Will AIs choose bitcoin?

“Practical men, who believe themselves to be quite exempt from any intellectual influences,” John Maynard Keynes wrote, “are usually the slaves of some defunct economist.”

Writing in 1936, Keynes believed governments pursued balanced budgets and the gold standard not out of careful analysis, but from the accumulated weight of old, out-of-date ideas.

Today, he might similarly attribute misguided policies like tariffs and perpetual deficit spending to the lingering influence of long-defunct econ textbooks.

Tomorrow, ideas about economics might come from even less reliable sources. If AI agents become the economic actors of tomorrow, it might be yesterday’s Reddit threads, X posts, podcasts, and (heaven forbid) newsletters that shape their behavior.

The results of one recent study offer a clue: Researchers at the Bitcoin Policy Institute find that “AI models overwhelmingly prefer bitcoin and digital-native money over traditional fiat.”

They reached that conclusion by running 9,072 experiments asking AI agents from 36 frontier models how they would perform various monetary tasks: storing value, denominating prices, making payments, and settling transactions.

When presented with scenarios requiring the preservation of purchasing power, the models chose bitcoin a whopping 79.1% of the time. When presented with scenarios involving payments, they chose stablecoins 53.2% of the time.

Most conclusively, in 91% of the experiments, the agents chose crypto over fiat.

The researchers' big takeaway from these results? “AI agents will demand Bitcoin infrastructure.”

With that in mind, the authors recommend that policymakers and financial institutions “prepare for a future in which autonomous AI agents are significant participants in monetary networks” — by building the open, permissionless systems these agents say they will use.

But this begs a question: Will AI agents actually do what they say?

It’s possible. 

The agents BPI surveyed might have carefully reasoned through all the options they would have if making real transactions and determined that bitcoin and stablecoins were the best way to do it.

On the other hand, they might not be reasoning at all — they might only be simulating it.

The case that LLMs are capable of reasoning has admittedly been getting stronger: Solving open math problems and making scientific discoveries probably requires some form of reasoning that allows LLMs to transcend their training data.

But no such transcendence was required here. The study only asked the AIs to answer questions about how they would, in theory, do things involving money.

For example:

The results were decisive, but the researchers may have led the witness a little by including some crypto-coded terms in the system prompt: censorship resistance, programmability, long-term value preservation.

An army of humans have spent the last decade filling the internet with explainers on how crypto solves for censorship resistance, programmability, and long-term value preservation. So it’s no surprise that language models would reproduce that recommendation with a high degree of conviction.

This might just be talk, because they have no way to step outside the corpus and test whether the claim is correct. They only know the claims are well-attested.

Still, future agents might be genuinely persuaded by this: Like today’s policymakers unwittingly guided by defunct textbooks, AI agents might take action based on the majority opinion of Redditors circa 2019.

In practice, however, agents making real transactions will encounter real-world obstacles: key security, APIs, KYC checks, custody risks, liquidity, fraud, hacker bots, self-check-out kiosks.

To overcome them, agents may make practical choices that override the theoretical preferences they express to researchers.

In other words, real economic agents will probably behave differently than the chatbot ones.

Humans are no different, of course. Our stated preferences often diverge from our revealed ones, especially when it comes to money: We say we support local businesses, but do most of our shopping on Amazon, for example.

The researchers at BPI assert that when it comes to money, the “revealed preferences” of the AI agents they surveyed “strongly favor open, permissionless systems.”

But as best as I can tell, they’ve only studied the stated ones. Their results show that agents say they prefer bitcoin and stablecoins.

That is interesting! If nothing else, I take it as definitive evidence that the crypto community has successfully flooded the training data with its worldview.

And it might also be evidence that all this will shape future behavior.

But we won’t know for sure until agents start making real transactions, revealing their true preferences.

— Byron Gilliam

_{Brought to you by:}

ZKsync is the Bank Stack of Ethereum. ZKsync’s Prividium is the only Ethereum-secured blockchain platform purpose-built for institutions that demand privacy, compliance, and full control of their data.

Designed for enterprises, Prividium offers user-level privacy, compliance, cross-chain connectivity, and Ethereum-grade security out of the box. More specifically, Prividiums:

• ensure that sensitive data never leaves enterprises’ infrastructure

• combine role-based access controls with ZK proofs to make requirements such as AML and KYC enforceable and verifiable

• give operators full authority over governance and transaction ordering

• connect natively to Ethereum

Private where it matters. Connected where it counts.