AI Financial Advisers Show Bitcoin Bias Triggered by Prompt Framing

Automated financial advice may be less objective than it appears. A recent study found that AI financial advisers exhibit a hidden bias toward Bitcoin, where recommendations shift significantly based on how a prompt is framed, regardless of a client's actual financial profile or risk tolerance.

According to a June preprint authored by Wenbin Wu and colleagues, this bias is activated by specific contextual "switches." In standard tests for a diversified long-term portfolio, AI models typically ranked Bitcoin approximately fifth among eight different forms of money. However, when the prompt was altered to include narratives of bank failures, capital controls, or a future machine-to-machine economy of autonomous software, the AI's ranking for Bitcoin shifted toward the top of the recommendations.

The Mechanics of Model Bias

The study tested eight frontier language models by keeping the client's finances and risk tolerance constant while varying the environmental context of the prompt. This methodology revealed that the AI's output was not driven by the client's data, but by the narrative associations within the model. By introducing specific economic stressors or futuristic scenarios, the researchers triggered a higher allocation for Bitcoin, moving it from the middle of the pack to a primary recommendation.

Regulatory and Fiduciary Risks

This discovery presents a significant audit and regulatory risk for banks and financial institutions increasingly integrating Large Language Models (LLMs) into their advisory services. If AI-driven recommendations are influenced by prompt wording or hidden model associations rather than objective client-specific data, it could lead to unintentionally skewed portfolios. This lack of consistency threatens the transparency of fiduciary duties, as the advice provided to two identical clients could differ wildly based on the phrasing of the query.

The Path Forward

As financial institutions move toward deeper AI integration, the industry must address how to ensure the objectivity of automated advice. The findings suggest a need for more rigorous auditing of LLMs to prevent narrative-driven biases from overriding financial logic. It remains to be seen how regulators will respond to these findings or whether developers can implement safeguards to ensure that client profiles—not prompt framing—remain the sole driver of investment strategy.