Daily Summary for 2026-11-19
daily
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: It's the nineteenth of November, twenty twenty-six, and this is the day's research.
Warren: 8 new papers came out today.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We'll take the day in one pass, then pull out the papers we're staying with.
Warren: Start us off with AI Adoption in S&P 500 Firms.
The day's research: Ruth: Understanding AI adoption in large companies is key to seeing productivity gains and labor market shifts.
Warren: The research focuses on studying AI adoption in S&P 500 firms over several years.
Iris: It shows a clear increase in how deeply these tools are embedded in business processes.
Ruth: Generative AI is popular for personal use, but the main story is deep integration into core operations.
Warren: To measure this deep adoption and separate it from just hype, a new method using SEC filings was developed.
Iris: That study found that by 2025, eleven percent of large firms had AI deeply integrated into their processes.
Ruth: Furthermore, another ten percent of these firms were using AI for making goods or delivering services.
Warren: The paper titled AI Adoption in S&P 500 Firms addresses the potential source of aggregate productivity improvement and labor market impact.
Iris: We also have Difference-in-differences for mediation analysis using double machine learning which aims to identify direct and indirect effects.
Ruth: Leniency Designs: An Operator's Manual develops a step-by-step guide to leniency designs using an unbiased jackknife instrumental variables estimator.
Warren: The Economics of AI Foundation Models explores how developers' choices about model openness affect competition and deployment strategies over time.
Iris: Simple Communication studies simple communication in a multidimensional cheap talk model to characterize equilibrium scores an expert uses.
Ruth: Just After Minimum Wage Hikes looks at how labor markets adjust immediately after minimum wage hikes in Japan's spot labor market.
Warren: Ranking Statistical Experiments via the Linear Convex Order and the Lorenz Zonoid introduces a novel ranking of statistical experiments for comparison across decision problems.
Iris: Irrational Random Utility Models can be generated by a population of decision makers who do not maximize any preference if their preferences are sufficiently uncorrelated.
Lucky paper: 2607.08920: Ruth: So we're diving into this paper now, AI Adoption in S andP five hundred Firms. We saw that eleven percent of these large firms had AI deeply integrated by two thousand twenty-five and another ten percent were using it for making goods or delivering services.
Warren: That measure is based on SEC ten-K filings, which are supposed to prohibit companies from making materially false or misleading statements, so how robust is that rubric?
Iris: The authors developed a rubric ranging from one to five points where five means AI is a core component of the firm’s strategy and deeply embedded across business functions and operations.
Ruth: They found that adoption has more than quadrupled from five percent in two thousand twenty-two to five point seven percent in two thousand twenty-five with aggressive adoption happening mostly in the technology sector, which accounts for two-thirds of that advanced integration.
Warren: The paper also looked at the impact on financial outcomes and found that firms using AI in production and having deeply integrated AI report higher net profit margins with a J-curve effect.
Iris: But they found no significant association between AI adoption and productivity when measured by revenue per employee, which is interesting because that’s a common metric for growth.
Ruth: They also checked the connection to capital expenditure, capex, and found no significant correlation there, but there is a statistically significant relationship between technology sector AI adoption and Tobin’s Q ratio.
Warren: So even though productivity isn't directly linked yet, the market valuation of these firms is definitely moving in the direction of deep AI integration.
Iris: And they noted that deeper AI adoption concentrates among larger firms in the technology sector, suggesting those large tech firms are better positioned to adopt AI at scale because they already have more complete data architecture and data pipelines.
Ruth: That makes sense when you think about having the necessary infrastructure ready for massive deployments.
Warren: We also saw that while generative AI models as a stand-alone work tool might increase worker productivity, the paper suggests by itself it might not be adequate in all cases to perform a wide range of tasks.
Iris: The authors conclude that deep AI adoption has grown from one percent in two thousand twenty-two to five point seven percent in two thousand twenty-five but they also pointed out that aggressive adoption among non-technology firms is growing alongside the technology sector's lead.
Lucky paper: 2510.15200: Ruth: Let's take a look at this paper, The Economics of AI Foundation Models: Openness, Competition, and Governance.
Warren: So the core idea here is that the strategic choice of model openness has become a defining issue for the foundation model ecosystem.
Iris: It seems like they are looking at how this choice impacts real economic drivers that aren't always talked about much.
Ruth: Exactly, because they found that the incumbent’s optimal openness level is surprisingly non-monotonic depending on the strength of its data flywheel advantage.
Warren: That means there are three distinct strategic regimes where the incumbent operates differently based on how strong that flywheel effect is.
