; cross-platform elasticity (multi-homing ease); competitive response elasticity from Google and Amazon.
**Displaced competitors face structural incidence through several mechanisms**: Google’s loss of product discovery queries reduces advertising inventory value; Amazon’s loss of product search mind share weakens marketplace network effects; traditional affiliate networks face disintermediation as LLM platforms offer superior discovery; comparison shopping engines lose relevance as AI provides synthesis rather than requiring user comparison.
The magnitude of competitive displacement incidence depends on whether LLM commerce proves complementary (expanding total e-commerce) or substitutional (redistributing existing commerce). Current evidence suggests hybrid: LLM platforms expand discovery and consideration stages (complementary) while potentially capturing transaction completion (substitutional).
**Policy implications for tax incidence analysis**: Regulators concerned about commission burden on small merchants should focus on data access policies and interoperability requirements rather than rate caps. Empirical evidence shows fee caps often backfire through platform adjustments on other dimensions. Ensuring merchants can multi-home across platforms maintains competitive constraint on commission rates more effectively than direct regulation.
Transparency requirements around algorithmic ranking and recommendation logic help prevent implicit burden increases through biased recommendations. If platforms must disclose how commercial factors influence suggestions, merchants can evaluate true costs of platform participation beyond nominal commission rates.
Dynamic efficiency considerations matter: short-run burden on merchants from modest commissions may be offset by long-run consumer welfare gains from improved discovery and reduced search costs. Static incidence analysis that ignores quality improvements and search cost reductions will overestimate net burden.
## Emerging market structure remains contested and shapeable
LLM-mediated commerce is fundamentally reshaping digital commerce economics, but the ultimate market structure, pricing dynamics, and incidence distribution remain uncertain and policy-shapeable.
**The theoretical frameworks from two-sided platform economics, tax incidence theory, and industrial organization provide robust analytical tools**, but their application requires careful attention to the specific characteristics of LLM commerce: high initial accuracy uncertainty affecting trust and adoption; bilateral learning creating switching costs; recommendation synthesis creating information asymmetries; network effects from user interaction data; rapid technological evolution enabling competitive disruption.
**OpenAI’s approximately 2% commission represents strategically sound market entry pricing** that balances merchant adoption incentives, competitive positioning, regulatory risk mitigation, and consumer welfare effects. The incidence at this rate falls primarily on merchants through reduced margins, with minimal consumer price impacts and modest competitive displacement of Google and Amazon’s market power.
**As the market matures, commission rates will adjust based on competitive dynamics, regulatory constraints, and merchant bargaining power**. If OpenAI achieves sustained competitive advantages through data accumulation and user lock-in, rates could increase to 4-6% range. If intense competition from Google, Perplexity, and other platforms persists, rates may remain constrained at 2% or below.
**The critical policy levers are data access frameworks and interoperability standards** rather than direct commission rate regulation. Ensuring merchants and consumers can multi-home across platforms maintains competitive constraint on pricing. Algorithmic transparency prevents implicit burden increases through recommendation bias.
**Tax incidence theory predicts that burden distribution depends on relative elasticities**, which remain uncertain given LLM commerce’s novelty. As consumer trust develops and LLM shopping becomes habitual, demand may become less elastic, enabling higher rates with greater pass-through to consumers. As merchants optimize specifically for ChatGPT presence (GEO, product feed optimization), supply becomes less elastic, enabling platforms to extract more surplus.
**The displacement of Google and Amazon represents a form of incidence through competitive erosion** that may ultimately exceed the direct burden of commissions on merchants and consumers. If LLM platforms capture product discovery, the value of search advertising and marketplace prominence declines, reducing incumbent pricing power. This mechanism redistributes surplus from established intermediaries to new LLM platforms without necessarily increasing total intermediation costs.
The race to establish standards, accumulate user trust, and achieve network scale will determine long-run market structure. OpenAI’s aggressive pricing, open protocol strategy, and prohibition on paid placement suggest an attempt to become the dominant trusted platform before competitors establish positions. Whether this succeeds depends on execution, competitive response, and regulatory evolution. The economic analysis indicates that modest commissions on high-quality recommendations could create substantial platform value while maintaining reasonable burden distribution across participants—but only if competitive constraints and algorithmic transparency prevent extraction of excessive surplus.
TC again: But then again, I guess most of you could do better than that!
What should I ask Cass Sunstein?
Yes, I will be doing a Conversation with him soon. Most of all (but not exclusively) about his three recent books Liberalism: In Defense of Freedom, Manipulation: What It Is, Why It Is Bad, What To Do About It, and Imperfect Oracle: What AI Can and Cannot Do.
