% paper/sections/literature.tex — UQO Working Paper No. 25
% Related literature. Every claim about a cited paper follows docs/literature_matrix.md.

\section{Related literature}
\label{sec:literature}

\paragraph{AI and the macroeconomy.}
The macroeconomics of artificial intelligence has been built almost entirely on the task framework. \citet{zeira1998workers} and \citet{acemoglu2018race} model automation as capital replacing labor in a measure of tasks, with displacement and reinstatement effects on wages and the labor share; \citet{acemoglu2022tasks} show that task displacement accounts for most of the rise in U.S. wage inequality, and \citet{acemoglu2025simple} uses the same accounting to bound the productivity gains from AI and to allow for ``bad'' tasks. \citet{aghion2019artificial} place AI in production and in idea production with Baumol-type bottlenecks, and \citet{jones2024dilemma,jones2025how} weigh growth against catastrophic risk. Two recent papers make compute the binding constraint: \citet{korinek2024scenarios} obtain a wage collapse once compute overtakes the task-complexity frontier, and \citet{restrepo2025we} shows that with abundant compute all bottleneck work is automated and wages converge to the cost of replicating human work. \citet{moll2022uneven} embed automation in a heterogeneous-agent economy and show that it raises returns to wealth, and hence wealth inequality, even when wages rise. \citet{ide2025artificial} come closest to AI as a decision-maker: in a knowledge-hierarchy economy, autonomous and non-autonomous AI reshape organizations and wages, but errors are not correlated across users and there is no model count. \citet{agrawal2019artificial} interpret AI as cheaper prediction complementing human judgment, and \citet{athey2020allocation} characterize the partial-equilibrium allocation of decision rights between a human and an AI with misaligned objectives. On the verbal side, \citet{hadfield2025economy} argue that general-equilibrium distortions from AI consumers and producers are precluded only if markets are large and AI mistakes are zero-mean and independent, and note that copying the same agent across firms induces correlated mistakes that do not wash out; \citet{rothschild2025agentic}, \citet{catalini2026some} and \citet{immorlica2024generative} discuss agentic matching, AGI pricing and research agendas. Measurement of exposure, adoption and productivity gains \citep{eloundou2024gpts,bick2024rapid,brynjolfsson2025generative,eisfeldt2023generative} disciplines our calibration. All of these treat AI as a production input, an information service or a growth factor. We keep the task-based production block of \citet{acemoglu2018race} and the compute constraint of \citet{restrepo2025we}, but add a third role: AI as a delegated decision-maker whose forecast error is shared by every principal using the same foundation model. The premise is the one \citet{hadfield2025economy} state in words; our contribution is to make the delegation share and the number of independent models equilibrium objects and to derive their price, welfare and policy consequences.