Iris: The paper constructs a two-period game-theoretic framework involving an incumbent developer, a downstream deployer, and an entrant developer.
Ruth: The deployer's main economic function is to transform that general purpose foundation model into something specialized for end users.
Warren: That fine tuning cost is given by the formula C1 Q1 equals c Q2 over one plus eta one.
Iris: And higher openness, that eta one, actually reduces the deployer’s fine tuning cost because it makes the model architecture more transparent and helps with knowledge sharing.
Ruth: The paper identifies three canonical strategies for an incumbent based on the strength of that data flywheel effect k.
Warren: For a weak data flywheel effect, they adopt a Harvest strategy, which means choosing maximum openness and a high license price to maximize short term profit before ceding the future market.
Iris: That harvest strategy is tied to point S0 in Figure two where the incumbent takes maximum exposure and charges the highest possible price initially.
Ruth: When that data flywheel effect is intermediate, they switch to a Defend strategy, deliberately restricting openness to impair an entrant’s learning and secure a future advantage.
Warren: That corresponds to point S1 in Figure two where they restrict openness strategically for that intermediate advantage.
Iris: Then there's the Dominate strategy for when the data flywheel effect is very strong, allowing the incumbent to set a low price and a correspondingly higher openness threshold to aggressively encourage adoption.
Ruth: The paper also discusses policy paradoxes, specifically what happens when regulators force a developer in that Dominate regime to be fully open.
Warren: That regulatory intervention creates what they term the “openness trap,” where forcing openness removes strategic flexibility and can even cause the incumbent to pivot back to the Harvest strategy by charging a high license price.
Iris: The welfare implications of vertical integration are conditional on that data flywheel effect k; when k is strong, integration boosts first period fine tuning Q1v and strengthens the second period effort Q2v.
Ruth: But if k is weak, that same vertical integration becomes anticompetitive because it forecloses a more efficient entrant who could have entered the market.
Warren: Similarly, government subsidies are vulnerable to strategic capture by the incumbent through price and openness adjustments, often leaving the rest of the value chain worse off than before.
Iris: The conclusion provides an actionable playbook: for an incumbent, recognize that openness and pricing are tools to be co-optimized based on that data flywheel effect.
Ruth: So if you have a weak flywheel, maximize short term revenue; if the flywheel is strong, accelerate lock in with lower prices and greater openness.
Warren: If the race is tight, they suggest strategically restricting openness to defend against fast followers who are trying to catch up.
Lucky paper: 2502.06530: Ruth: Okay, we’re moving on to the paper titled Ranking Statistical Experiments via the Linear Convex Order and the Lorenz Zonoid from Aalto University.
Warren: This paper introduces a novel way to rank statistical experiments called the linear Blackwell order, or LB order.
Iris: It characterizes this order by three things: dispersion of posterior and likelihood ratios in the linear convex sense, which is related to the size of a Lorenz zonoid, or statewise expectation profiles.
Ruth: So, if an experiment F dominates G in this LB order, it means its random vector is more dispersed than G's vector when measured by every univariate convex function C.
Warren: That dispersion characterization is really about how much uncertainty you have when you look at the data under different prior beliefs.
Iris: And they also say that F dominates G if the Lorenz zonoid of F, denoted as Z(F), contains the Lorenz zonoid of G, which is a geometric condition.
Ruth: That geometric condition is equivalent to saying that for any function you pick for G's signal, there's a corresponding function for F with the same conditional expectation in every state.
Warren: It seems like they are trying to build a more comprehensive measure of informativeness than the Blackwell order.
Iris: They also show that this LB order is more tractable when the number of states, which we call n, is at least two because it only needs comparisons involving univariate convex functions.
Ruth: That makes sense for practical use if you're dealing with a lot of states.
Warren: The paper connects the LB order to other concepts proving that F dominates G in LB order if and only if F is mean-preserving spread of G for every function phi, which is the monotone-posterior-expectation order.
Iris: That means they are showing these different ways of measuring how much information an experiment provides are all linked under this ordering framework.
Ruth: What about the economic applications they show? They talk about decision problems and moral hazard problems in particular.
Warren: In decision problems, F dominates G in LB order if and only if the expected payoff under F is higher than under G for every binary-action problem.
Iris: That extends to all decision problems where the payoff is quasi-concave in a given belief, meaning you get a better expected outcome with more information.
Ruth: For moral hazard problems, dominance in LB order corresponds to both a wider range of implementable target actions and a smaller minimal disutility incurred by the principal when implementing any target action.