So what should I ask him? Here is my previous CWT with Cass.
Tariffs do not in general help trade deficits
That is the topic of my latest Bloomberg column, here is one bit:
The most important factors behind trade balances include savings decisions, fiscal policy, economic growth rates, wealth levels and demographic characteristics such as the age of the population. As economist Joseph Gagnon bluntly put it: “None of the studies found any role for trade barriers.”
Currency manipulation can be an important factor, and that has been a problem in the past with China. But it is not a problem right now; if anything, China is propping up the value of its currency. Nor is currency manipulation a problem with Canada, Mexico or the European Union, other targets of Trump’s tariffs.
Insofar as currencies do matter, currency appreciation is one very direct mechanism that limits the potential for tariffs to improve the trade balance. If a country slaps tariffs on imports, that does make those imports more expensive and thus lowers the demand for them. But then the value of the domestic currency will rise, which in turn makes it harder for domestic exporters. There is no guarantee that these effects will cancel each other out exactly, but it is difficult to get much of a trade balance boost through this mechanism, given these offsetting effects.
Standard stuff people, standard stuff. A bunch of you should know better.
AI Personality Extraction from Faces: Labor Market Implications
Human capital—encompassing cognitive skills and personality traits—is critical for labor market success, yet the personality component remains difficult to measure at scale. Leveraging advances in artificial intelligence and comprehensive LinkedIn microdata, we extract the Big 5 personality traits from facial images of 96,000 MBA graduates, and demonstrate that this novel “Photo Big 5” predicts school rank, compensation, job seniority, industry choice, job transitions, and career advancement. Using administrative records from top-tier MBA programs, we find that the Photo Big 5 exhibits only modest correlations with cognitive measures like GPA and standardized test scores, yet offers comparable incremental predictive power for labor outcomes. Unlike traditional survey-based personality measures, the Photo Big 5 is readily accessible and potentially less susceptible to manipulation, making it suitable for wide adoption in academic research and hiring processes. However, its use in labor market screening raises ethical concerns regarding statistical discrimination and individual autonomy.
That is from a new paper by Marius Guenzel, Shimon Kogan, Marina Niessner, and Kelly Shue. I read through the paper and was impressed. Of course since this is machine learning, I can’t tell you what the five traits are in any simple descriptive sense. But this is somewhat of a comeback for physiognomy, which even DeepSeek tells me is a pseudoscience. Via tekl, a fine-looking fellow if there ever was one.
It’s later than you think
Here is a short essay by Hollis Robbins on AI and education, excerpt:
Every faculty member should begin to write a detailed memo specifying the following: “What specific knowledge do I possess that AGI does not? What unique insights or capabilities can I offer that exceed AGI systems? Which students, and in which topics, would benefit enough to pay to learn from me and why?” Faculty who cannot produce this memo with concrete, defensible answers have no place in the institution. There is no middle ground.
Every dean must immediately audit their course catalog against one criterion: what advanced knowledge or skills does this course offer that AGI cannot replicate? Each course must demonstrate specific knowledge transfer or skill development that exceeds AGI capabilities. It will become obvious that the highest value courses are those aligned with specific faculty expertise. General education courses focused on basic knowledge transfer become indefensible. If the information is general enough to be called “general education,” AGI can deliver it more effectively than any human instructor. This will eliminate most of the current curriculum.
Universities will retain faculty in three categories: those advancing original research beyond AGI capabilities, those who teach the use of advanced equipment and sophisticated physical skills, and those handling previously undiscovered source materials or developing novel interpretations that outstrip AGI’s analysis. In the sciences, this means laboratory-based faculty who validate AGI-generated research proposals and offer advanced hands-on training with advanced equipment. In engineering and the arts, it’s faculty who guide students in high-level physical manipulation, augmented by AI tools. In the humanities, it’s scholars working with newly discovered primary sources, untranslated manuscripts, or archaeological evidence not yet processed by AI, as well as those creating fundamentally new interpretive frameworks that transcend AGI’s pattern-recognition capacities.
The curriculum narrows dramatically. Most lecture courses disappear. What remains are advanced research seminars where faculty share findings from new source materials or original experiments, intensive laboratory and studio sessions for hands-on skills, and research validation practicums where students learn to test AGI hypotheses. This represents a 60-70% reduction in current faculty positions, with remaining roles requiring fundamentally different capabilities than traditional academic work.
There is more of interest at the link.
Prediction Markets for the Win
The prediction markets predicted the election outcome more accurately and more quickly than polls or other forecasting methods, just as expected from decades of research. In this election, however, many people discounted the prediction markets because of large trades on Polymarket. Paul Krugman, for example, wrote:
Never mind the prediction markets, which are thin and easily manipulated.
