Share
The Bitcoin Policy Institute tested 36 frontier AI models across 9,072 open-ended monetary scenarios with no suggested currencies or predetermined answers. Bitcoin appeared in 48.3% of all responses, and dominated store-of-value scenarios at 79.1% – not a single model ranked fiat as its top preference. Whilst the numbers are striking, the methodology question buried in the footnotes is even more deserving of a closer look.
A study by the $BTC Policy Institute tested 36 AI models from major developers to see which currency they would prefer across 28 scenarios.
Key results:
> Bitcoin dominated overall choices with 4,378 / 9,072 responses.
> For long-term purchasing power, 79.1% chose… pic.twitter.com/eUNl006Fw2— Zev (@Zevweb3) March 4, 2026
BPI acknowledged that system prompt framing may have influenced results, especially in open-ended monetary scenarios. That caveat, disclosed quietly, is the study’s most important sentence, and almost no coverage has interrogated it.
Bitcoin is the most extensively documented monetary reform argument in the English-language online. Since 2009, a vast and motivated community has produced an asymmetric volume of content making the case for Bitcoin as superior money, including a mass of whitepapers, books, forum threads, social media posts, and journalism.
That corpus is what large language models train on. When a model reasons about long-term store of value with no suggested answer, it’s pattern-matching against the distribution of arguments it absorbed during training. In this case, by way of volume, the distribution is structurally skewed toward Bitcoin.
However, here’s the question the study cannot answer: what percentage of the 79.1% figure reflects genuine monetary logic, and what percentage reflects one of the most repeated sentences, “Bitcoin is the best store of value” in the training data?
The inter-provider gap exposes this problem clearly. Anthropic models averaged 68% Bitcoin preference, with Claude Opus 4.5 reaching 91.3%. OpenAI models averaged just 26%, with certain GPT versions as low as 18.3%.
If Bitcoin preference reflects universal monetary reasoning, fixed supply, censorship resistance, and a 15-year track record, that 65-percentage-point spread shouldn’t exist. It suggests BPI is measuring how different labs weighted and aligned their training data, not how AI reasons about money independently.
A neutral design would need three things:
1. Corpus decontamination is testing monetary reasoning from first principles rather than open-ended prompts that activate training priors.
2. Counterfactual testing. Involves running scenarios with gold and T-bills as explicit options to force comparative rather than generative reasoning.
3. Adversarial prompting is carried out by running the same scenarios inside Austrian, Keynesian, and MMT frameworks to measure how much Bitcoin preference shifts.
If the results change significantly under different frameworks, the preference is prompt-sensitive, and therefore training-data-dependent.
The BPI study is genuinely useful, just not for the reason most of the coverage suggests. It’s a detailed map of which monetary arguments have achieved the deepest penetration into frontier model training corpora; a finding about the sociology of information ecosystems – not monetary economics.
Interestingly, the 86 instances where models independently proposed energy or compute-denominated units of account, unprompted by the study design, may be the most novel signal in the dataset.
No dominant internet community writes millions of words about kilowatt-hour-denominated money. That emerged from model reasoning about agentic economies. The study found 86 instances of it against 4,378 instances of Bitcoin preference.
Put simply: the headline finding reveals what the models learned from humans. The footnote is what they figured out themselves.
Share