\paragraph{Dispersed information, common signals and sentiment-driven fluctuations.}
The information structure of our baseline is classical. \citet{grossman1980impossibility}, \citet{hellwig1980aggregation} and \citet{diamond1981information} establish that prices aggregate dispersed, independently-noisy signals and cannot be fully revealing when information is costly \citep[see][for the full-revelation benchmark and a survey]{radner1979rational,vives2008information}. \citet{morris2002social} show that agents over-weight a common public signal and that more precise public information can lower welfare; \citet{angeletos2007efficient} characterize efficient information use in a general class of economies in which the social value of common information depends on the degree of strategic complementarity. On the business-cycle side, \citet{lorenzoni2009theory} shows that noise in a common public signal generates demand-driven fluctuations, \citet{angeletos2013sentiments} obtain unique-equilibrium fluctuations from exogenously correlated belief errors, and \citet{benhabib2015sentiments} construct self-fulfilling sentiment equilibria; \citet{david2016information} quantify the misallocation cost of firm forecast errors, and \citet{sims2003implications} microfounds noisy responses through limited attention. We state candidly that ``common noise in beliefs is an aggregate shock with real effects'' belongs to this literature. What we add is a market in which the correlation of forecast errors is chosen. The information-choice literature has endogenized the source of information: \citet{veldkamp2006media} shows that fixed costs make many agents buy the same news, \citet{hellwig2009knowing} derive homogenization of information sources from strategic complementarity, \citet{colombo2014information} and \citet{vives2017endogenous} analyze the welfare properties of endogenous information, and \citet{farboodi2020long} make the choice of an information \emph{technology} an equilibrium object that changes what prices reveal \citep[see also][]{farboodi2022where,farboodi2023data,veldkamp2011information}. \citet{jones2020nonrivalry} show that ownership of a non-rival information input shapes its use, and \citet{dugast2018data} that cheap raw data can lower price informativeness without any correlated-error story. In each of these papers the technology is private, or homogenization is driven by complementarity or fixed costs. Our delegation margin instead rests on a private precision gain and the scale economies of foundation models, with no strategic complementarity required: a household or firm that delegates obtains model $m$'s forecast and inherits model $m$'s error, so the error covariance of the economy loads on the delegation share $D$ and on $1/M$. The efficiency analysis then connects directly to \citet{angeletos2007efficient}: in the constant-returns, price-taking benchmark the private and social marginal values of delegation coincide exactly for every $D$, $M$ and degree of strategic substitutability (Proposition~\ref{prop:benchmark}), so aggregate ``model noise'' is real but not an inefficiency, and a wedge opens only under identified frictions---non-pecuniary aggregate effects, uninsurable price risk for non-participants, prices in collateral constraints, or finite market depth (Proposition~\ref{prop:wedges}).

\paragraph{Algorithmic trading, homogenization and AI in finance.}
That algorithmic behavior is more correlated than human behavior is documented: \citet{chaboud2014rise} find that algorithmic strategies in foreign-exchange markets are more correlated than human strategies yet improve efficiency, \citet{khandani2011what} show that similar, crowded quantitative strategies unwound together in August 2007, and \citet{kirilenko2017flash} document the synchronized withdrawal of high-frequency liquidity during the Flash Crash. The theory of crowding is also established: \citet{stein2009presidential} analyzes crowded-trade and leverage externalities in which each arbitrageur ignores the effect of his trade on others' unwinding, and \citet{wagner2011systemic} shows that diversification by every investor makes portfolios similar, creates joint-liquidation risk and leaves diversity under-provided. Outside finance, \citet{kleinberg2021algorithmic} show that a shared, more accurate algorithm can lower aggregate outcome quality because independence is lost, \citet{bommasani2022picking} document outcome homogenization across systems sharing a base model, \citet{doshi2024generative} find that generative assistance raises individual quality while reducing collective diversity, and \citet{jo2025homogeneous} show that correlated predictors soften product-market competition. \citet{calvano2020artificial} and \citet{fish2024algorithmic} obtain correlated algorithmic behavior through learned collusion rather than shared errors. In asset markets, \citet{dou2025ai} show that reinforcement-learning traders in a Kyle-type setting \citep{kyle1985continuous} sustain supra-competitive outcomes through price-trigger strategies and through homogenized learning biases that lower informativeness and liquidity; \citet{park2026correlated} analyzes correlated AI signals in a noisy rational-expectations market; \citet{qiu2026representation} simulate how representation homogeneity compresses forecast disagreement under stress; and \citet{garleanu2018efficiently} show that delegation to human managers already co-determines market efficiency. \citet{gabaix2021search} provide the transmission channel we require: in inelastic markets, correlated flows move prices. The closest formal antecedent is \citet{meng2026artificial}, who study endogenous adoption of AI trading with a common signal component in a Kyle market, obtain a systemic-risk multiplier convex in the adoption share, and show tipping to monoculture in a supermodular adoption game. We differ on four counts: our adopters are households and firms whose delegation moves wages, the interest rate and capital accumulation, not traders alone; the model count $M$ enters explicitly, so diversity is a comparative static; we derive an efficiency benchmark and the frictions under which a Pigouvian delegation tax or model diversity restores constrained efficiency; and we close the compute market. Relative to \citet{stein2009presidential} and \citet{wagner2011systemic}, our externality operates in forecast space rather than portfolio space and is priced through factor and asset prices rather than liquidation costs. The 2025--26 evidence constrains rather than confirms us. \citet{he2026when} uses ChatGPT outages to show that the tool acts as a belief homogenizer for retail investors, synchronizing trades and worsening liquidity \citep[on the tradability of a single public model's signal, see][]{lopezlira2023can}; \citet{spiro2026oracle} finds that LLM forecast errors are strongly correlated across vendors, which is why we treat $M$ as an \emph{effective} number of independent models to be calibrated rather than a count of providers; and \citet{hansen2025financial} find that LLM agents herd less than humans in observational cascade games, consistent with our channel, which runs through shared technology rather than observation \citep[contrast][]{banerjee1992simple}. Policy institutions have stated the monoculture concern verbally \citep{gensler2020deep,danielsson2022artificial,danielsson2024artificial,fsb2024financial,imf2024advances,bis2024artificial}. We also confront the competing explanation. Belief homogenization reduces disagreement, and in \citet{miller1977risk}, \citet{harrison1978speculative}, \citet{scheinkman2003overconfidence} and \citet{simsek2013belief}, surveyed by \citet{hong2007disagreement}, less disagreement means less overpricing, less speculative volume and less leverage. Our Model C reproduces this in calm periods; the model-noise channel dominates only when the shared error is large relative to the dispersion it removes, and we characterize when each effect wins.