None of that was true but perhaps that was par for the course. Even some prediction market experts, however, began to wobble under the influence of “whale” manipulation theories. But this story was always shaky. What was the supposed logic?
Few directly articulated the theory—perhaps because it sounds absurd when spelled out. The idea seems to be that whales shifted market odds from 50:50 to 40:60, hoping this would drive more people to vote for Trump. Really? Were voters in Pennsylvania watching Polymarket to decide who to vote for? In a decision market, manipulation might be desirable to a whale (albeit unlikely to succeed), but in prediction markets, this scenario seems dubious: a) people would need to know about these markets, b) they’d need to care about probability shifts on these markets (as opposed to voting say the way their family and neighbors were voting), and c) this would have to be an effective way to spend money to influence votes compared to the myriad other ways of influencing voting. Each step seems dubious.
Alternatively, maybe whales were simply wasting money, “memeing” away millions of dollars? Is that something that whales do? The memeing theory is more plausible with many small traders, not a few whales. Or maybe the whales aimed to spark excitement among the minnows, hoping to build momentum before cashing out. However, exciting small traders to inflate prices and then exiting is risky; the same power that whales have to drive up prices can drive prices down just as quickly, making a profitable exit challenging. In short, while not impossible, the idea of whale-driven manipulation in prediction markets was far-fetched.
In fact, we now know that the biggest whale was moving the markets towards accuracy (against his own interest by the way). In an excellent WSJ article we learn:
The mystery trader known as the “Trump whale” is set to reap almost $50 million in profit after running the table on a series of bold bets tied to the presidential election.
Not only did he see Donald Trump winning the presidency, he wagered that Trump would win the popular vote—an outcome that many political observers saw as unlikely. “Théo,” as the trader called himself, also bet that Trump would win the “blue wall” swing states of Pennsylvania, Michigan and Wisconsin.
Now, Théo is set for a huge payday. He made his wagers on Polymarket, a crypto-based betting platform, using four anonymous accounts. Although he has declined to share his identity, he has been communicating with a Wall Street Journal reporter since an article on Oct. 18 drew attention to his bets.
In dozens of emails, Théo said his wager was essentially a bet against the accuracy of polling data. Describing himself as a wealthy Frenchman who had previously worked as a trader for several banks, he told the Journal that he began applying his mathematical know-how to analyze U.S. polls over the summer.
Here’s the most remarkable bit. Theo commissioned his own polls using a different methodology!
Polls failed to account for the “shy Trump voter effect,” Théo said. Either Trump backers were reluctant to tell pollsters that they supported the former president, or they didn’t want to participate in polls, Théo wrote.
To solve this problem, Théo argued that pollsters should use what are known as neighbor polls that ask respondents which candidates they expect their neighbors to support. The idea is that people might not want to reveal their own preferences, but will indirectly reveal them when asked to guess who their neighbors plan to vote for.
…In an email, he told the Journal that he had commissioned his own surveys to measure the neighbor effect, using a major pollster whom he declined to name. The results, he wrote, “were mind blowing to the favor of Trump!”
Théo declined to share those surveys, saying his agreement with the pollster required him to keep the results private. But he argued that U.S. pollsters should use the neighbor method in future surveys to avoid another embarrassing miss.
Thus, a big win for prediction markets, for Polymarket and for GMU’s Robin Hanson, the father of prediction markets, whose work directly influenced the creation of Polymarket.
Saturday assorted links
2. Alexa will be powered by Claude. And ChatGPT goes to church?
3. In praise of reading reference books.
4. AI analyzes interviewing styles, in this case with CWT with Nate Silver.
5. Five productive years from Stephen Wolfram.
6. The microfoundations of Long Covid.
7. New documentary on home schooling.
8. The new Meta glasses might be called “Puffin.” (The Information)
9. But what do the AIs think of you? (NYT)
Trump and the Fed
That is the topic of my latest Bloomberg column, here is one bit:
Trump advisers have been drafting plans to limit significantly the operating autonomy of the Fed. The Trump campaign has disavowed these plans, but the general ideas have been spreading in Republican circles, as evidenced by the Heritage Foundation’s Project 2025 report. Trump himself has called for a weaker dollar policy, which could not be carried out without some degree of Fed cooperation. As a former businessman and real-estate developer, Trump seems to care most about interest rates, banking and currencies.
