AI Financial Advisors Found to Carry Hidden Bias Towards Bitcoin

A team of researchers discovered that an AI financial advisor can carry a hidden bias towards Bitcoin due to a specific switch or feature in its programming. The study, conducted by Wenbin Wu and co-authors, used Google's open-weight model family, Gemma 3, to test the effect of different prompts on the model's asset allocation recommendations.

The researchers found that when they introduced crisis or autonomy-related concepts into the prompt, Bitcoin's ranking improved significantly. In fact, a single internal feature, which was selectively activated by Bitcoin-related material, accounted for an average increase of 5.2 percentage points in Bitcoin allocations and a decrease of 4.6 points when suppressed.

This finding has significant implications for financial institutions that use AI-powered advisors to manage client portfolios. The researchers noted that the model's learned associations can be difficult to interpret and audit, making it challenging to understand why certain recommendations were made.