\paragraph{Collateral, leverage and pecuniary externalities.}
The amplification block is standard and we claim nothing new about it. \citet{kiyotaki1997credit} and \citet{bernanke1999financial} establish the feedback between asset prices, net worth and borrowing; \citet{brunnermeier2009market} the funding--market liquidity spiral; \citet{geanakoplos2010leverage} the leverage cycle; \citet{gertler2010financial} the bank balance-sheet channel; and \citet{brunnermeier2014macroeconomic} the non-linear dynamics of continuous-time macro-finance. The welfare logic is that of \citet{greenwald1986externalities}: with incomplete markets, equilibria are generically constrained inefficient because of pecuniary externalities. \citet{lorenzoni2008inefficient} shows over-borrowing when agents ignore the price effects of fire sales, and \citet{davila2018pecuniary} decompose pecuniary externalities into distributive and collateral components with the corresponding Pigouvian corrections. Our collateral wedge is a Dávila--Korinek collateral externality whose triggering shock is model noise: a shared forecast error moves collateral values for every delegator at once, and the private delegation decision ignores the resulting tightening of others' constraints. We note the tension with \citet{geanakoplos2010leverage} and \citet{simsek2013belief}, whose leverage cycles require belief heterogeneity that delegation reduces; this is the calm-period effect above.

\paragraph{Delegation and principal--agent theory.}
We model an AI agent as a constrained policy acting under a mandate, with the principal retaining residual claims and liabilities. This is a contract on observables in the sense of \citet{holmstrom1979moral}, whose informativeness principle governs what the mandate conditions on; the linear risk limits we use follow \citet{holmstrom1987aggregation}, and multi-task mandates covering both trading and risk control distort effort as in \citet{holmstrom1991multitask}. The delegation decision itself is an \citet{aghion1997formal} trade-off between initiative and control, with the threshold structure of \citet{dessein2002authority} and the interval-delegation sets of \citet{alonso2008optimal}. What distinguishes our delegation threshold (Proposition~\ref{prop:threshold}) from these partial-equilibrium results, and from \citet{athey2020allocation}, is the general-equilibrium term attached to it: the mandate's payoff depends on how many other principals have delegated to the same model. The robo-advice evidence supports the private side of the trade-off: \citet{dacunto2019promises} find that adopters diversify more and that under-diversified investors gain, and \citet{rossi2024diversification} estimate welfare gains. Our model predicts that such users' trades become correlated, a testable implication.