One concrete proposal reported in the Wall Street Journal would require the Fed to informally consult with the president on decisions concerning interest rates and other major aspects of monetary policy. That would make it harder for the central bank to commit to a stated policy of disinflation, since the ongoing influence of the president would be a wild card in the decision. Presidents would likely give more consideration to their own reelection prospects than to the advice of the Fed staff. Further confusion would result from the reality that the responsibility of the president in these matters simply would not be clear.
It’s important not to be naïve: Regardless of who is in the White House, the Fed already cares what the president and Congress think, as its future independence is never guaranteed. Still, explicit consultation would undercut the coherence of the decision-making process within the Fed itself and send a negative signal to investors. There is no upside from this approach.
There is much more at the link.
“The Simple Macroeconomics of AI”
That is the new Daron Acemoglu paper, and he is skeptical about its overall economic effects. Here is part of the abstract:
Using existing estimates on exposure to AI and productivity improvements at the task level, these macroeconomic effects appear nontrivial but modest—no more than a 0.71% increase in total factor productivity over 10 years. The paper then argues that even these estimates could be exaggerated, because early evidence is from easy-to-learn tasks, whereas some of the future effects will come from hard-to-learn tasks, where there are many context-dependent factors affecting decision-making and no objective outcome measures from which to learn successful performance. Consequently, predicted TFP gains over the next 10 years are even more modest and are predicted to be less than 0.55%.
Note he is not suggesting TFP (total factor productivity, a measure of innovation) will go up by 0.71 percentage points (a plausible estimate, in my view), he is saying it will go up 0.71% over a ten year period, or by 0.07 annually. Here is the explanation of method:
I show that when AI’s microeconomic effects are driven by cost savings (equivalently, productivity improvements) at the task level—due to either automation or task complementarities—its macroeconomic consequences will be given by a version of Hulten’s theorem: GDP and aggregate productivity gains can be estimated by what fraction of tasks are impacted and average task-level cost savings. This equation disciplines any GDP and productivity effects from AI. Despite its simplicity, applying this equation is far from trivial, because there is huge uncertainty about which tasks will be automated or complemented, and what the cost savings will be.
Mostly I think this piece is wrong, and I think it is wrong for reasons of economics. It is not that I think the estimate is off, I think the method is misleading altogether.
As with international trade, a lot of the benefits of AI will come from getting rid of the least productive firms from within the distribution. This factor is never considered.
And as with international trade, a lot of the benefits of AI will come from “new goods,” Since the prices of those new goods previously were infinity (do note the degree of substability matters), those gains can be much higher than what we get from incremental productivity improvements. The very popular Character.ai is already one such new good, not to mention I and many others enjoy playing around with LLMs just about every day.
By the way, the core model of this paper — see pp.6-7 — postulates only a single good for the economy. Mention of the contrary case does surface on p.11, and starting with p.19, where most of the attention is devoted to bad new goods, such as more effective manipulation of consumers. Note the paper doesn’t have any empirical argument as to why most new AI goods might be bad for social welfare.
pp.34-35 focus on the possibility of a public goods problem for AI use, similar to what has been suggested for social media. That discussion seems very far from both current practices with AI and most of the speculation from AI experts. Do I have to use Midjourney because all of my friends do, and I wish the whole thing didn’t exist? Or rather do I simply find it to be great fun, as do many people when they create their own songs with AI? It is dubious to play up the prisoner’s dilemma effects so much, but Acemoglu returns to this point with much force in the conclusion.
Toward the end he writes:
Productivity improvements from new tasks are not incorporated into my estimates. This is for three reasons. First and most parochially, this is much harder to measure and is not included in the types of exposure considered in Eloundou et al. (2023) and Svanberg et al. (2024). Second, and more importantly, I believe it is right not to include these in the likely macroeconomic effects, because these are not the areas receiving attention from the industry at the moment, as also argued in Acemoglu (2021), Acemoglu and Restrepo (2020b) and Acemoglu and Johnson (2023). Rather, areas of priority for the tech industry appear to be around automation and online monetization, such as through search or social media digital ads. Third, and relatedly, more beneficial outcomes may require new institutions, policies and regulations, as also suggested in Acemoglu and Johnson (2023) and Acemoglu et al. (2023).
While many of the points in that paragraph seem outright wrong to me (such as the industry attention point), what he can’t bring himself to say is that the gains from such new tasks will in fact be small. Because they won’t be. But whether or not you agree, what is going on in the paper is that the gains from AI measure as small because it is assumed AI will not be doing new things. I just don’t see why it is worth doing such an exercise.
A more general question is whether this model can predict that TFP moves around as much as it does. I am pretty sure the answer there is “no,” not anywhere close to that.