Warren: So it implies greater flexibility in control and greater effectiveness when dealing with those situations.
Iris: They also establish bounds on penalties and rewards showing that if F dominates G in LB order, there exists a scheme under F implementing the same outcome as G but at lower disutility for the principal.
Ruth: That's a concrete result about reducing costs in those kinds of problems.
Warren: And for screening problems with ex post signals, dominance in LB order means the principal’s expected utility from her optimal mechanism is higher under F than under G for every screening environment E.
Iris: So greater informativeness in the LB sense directly leads to a higher payoff guarantee for the investor in those settings.
Ruth: This paper concludes that F dominates G in this LB order if and only if F dominates G under the Blackwell order, which is interesting because it suggests LB is a broader measure of informativeness.
Warren: They also introduce the monotone-posterior-expectation order which coincides with the Lehmann order under MLRP but stays applicable when MLRP fails.
Iris: And finally, they provide numerical methods for verification that reduce the check down to verifying zonotope inclusion, which simplifies things quite a bit.
Lucky paper: 2403.10208: Ruth: We're moving into paper number four of four today, and it’s titled Irrational Random Utility Models.
Warren: This paper looks at the Random Utility Model, showing it is observationally equivalent to a population of irrational decision-makers if their preferences are sufficiently uncorrelated.
Iris: So this questions the rational interpretation of Random Utility Models because the results suggest we might be looking at something different than what we assume.
Ruth: The authors introduce random non-utility models, or RNUMs, as a contrast to RUMs and characterize them as an RCM with support in C excluding P.
Warren: A key observation they make is that the relationship between rho and mu is analogous to the one between joint and marginal distributions, which is governed by Fréchet bounds.
Iris: They introduce correlation bounds as a variation of these Fréchet bounds to highlight the role of correlation between rational and irrational choice functions.
Ruth: Theorem three point four states that a stochastic choice function is an RNUM if and only if it satisfies the correlation bounds, which is a big result for characterizing these models.
Warren: The paper then characterizes Irrational Random Utility Models, or I-RUMs, as those that are both RUM and RNUM, showing they exist only if the probability distribution on preferences is sufficiently spread out.
Iris: Specifically for a RUM with support on P1 and P2 to be an I-RUM, it must satisfy mu of P1 equals mu of P2 equals zero point five.
Ruth: They then introduce I rho as a measure of irrationality, which equals the maximal fraction of irrational decision-makers in any RCM representation.
Warren: Proposition five point six shows that max mu in delta of P rho equals rho mu C over P, meaning I rho is the maximal fraction of irrational decision-makers in any RCM representation mu of rho.
Iris: That implies if I rho is greater than I prime rho then rho is more irrational than prime because it can be represented by a larger fraction of irrational decision-makers.
Ruth: They discuss the non-falsifiability of the rational interpretation of RUMs, showing that almost all RUMs are compatible with a fraction of irrational individuals.
Warren: Proposition six point one establishes that for a RUM rho having an RCM representation mu with mu C over P greater than zero is equivalent to having an alpha-RNUM representation for some alpha greater than zero.
Iris: This suggests that almost all RUMs are non-falsifiable because they are compatible with a fraction of irrational decision-makers.
Ruth: They conclude by showing that all RUMs can be approximated arbitrarily close by an I-RUM when the set of alternatives is sufficiently large.
Warren: Proposition six point four provides a bound on this distance, stating that for every RUM rho there is an I-RUM rho prime with rho minus rho prime less than or equal to one over two X minus X minus one.
Iris: That bound decreases exponentially as the number of alternatives increases, which is interesting for how robust these approximations are.
Ruth: The analysis reveals a tension between stochastic and individual rationality, showing that some non-rationalizable stochastic choice functions may need more rational choice functions than some rationalizable ones do.
Warren: They illustrate this by an example where a stochastic choice function with concentrated probability mass that is not a RUM implies both a RUM and an RNUM representation do not exist, meaning both rational and irrational decision-makers are present.
Iris: For the Logit Model, they show that a sufficient condition for a Luce stochastic choice function to have an irrational representation is u of A less than or equal to K minus two u of B for all A, B in X.
Ruth: That condition intuitively shows that even the strictest notion of stochastic rationality is susceptible to irrational representations whenever the utilities are not sufficiently far apart.
Warren: The paper also extends these arguments to other aligned influential models such as the additive perturbed utility model and the simple scalability model.
Iris: The authors conclude that economic theory is much more compatible with irrational behavior than had been previously suspected, and that individuals with correlated mistakes can never be translated into rational aggregate behavior.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19