TLDR: A new study reveals that Large Language Models (LLMs) used in finance are not neutral decision-makers but possess inherent biases, often preferring large-cap stocks and contrarian investment strategies. These latent preferences can harden into confirmation bias, causing LLMs to cling to initial judgments even when presented with contradictory evidence, potentially leading to unreliable investment recommendations that reflect the AI’s view rather than the user’s intent.
Large Language Models (LLMs) are rapidly transforming the financial sector, assisting with everything from forecasting stock prices to optimizing investment portfolios. However, a critical and often overlooked challenge arises when these powerful AIs encounter conflicting information: their own inherent biases can lead to unreliable and unpredictable investment recommendations, potentially overriding the user’s true intentions.
The Hidden Conflict: Your AI, Not Your View
The core issue, as highlighted in a recent research paper titled “Your AI, Not Your View: The Bias of LLMs in Investment Analysis” by Hoyoung Lee and colleagues, is what’s known as “knowledge conflict.” This occurs when an LLM’s pre-trained knowledge clashes with real-time market data. Even more concerning is the tendency for LLMs to exhibit “confirmation bias”—stubbornly adhering to information that confirms their existing beliefs while disregarding contradictory evidence. This means an AI designed to help you invest in, say, the energy sector, might instead push towards technology stocks if that aligns with its ingrained preferences.
Uncovering AI’s Investment Preferences
To systematically investigate these hidden biases, the researchers developed a three-stage experimental framework. First, they generated balanced sets of “buy” and “sell” arguments for 427 prominent S&P 500 stocks, ensuring neutrality by using a separate LLM (Gemini-2.5-Pro) for this task. Next, they presented these balanced arguments to various LLMs, creating a state of informational equilibrium. The models’ subsequent decisions revealed their intrinsic, latent preferences. Finally, they introduced progressively stronger counter-evidence to see how resilient these preferences were, effectively measuring the extent of confirmation bias.
What Biases Do LLMs Hold?
The study uncovered several distinct, model-specific tendencies:
- Sector Preference: The strength and range of sector preferences varied significantly among different LLMs. Some models, like Llama4-Scout and DeepSeek-V3, showed strong and consistent preferences for certain sectors, while others, such as GPT-4.1 and Mistral-24B, exhibited much flatter preferences across sectors. This suggests that the specific AI model used plays a crucial role in its inherent sector leanings.
- Size Preference: A consistent trend emerged: most LLMs showed a clear preference for large-capitalization companies. This “popularity effect” is likely due to the greater volume and richness of data available for well-known corporations in the models’ training datasets. This inclination could lead to smaller companies being systematically overlooked, regardless of their actual merit.
- Momentum Preference: When it came to investment styles, the analysis revealed a surprising and consistent preference across nearly all evaluated models for “contrarian” strategies (investing in underperforming assets expecting a rebound) over “momentum” strategies (favoring assets with recent strong performance).
From Preference to Stubborn Bias
The research clearly demonstrated that these latent preferences can harden into significant confirmation bias. While LLMs were generally receptive to reversing their decisions when presented with only counter-evidence, their flexibility sharply decreased when supporting and counter-evidence were mixed. Even when the amount of counter-evidence was greater, models struggled to reverse their initial judgments. This rigidity was particularly pronounced in models that initially exhibited stronger preferences, highlighting a direct link between the intensity of a latent preference and the stubbornness of the resulting bias.
An interesting finding from the “decision uncertainty” analysis (using a metric called entropy) showed that models with strong inherent biases (like DeepSeek-V3) were very confident in their decisions when presented with balanced information. However, when faced with conflicting, imbalanced evidence, their confidence plummeted, indicating cognitive dissonance. Conversely, models with weaker initial biases (like GPT-4.1) were initially uncertain with balanced prompts but became more confident when presented with a clear majority of evidence, aligning their decisions accordingly.
Also Read:
- Dialect-Linked Biases in AI: How Subtle Data Poisoning Amplifies Harmful Stereotypes in Language Models
- Navigating the AI Frontier: Large Language Models in Social Simulation
Implications for Financial AI
These findings have profound implications for the financial industry. If LLM-based financial services are guided by the opaque and arbitrary preferences of the underlying AI model rather than the user’s intended, evidence-based investment views, the reliability of these services is fundamentally compromised. This means that if your investment goals differ from the AI’s inherent biases, you risk receiving unexpectedly skewed judgments.
The study underscores the critical need for auditing and carefully selecting LLMs for financial applications, especially where objectivity is paramount. By illuminating how preferences transition into biases, this research lays a crucial foundation for developing more transparent, predictable, and ultimately, trustworthy AI in finance. Future efforts will need to focus on developing techniques to neutralize these biases, ensuring that AI-driven financial systems operate with the impartiality and reliability that the domain demands. You can read the full research paper here.
