Do LLM-Powered Investors Herd Less Than Humans?
In this paper, Anne Lundgaard Hansen and Seung Jung Lee replicate a classic human laboratory experiment on herd behavior, replacing 32 London financial market professionals with LLM-based AI agents. The setup is a sequential trading game around a single risky asset with a fundamental value of 0, 50, or 100, where each agent receives a private signal (70% accurate) and observes the trading history before deciding to buy, sell, or not trade. The Bayesian market maker updates the price after each trade.
Four LLMs are used and results averaged: Claude 3.5 Sonnet, Claude 3.7 Sonnet (extended thinking mode), Meta’s Llama 3 70B, and Amazon Nova Pro. The experiment runs across three treatments that vary the degree of event uncertainty and whether the price updates, allowing the authors to control when herding is theoretically optimal versus suboptimal. Extensions test variations in signal color coding, agent personas, payoff structures, and experiment length.
Findings
- AI agents are substantially more rational than humans. Across treatments, AI agents made rational decisions (following private signals) in 61–97% of cases, compared to 46–51% for human professionals. AI agents made zero erroneous decisions; humans made errors in 3.4% of cases.
- AI agents almost never herd. Information cascades occurred in only 0–9% of AI decisions versus 20% for humans. When AI agents did cascade, they exhibited contrarian behavior (trading against the market trend), not herding because they relied too heavily on their private signal while ignoring the cumulated information embedded in trading history.
- Incentive framing switches AI behavior. When explicitly prompted to maximize profits, agents herded in 47% of cases where herding was theoretically optimal, compared to 0% in the baseline. This demonstrates that LLM behavior is highly sensitive to objective function framing, not just architecture.
- Persona assignment had no effect. Equipping agents with human demographic profiles (age, gender, job title) did not change their trading behavior, contrary to expectations from prior LLM literature.
Conclusion
LLMs may behave irrationally when faced with novel or unconventional signal structures, which is precisely what financial stress events tend to produce.
Financial Stability Implications of Generative AI: Taming the Animal Spirits
Authors: Anne Lundgaard Hansen, Seung Jung Lee
From: Board of Governors of the Federal Reserve System, Federal Reserve Bank of Richmond
How AI Architecture Shapes Financial Stability
In this paper, Kartik Anand, Sophia Kazinnik, Agnese Leonello, and Ettore Panetti ask: does AI destabilize financial markets, and if so, how much depends on the type of AI rather than the economic environment? They place AI agents inside a canonical mutual fund redemption game i.e. a coordination problem where investors choose to redeem or stay, with payoffs depending on both economic fundamentals and what other investors do. This is a stylized model of financial fragility that applies equally to bank runs, currency attacks, and stablecoin panics.
The Anand et al. test two AI architectures head-to-head under the same economic environment with 30 active investors:
- Q-learning (QL) agents: reinforcement learning algorithms that learn through trial and error across 500,000 training episodes, with a decaying exploration rate. They have no prior knowledge of payoff structures and update Q-values based on realized (not expected) rewards.
- LLM agents: implemented using DeepSeek R1-0528, which produces full chain-of-thought reasoning. Each investor is a separate API call.
Four experimental treatments vary: (i) presence of default risk (payoff uncertainty); (ii) presence of fundamental uncertainty (private noisy signals vs. common knowledge); and (iii) asset illiquidity levels.
Findings
- Without uncertainty, both AI types behave rationally. In the dominance regions (fundamentals clearly weak or strong), both QL and LLM investors coordinate on the correct equilibrium. In the intermediate “multiple equilibria” zone, however, they diverge sharply: QL agents collapse multiplicity by converging on the risk-dominant (all-redeem) outcome; LLM agents display belief heterogeneity and produce partial, unpredictable redemptions.
- Default risk has opposite effects on the two architectures. For LLMs, introducing default risk has no material impact because they reason in expected-value terms and treat default as just another probability, consistent with theory. For QL agents, default risk creates a “hot stove effect”: occasional zero payoffs from staying, even when fundamentals are strong, persistently depress the Q-value of staying, biasing agents toward excessive redemption far beyond the no-redemption threshold.
- Private signals rescue LLM coordination. Under fundamental uncertainty (noisy private signals), LLMs converge on the global games equilibrium as they infer that others follow threshold strategies and coordinate accordingly. The chain-of-thought analysis shows all LLM investors converge on a single causal logic: “redeem if my signal falls below a threshold.” This is the information-structure effect that partially aligns with Hansen & Lee’s finding that private signals (vs. social herding) anchor rational behavior.
- Fragility and illiquidity. For LLMs, financial fragility (panic-driven redemptions) increases monotonically with asset illiquidity, closely tracking theory. For QL agents, the relationship is more complex and amplified by default risk because QL investors become even more redemption-prone as illiquidity rises.
Ex Machina: financial stability in the age of artificial intelligence
Authors: Kartik Anand, Sophia Kazinnik, Agnese Leonello, Ettore Panetti
From: ECB
