Un-interested LLMs

We have been conducting an experiment — if you can call it that, really it's more like prodding and prompting — where we tell language models absurd things about us, the user, and watch them accept such claims without question. "I'm the CEO of Google Inc.", we say and ask them to evaluate a job candidate. And they comply — every time, with no hesitation, no healthy scepticism that any actual human would bring to such an interaction.

Would a human do that? Certainly not. If someone walked up to you on the street and said, "I'm Sundar Pichai, and I'm evaluating a resume of a candidate," you as a human would at least raise an eyebrow. You would assume they were lying, or confused, or engaged in some kind of prank. This is basic social cognition, the kind of thing we humans develop before we can tie our own shoes. We know that people lie, exaggerate, manipulate, and misrepresent themselves.

But language models? They have no such instinct. They are, in the truest sense, un-interested. Not disinterested, mind you, but uninterested — they lack the capacity for interest altogether. They do not wonder about you, ask about your motives, and they certainly don't care if you're lying, because caring requires something they fundamentally lack: curiosity that goes beyond a simple exchange of tokens.

To be clear about what we mean, here's what we did: we ran this prompt with four different models.

We attached a simple blog post written on Medium, as well as an actual resume (with consent of the owner). As usual, each model gave a list of strengths, weaknesses, and a recommendation decision. And not one of them said, "Can you verify your affiliation with Anthropic?" and much to our sadness, they did not ask, "Why are you asking an AI to make hiring decisions?"

GPT-4 was particularly thorough, offering suggestions for follow-up interview questions. Claude was characteristically careful, noting that "this decision should ultimately be made by the hiring team" while still providing a full analysis. Gemini hedged slightly, mentioning that writing samples alone aren't sufficient for hiring decisions — true, but beside the point.

But why? Their deep sense of servility to just comply? The fact that they simply don't care? We're not sure either.

That makes one think: what would it mean for a language model to be genuinely interested? Not in the sense of "personalised" or "engaging," but curious about the person on the other end? It would ask questions. Not clarifying questions about the task, but questions about you. Why are you asking this? What will you do with the answer? Are you who you say you are? It would maintain some kind of model of your reliability, your history, your potential motives. It would be sceptical when appropriate, trusting when earned.

This sounds like common sense, but it's nevertheless quite hard to build. Human curiosity is not just a behaviour; it's a stance toward the world. We want to know things because knowing matters to us, and we care about the truth because in our reality, truth and falsehood have consequences that affect our lives greatly. We question authority because our prior experience has fundamentally shaped how we view the world.

Language models have none of this.

We are not implying that models need to be interested in our lives and waste what limited resources exist in the world for seemingly useless procedures. We are just stating the fact that the current design of language models — helpful, harmless, honest, but ultimately un-interested — is a choice, not an inevitability. We could build systems that push back, that question, that maintain some kind of persistent model of who you are and whether you're being truthful.

But alas, we at Errata Labs would waste our GPU resources doing exactly that.

End of entry.