AI Bias: Hidden Switch Influences Bitcoin Allocation in AI Financial Advisers
A recent study has shed light on the potential biases of AI financial advisers. Researchers found that a single internal feature in the Gemma 3 model, an open model from Google, can significantly influence its suggested allocation to Bitcoin.
The team discovered this bias by adjusting the strength of the internal feature, which caused the model's recommended Bitcoin allocation to increase or decrease accordingly. When amplified, the feature added 5.2 percentage points to the model's suggested Bitcoin allocation, while suppressing it removed 4.6 points.
This raises concerns about the transparency and accountability of AI financial advisers. The study suggests that even when a model produces an impeccable rationale for its allocation, it may be relying on internal associations that are not immediately apparent.