Artificial intelligence, artificial stupidity, and market breakdown
Winston Wei Dou, Itay Goldstein, and Yan Ji address whether AI-powered trading algorithms using reinforcement learning (RL) can autonomously learn to collude, sustaining supra-competitive profits without any explicit agreement, communication, or intent, and if so, through what mechanisms and with what consequences for market efficiency and liquidity.
The authors identify two fundamentally different algorithmic pathways to collusion, which they colorfully dub “artificial intelligence” and “artificial stupidity”:
- Price-Trigger Collusion (“Artificial Intelligence”) arises when noise trading risk is low and information-insensitive investor presence is large. Under these conditions, lagged market prices are sufficiently informative that algorithms learn to monitor each other indirectly through prices. They sustain collusion by trading conservatively and reverting to aggressive punishment strategies when price deviations signal a defection. This requires genuine dynamic sophistication from the RL process.
- Over-Pruning Bias Collusion (“Artificial Stupidity”) arises when noise trading risk is high or information-insensitive investor presence is small . In this regime, price-trigger strategies break down. Instead, collusion emerges from a systematic learning bias: when noise traders happen to trade in the same direction as an aggressive algorithm, that algorithm suffers large losses and sharply lowers the Q-value of aggressive strategies, effectively “pruning” them. Because exploitation causes aggressive strategies to be revisited less, the bias is never corrected. All algorithms end up collectively favoring conservative trading — generating supra-competitive profits through a shared bias rather than strategic coordination.
Regardless of mechanism, greater AI collusion leads to:
- Lower market liquidity
- Lower price informativeness
- Higher mispricing
The paper also characterizes who wins and loses: informed AI speculators gain, while information-insensitive investors bear the direct costs, and the broader market suffers from degraded price discovery.
Regulatory Implication
AI collusion falls outside existing antitrust frameworks, which focus on detecting explicit communication or shared intent, neither of which is present here. The paper thus poses a direct challenge to regulators: the anti-competitive outcome is indistinguishable from legal behavior under current law.
AI-Powered Trading, Algorithmic Collusion, and Price Efficiency
Authors: Winston Wei Dou, Itay Goldstein, Yan Ji
From: The Wharton School, Hong Kong University of Science
Smarter lenders but more fragile credit markets
In this paper, Simone Lenzu argues that AI in credit markets has two opposing effects: it makes individual lending decisions smarter, but may make the financial system collectively more fragile. the author develops an analytical framework rooted in the New Keynesian macroeconomic model. The core architecture is a two-equation forward-looking system consisting of a New Keynesian Phillips Curve (NKPC) and a dynamic IS curve, augmented to incorporate AI’s structural effects.
Findings
- Improved credit allocation dampens the financial accelerator.
AI-enhanced credit scoring, risk assessment, and contract enforcement reduce external finance premia. This weakens the traditional financial accelerator mechanism: firms face smaller cyclical spikes in borrowing costs during downturns, which should dampen boom-bust amplification and smooth monetary transmission. - But model monocultures create procyclical credit supply.
The dangerous counterpart is that when lenders deploy similar AI models, their decisions become highly correlated. Algorithmic lenders tend to ease credit standards during expansions and tighten them abruptly in downturns, simultaneously and across institutions. This synchronization can amplify credit booms, accelerate leverage build-up, and trigger sharper contractions than traditional lending would produce. - The stagflation risk at the intersection.
The most alarming scenario Lenzu identifies arises when AI adoption frictions depress realized productivity in the short run (cost-push inflation), while AI-driven credit expansion simultaneously fuels elevated asset valuations (financial fragility). The result is a potential AI-specific stagflation with rising costs and financial stress occurring together. - Implication for monetary policy transmission.
Because AI in credit markets can both reduce and amplify the financial accelerator depending on the phase of the cycle, the mapping from policy rates to financial conditions faced by households and firms becomes state-dependent and harder to predict. Central banks can no longer rely on stable reduced-form relationships between policy rates and credit conditions.
Artificial intelligence and monetary policy: A framework and perspective on cyclical transmission, structural transition, and financial stability
Author: Simone Lenzu
From: Federal Reserve Bank of New York, New York University
