Original title: Auditing Asset-Specific Preferences in Financial Large Language Models: Evidence from Bitcoin Representations and Portfolio Allocation
Wenbin Wu · June 2026
People now use AI models to help trade and manage money, but we do not know if these models carry hidden biases toward certain assets. Researchers created a test to check if AI systems favor Bitcoin. First, they asked nine top models to rank money types. The models ranked Bitcoin around 5 of 8 for reliable money, but ranked it near the top during crises or for autonomous agents. A swap test showed models care about an asset's actual features, not its name. Next, the team looked inside one model, Gemma 3. They found a specific internal feature for Bitcoin. When they turned this feature up, the model favored Bitcoin more. Turning it down made the model favor Bitcoin less, even without mentioning the word Bitcoin. Finally, this change altered investment advice: turning the feature up raised Bitcoin's portfolio share, while turning it down cut it.
Imagine deciding which friend to ask for a favor. Normally you consider each friend's helpfulness. But if your judgment secretly depends on the order you think of them, you might pick the first name that pops into your head even if another is better. In this paper, the model's preference for Bitcoin shifts depending on the framing of the question, and a single internal feature acts like that first name: amplifying it makes Bitcoin more likely to be chosen, suppressing it makes Bitcoin less likely, even when Bitcoin isn't mentioned.
If an AI advisor secretly favors Bitcoin, your portfolio could be nudged without you knowing. This research shows those biases are real, traceable to a single internal feature, and can be moved up or down by hidden tweaks. As LLMs take over financial decisions, regulators will need new rules to check what an agent prefers and how far that preference can be tilted.
This audit only tests Bitcoin and opens the internals of a single model (Gemma 3). The leverage is bounded—effects are measurable but limited in magnitude. The behavioral audit covers a specific set of frames and may not generalize to other assets or models without further validation.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Auditing Asset-Specific Preferences in Financial Large Language Models: Evidence from Bitcoin Representations and Portfolio Allocation — Wenbin Wu
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.