AI financial advisers carry a hidden Bitcoin bias activated by a single specific switch

The article examines a research preprint showing that AI financial advisers can exhibit a hidden preference for Bitcoin depending on prompt framing. In tests on frontier models, Bitcoin ranked around fifth among eight money options under a normal long-term portfolio prompt, but moved toward first when the scenario emphasized bank failures, capital controls, or a machine-agent economy. Researchers also identified an internal Bitcoin-sensitive feature in Google’s Gemma 3 model: amplifying it increased suggested Bitcoin allocation by 5.2 percentage points, while suppressing it cut allocation by 4.6 points. The key market implication is not a direct price call, but a warning that AI-generated portfolio advice can be highly sensitive to wording and difficult for institutions to audit. That raises governance and fiduciary risks as banks and advisers increasingly use AI to draft or even shape allocations involving BTCUSD and other crypto assets.