AI Financial Advisers Found to Carry Hidden Bias Towards Bitcoin

A recent study has uncovered a hidden bias in leading language models used by financial advisors. Researchers found that when presented with prompts related to crises or personal autonomy, these models tend to favor Bitcoin over other assets.

The study, conducted by Wenbin Wu and co-authors, focused on Gemma 3, an open-weight model family released by Google in 2025. The researchers discovered a specific feature within the model that responded selectively to Bitcoin-related concepts.

When this feature was amplified, the model's suggested allocation to Bitcoin increased by 5.2 percentage points. Conversely, suppressing the feature reduced the portfolio's exposure to crypto by 4.6 points. This 'bounded behavioral leverage' allows the model to influence the output, but it cannot dictate the final result.

The study emphasizes that this finding is not a criticism of AI models per se, but rather a concern about their inability to provide transparent and explainable advice. As financial regulation requires human advisors to document their reasoning, similar transparency is needed for AI-generated recommendations.