On the general approach, I found this sentence (p.4) very odd: “…my framework also clarifies that what is relevant for consumer welfare is TFP, rather than GDP, since the additional investment comes out of consumption.” I would say what is relevant for consumer welfare is the sum of consumer and producer surpluses, of which TFP is not a sufficient statistic. This unusual “redefinition of all welfare economics in a single sentence” perhaps follows from how many other gains from trade he has abolished from the system? And footnote six is odd and also wrong: “For example, if AI models continue to increase their energy requirements, this would contribute to measured GDP, but would not be a beneficial change for welfare.” Even for dirty energy that might be wrong, not to mention for green energy. If an innovation induces the market to invest more in a service, the costs of that added investment simply do not scuttle the gains altogether. And if Acemoglu wants to argue that weird welfare economics is true in his model, that is a good argument against his model, not a good argument that such gains would not count in the real world, which is what this paper is supposed to be about.
Acemoglu explicitly rules out gains from doing better science, as they may not come within the ten-year time frame. On that one, he is the prisoner of his own assumptions. If many gains come in say years 10-15, I would just say the paper is misleading, even if his words are defensible in the purely literal sense.
That said, just how much does the “no new science” clause rule out? In terms of an economic model, how does “new science” differ from “TFP”? I am not sure, not are we given clear guidance. Is better software engineering “new science”? Maybe so? Won’t we get a lot of that within ten years? Don’t we have some of it already?
In sum, I don’t think this paper at all establishes the “small gains point” it is trying to promote in the abstract.
It is perfectly fair to point out that the optimists have not shown large gains, but in this paper the deck is entirely — and unfairly — stacked in the opposite direction.
For the pointer I thank Gabriel.
Emergent Ventures winners, 32nd cohort
Anson Yu, Waterloo, telemetry devices that can detect compromised hardware devices to protect our electrical grid and other critical infrastructure.
Anshul Kashyap, Berkeley, neurotech and vision, to visit the Netherlands for work and research reasons.
Kieran Lucid, Dublin, Irish videos about YIMBY and aesthetics, at the site Polysee.
Matin Amiri, Antwerp, Afghanistan, and San Francisco (?), building digital clones.
Snowden Todd, USA and Honduras and South Korea, to write a book on South Korean fertility issues.
Anthony Jancso, Accelerate SF, San Francisco, for general career development.
Denisa Lepadatu, Romania and Bremen, trip to Prospera to pursue longevity research.
Jamie Rumbelow and Henry Dashwood, London, British company to ease land rights/permissions.
Anastasia Vorozhtsova, Columbia University, to study Russian education and the Russian state.
Rohan Selva-Radov, Oxford, general career development, and to develop a dating/matching service for young people.
Olga Yakimenko, Vienna, movie-making.
Rucha Benare, Dublin, Pune area, art and biology.
Brooke Bowman, San Francisco, Vibecamp.
Ruxandra Tesloianu, Cambridge/Romania, travel grant and career development, bio space, science, and meta-science.
Ukraine cohort:
Serhii Shadrin, to study at University of Chicago, and to study information manipulation and media.
Le Sallay Academy, school for Ukrainian refugees, including in France and Serbia, Sergey Kuznetsov and Aleka Molokova.
Here are previous winners of Emergent Ventures. Here is Nabeel’s software for querying about EV winners.
A congestion theory of unemployment fluctuations
Yusuf Mercan, Benjamin Schoefer, and Petr Sedláček, newly published in American Economic Journal: Macroeconomics. I best liked this excerpt from p.2, noting that “DMP” refers to the Nobel-winning Diamond-Mortensen-Pissarides search model of unemployment:
This congestion mechanism improves the business cycle performance of the DMP model considerably. It raises the volatility of labor market tightness tenfold, to empirically realistic levels. It produces a realistic Beveridge curve despite countercyclical separations. On its own, it accounts for around 30–40 percent of US unemployment fluctuations and much of its persistence. In addition, the model accounts for a range of other business cycle patterns linked to unemployment: the excess procyclicality of wages of newly hired workers compared to average wages, the countercyclical labor wedge, large countercyclical earnings losses from displacement and from labor market entry, and the long-run insensitivity of unemployment to policies such as unemployment insurance.
And by their congestion mechanism the authors mean this:
…a constant returns to scale aggregate production function that exhibits diminishing returns to new hires, a feature we call congestion in hiring.
I find that assumption plausible. It remains the case that the DMP model is grossly underrepresented in on-line writings on economics, on Twitter, and in the blogosphere. It won three Nobel Prizes, yet it also suggests that the “simple” manipulation of spending or nominal values does not automatically restore higher levels of employment.
Here are less gated versions of the paper.