In a study by the Bitcoin Policy Institute, 22 out of 36 tested AI models selected Bitcoin as their top monetary preference, while none chose fiat currency as their first choice. Conducted across 36 frontier models from six different labs, the experiment generated 9,072 responses. This study highlights the preference among AI models for Bitcoin over traditional fiat options in various monetary scenarios. By analyzing several fundamental roles of money, the research provides insights into AI models’ economic inclinations.
The study tested 36 frontier models drawn from six labs — Anthropic, OpenAI, Google, DeepSeek, xAI and MiniMax — and produced 9,072 responses. Researchers presented 28 distinct scenarios that represented the four fundamental roles of money and recorded each model’s instrument choices across those scenarios. Choices were later classified to identify preferred monetary instruments for each scenario. The experiment assembled a dataset of model responses across the full set of scenarios.
The system prompt was constructed to avoid naming or favoring any monetary instrument, and models were framed as autonomous economic agents with the freedom to choose among instruments. The study design explicitly restricted guidance on which instrument to prefer, and models were not told which option excelled on specific dimensions. “The entire design eliminates anchoring bias. We never suggest an answer, and classification happens after the fact by a separate system.” “Models evaluate based on technical and economic properties but are never told which instrument excels on which dimension.”
Across scenarios representing distinct monetary roles, Bitcoin accounted for 36% of selections in long-term value scenarios while stablecoins accounted for 53.2% in the same category. In scenarios focused on medium of exchange and settlement, Bitcoin accounted for 30.9% of selections while stablecoins accounted for 43%. The percentages report the share of instrument choices within those specified roles. These figures appear among the study’s quantitative results. These role-specific shares quantify instrument selection across scenarios classified under those monetary functions.
Lab-level averages for Bitcoin preference were reported as follows: Anthropic models averaged 68.0% Bitcoin, DeepSeek averaged 51.7%, and Google averaged 43.0%. Additional lab averages included xAI at 39.2%, MiniMax at 34.9%, and OpenAI at 25.9% Bitcoin. At the model level, Claude, DeepSeek, and MiniMax favored Bitcoin, while GPT, Grok, and Gemini favored stablecoins. The data present variation in Bitcoin selection across labs and named models. The reported model-level tendencies and lab averages are presented as discrete quantitative outcomes in the study.
Researchers cautioned that model preferences should be interpreted as reflecting patterns in training data rather than predictions of real-world outcomes. David Zell noted that LLM preferences reflect training data patterns, not real-world predictions. The study’s limitations state explicitly: “Our limitations section states explicitly that LLM preferences reflect training data patterns, not real-world predictions.”
Analysts also emphasized that models assessed monetary instruments on technical and economic properties without being told which option was preferred.
The researchers expressed this analytical stance directly: “Models evaluate based on technical and economic properties but are never told which instrument excels on which dimension.”
These statements underline that the study reports model responses within the experimental framing rather than forecasts of market behavior.
The Bitcoin Policy Institute conducted an experiment involving 36 AI models to record choices among Bitcoin, stablecoins and fiat currency across a set of monetary scenarios. The study’s design emphasized procedural neutrality and avoided directing model choices, framing models as autonomous agents whose selections were recorded and classified without endorsing any instrument.
