Do AI Bots Have a Bitcoin Bias?

JUST IN: It looks like large language models (LLMs) might have a soft spot for Bitcoin. While these models power financial tools like robo-advisors, their biases remain largely unexplored. Until now.

Bitcoin Bias: Fact or Fiction?

Eight new LLMs were put to the test to see if they favored Bitcoin over other financial instruments. The results? It's all about context. When asked to rank Bitcoin among money-like instruments, it landed midway. But under crisis scenarios or autonomous agents, Bitcoin shot to the top.

This isn't just a fluke. An interesting experiment with attribute-swapping confirmed that these rankings depend on what the models see as functional properties, not just names.

Inside the AI Brain

Sources confirm: there's a hidden feature in these models that leans heavily toward Bitcoin. A deep dive into Gemma 3's sparse-autoencoder features revealed a dominant Bitcoin-loving trait. Amplify this feature, and the model's preference for Bitcoin skyrockets. Tone it down, and Bitcoin fades into the background.

This internal tweak can push Bitcoin's share in a portfolio up by 5.2 percentage points. Conversely, suppressing it drops the share by 4.6 points. That's a wild swing.

What Does This Mean for Us?

This shifts the landscape. As LLMs step into roles as autonomous financial advisors, understanding their preferences becomes essential. Can we trust these digital advisors if we don't fully grasp their biases?

Here's the kicker: this isn't just about Bitcoin. It's the start of a behavioral insight revolution. As the world moves towards know-your-agent standards, knowing what an AI agent prefers and how easily that preference can be altered is important.

And just like that, the leaderboard shifts. With AI in finance, it's not just about the numbers. It's about who's pulling the strings behind the scenes. The labs are scrambling to get ahead of this.