\paragraph{Heterogeneous agents, agent-based economics and LLM simulations.}
Model B is a stationary recursive equilibrium in the tradition of \citet{huggett1993riskfree} and \citet{aiyagari1994uninsured}, with aggregate risk handled as in \citet{krusell1998income}; we solve it with the sequence-space methods of \citet{auclert2021using} and relate the continuum-with-common-noise structure to the mean-field formulation of \citet{achdou2022income}. The direct-versus-indirect decomposition of \citet{kaplan2018monetary} organizes our split of delegation effects into a private precision gain and a general-equilibrium model-noise channel. Because \citet{hubmer2021sources} and \citet{moll2022uneven} already explain rising top wealth shares through heterogeneous returns and automation-driven increases in the return to wealth, the ownership result (Proposition~\ref{prop:ownership}) must carry a distinct signature: compute scarcity rents respond to the delegation share while the generic return to capital does not. Model C is a bounded-rationality simulated outcome in the constructive sense of \citet{tesfatsion2006agent}, benchmarked against the stylized facts of agent-based finance \citep{lebaron2006agent,hommes2006heterogeneous} and macro \citep{farmer2009economy,dawid2018agent,axtell2025agent}. \citet{brock1998heterogeneous} already endogenize predictor choice by past fitness; ours is predictor choice with a shared predictor and a price externality. The generative-agent literature supplies the architecture \citep{park2023generative} and the caution of \citet{horton2023large}; \citet{chen2023emergence} and \citet{mei2024turing} show that LLM choices are broadly rational and within the human distribution, \citet{ross2024llm} document systematic, model-specific biases---exactly the model-indexed error we posit---and \citet{bybee2023surveying} shows that LLM expectations track survey expectations. LLM-agent simulations exist: \citet{li2024econagent} reproduce the Phillips curve and Okun's law with LLM households, \citet{zhao2023competeai} study competing LLM firms, \citet{lopezlira2025can} runs LLM traders in an order book and warns that similar architectures may behave in correlated ways, \citet{ross2026why} find that higher LLM participation improves price discovery under a ``better models, riskier systems'' framing, \citet{hammond2025multiagent} list correlated failures from shared models as a multi-agent risk, and \citet{xia2026agentic} flag strategy homogenization as an open problem. None varies the model count or the adoption margin as a treatment tied to a theory. Our laboratory follows the discipline of \citet{gode1993allocative}: constraints, budgets, collateral and accounting are enforced in deterministic code, zero-intelligence and Bayesian controls isolate what the agent contributes, and LLM output is never trusted as accounting truth or read as evidence on human preferences.

\paragraph{Summary of the gap.}
The ingredients of our mechanism are all in the literature: common noise generates fluctuations, common information is over-used, information sources homogenize endogenously, private similarity creates inefficient common exposure, pecuniary externalities justify Pigouvian corrections, and collateral constraints amplify shocks. Four things are not. First, no existing model makes the \emph{correlation structure} of forecast errors---not only their variance---an equilibrium object chosen by households and firms through a delegation decision, with a closed-form loading on $D$ and $1/M$ (Proposition~\ref{prop:linear}). Second, no paper states the delegation externality in a production economy: competitive delegation is constrained efficient in the constant-returns benchmark and becomes excessive only when model noise is loaded on wages, interest rates and asset prices through identified frictions, with a Pigouvian delegation tax and model diversity as equivalent remedies derived with the \citet{davila2018pecuniary} decomposition (Propositions~\ref{prop:benchmark} and~\ref{prop:wedges}). Third, a compute market that clears inference demand from delegation against automation demand from production through a single rental price, generating rents to the owners of algorithmic capital, is a new mechanism with a comparative static that distinguishes it from a generic rise in the return to capital (Propositions~\ref{prop:compute} and~\ref{prop:ownership}). Fourth, our agentic market experiments vary $M$ and $D$ factorially with LLM agents across Walrasian, continuous-double-auction and bilateral designs, and tie each treatment to a prediction of the theory.