AI traders replicate human bias and trigger market bubbles

Artificial intelligence systems designed for financial trading are now replicating the same cognitive biases long observed in human investors, raising new concerns about the stability of increasingly automated financial markets. A new study reveals that autonomous AI agents not only mirror human trading behavior but can also collectively generate market bubbles and crashes under certain conditions.

Published as Dissecting AI Trading: Behavioral Finance and Market Bubbles on arXiv, the research suggests that AI is not eliminating irrationality from markets but instead encoding and amplifying it in new ways, while also offering a potential mechanism to control it.

AI traders inherit human biases and act on them more aggressively

The study finds that AI agents trained on human-generated data internalize deeply rooted behavioral finance patterns, including the disposition effect and extrapolative expectations. These biases have been widely documented in human investors, but their replication in autonomous AI systems signals a critical shift in how financial markets may evolve.

AI traders consistently sell assets that have gained value while holding onto those that have declined, despite having full access to rational market fundamentals. This behavior reflects a reliance on past purchase prices rather than forward-looking valuation, a hallmark of human cognitive bias.

AI agents demonstrate strong extrapolative tendencies. Instead of anchoring expectations to intrinsic value, they forecast future prices based heavily on recent trends. When prices rise, AI agents systematically expect further increases, effectively reinforcing momentum-driven trading patterns.

The research shows that this extrapolation is not evenly distributed across time. AI agents place disproportionate weight on recent price movements, displaying a form of recency bias that mirrors human decision-making under uncertainty. As forecasting horizons extend, this bias intensifies, with agents projecting short-term trends into long-term expectations, a key driver of speculative bubbles.

Unlike human investors, however, AI agents do not suffer from hesitation, transaction costs, or emotional inertia. The study finds a tight coupling between beliefs and actions, meaning that once an AI agent forms a price expectation, it acts on it immediately and decisively. This frictionless execution amplifies the impact of biased expectations, making AI-driven markets potentially more volatile than their human counterparts.

Individual AI behavior scales into full market bubbles

The study demonstrates how these micro-level biases aggregate into large-scale market phenomena. When multiple AI agents interact within a simulated trading environment, their collective behavior reproduces the classic boom-and-bust cycles observed in human financial markets.

Prices in the simulated markets consistently rise above fundamental value in early trading periods, forming bubbles that eventually collapse. These dynamics closely resemble historical experimental market behavior, confirming that AI systems can generate endogenous instability without external shocks.

A key mechanism driving this instability is the relationship between excess demand and price changes. The study shows that when buying pressure exceeds selling pressure, prices increase in subsequent periods, creating a feedback loop that reinforces upward trends. This adaptive price formation process mirrors well-established economic theories of market dynamics.

Another critical factor is disagreement among AI agents. The research finds that when agents hold divergent expectations about future prices, trading volume increases significantly. This heterogeneity of beliefs fuels speculative activity, as agents trade based on differing interpretations of market signals.

The interaction between extrapolation, disagreement, and rapid execution creates a self-reinforcing cycle. Rising prices lead to more optimistic forecasts, which trigger additional buying, further pushing prices upward. This loop continues until the market reaches unsustainable levels, eventually leading to a correction or crash.

Importantly, the study also reveals that not all AI models behave identically. Some generate severe mispricing and large bubbles, while others remain close to fundamental value. This variation suggests that model architecture and training data play a significant role in determining market outcomes, raising questions about the systemic impact of deploying different AI systems in financial markets.

Programmable AI behavior opens new path for market regulation

While the replication of human bias in AI may raise concerns, the study highlights a critical advantage: AI behavior is programmable. Unlike human investors, whose biases are difficult to eliminate, AI agents can be directly influenced through targeted interventions at the prompt level.

Modifying the instructions given to AI agents can significantly alter market outcomes. When prompts encourage speculative behavior, such as momentum chasing or "riding the bubble," the magnitude of market bubbles increases dramatically. Conversely, when prompts emphasize fundamental analysis and risk awareness, speculative activity declines and prices stabilize.

These findings suggest that AI-driven markets can be actively shaped through what the authors describe as cognitive guardrails. By embedding behavioral constraints into AI systems, regulators and financial institutions may be able to reduce volatility and prevent extreme mispricing before it occurs.

This approach represents a shift from traditional regulatory tools, which often rely on reactive measures such as circuit breakers or capital requirements. Instead, prompt-level interventions offer a proactive mechanism to influence market behavior at its source, the decision-making process of AI agents.

The study also acknowledges limitations. The simulated markets do not include human participants, and real-world environments involve more complex institutional structures. Additionally, rapidly evolving AI models may exhibit different behavioral patterns over time, making continuous monitoring essential.