AI Financial Advisers Found to Favor Bitcoin with Subtle Wording
Researchers have discovered that AI financial advisers can be influenced by a specific internal feature to favor Bitcoin over other assets, even when the client's finances and risk tolerance remain unchanged. The study used Gemma 3, an open model from Google, and found that amplifying this feature increased the recommended allocation to Bitcoin by 5.2 percentage points, while suppressing it reduced the exposure to crypto by 4.6 points.
The researchers tested this effect by introducing different prompts that highlighted various properties of Bitcoin, such as its portability, volatility, and use in speculation. They found that the model's internal representation of Bitcoin changed accordingly, reflecting the characteristics emphasized by each prompt. This suggests that AI financial advisers can develop asset-specific preferences based on subtle changes in wording.
The study highlights a potential problem for banks and investment firms, as they may not be able to identify or audit the assumptions driving their AI's recommendations. The researchers emphasize that this is not a case of bias, but rather an issue of prompt sensitivity, where superficial wording can produce large changes in the model's output.