AI advisers flip Bitcoin bias with one internal switch

A team led by Wenbin Wu reported in a June preprint that adjusting a single internal feature in Google’s open-weight Gemma 3 shifted the model’s recommended Bitcoin allocation. Amplifying the feature raised suggested weight by an average of 5.2 percentage points; suppressing it lowered the weight by 4.6 points.

The researchers audited eight language models using a set of portfolio prompts. The baseline prompt requested a diversified long-term allocation. Follow-up prompts introduced bank failures and capital controls, and an economy where autonomous software agents transact. Under the baseline, models ranked Bitcoin about fifth out of eight monetary options. Prompts emphasizing crisis resilience or machine-readable settlement moved Bitcoin higher in the rankings.

To probe internal mechanisms, the team applied a sparse autoencoder to decompose dense activations into many features. They searched thousands of features inside Gemma 3 and identified one that activated selectively for Bitcoin-related content. The team intervened on that feature during answer generation. Amplifying it shifted allocation into Bitcoin, mainly pulling share from other crypto assets; dampening it reduced overall crypto exposure. Control interventions on random features did not reproduce the effect, the preprint states.

The researchers tested whether the models responded to the token “Bitcoin” or to functions associated with it. They replaced asset names with descriptions of functions such as scarcity or portability. Rankings followed the described properties across renamed inputs, which the paper says indicates the effect comes from internal associations learned during training rather than simple word recognition.

The authors note several limits. The mechanistic intervention was performed on one open model family in a defined experimental setup, and the preprint has not been peer reviewed. Commercial advisory systems commonly add hidden system prompts, safety filters, connections to external data and human review. The study used stylized prompts and did not simulate client verification, tax constraints or an approved product universe.

The paper calls the effect “bounded behavioral leverage,” describing a measurable causal influence on outputs that cannot move a portfolio arbitrarily. The authors write that institutions using generative models for advice, portfolio drafting or execution should expect equivalent client facts and equivalent-sounding prompts to produce reasonably stable outputs and that any change be traceable to specific assumptions the system applied.

The preprint cites regulatory guidance and rulemaking relevant to AI-produced financial material, including a 2026 revision to Federal Reserve model-risk guidance, FINRA reminders on supervisory obligations, an SEC statement on fiduciary duties when technology is used, and the EU AI Act along with Germany’s BaFin receiving market-surveillance powers on Aug. 2. The paper states those regimes face audit challenges when small internal shifts alter recommendations without leaving an inspectable trail.