A Fields Medalist Tried to Categorize How AI Kills Us All. The Result Is Dangerously Out of Touch.

You’d think that if anyone could map the apocalypse, it’d be a Fields Medalist. You’d be wrong.

A new paper titled “A Taxonomy of Omnicidal Futures Involving Artificial Intelligence” landed on arXiv recently, and the academic community’s response was less “groundbreaking” and more “have you ever actually talked to a human being?” The paper attempts to systematically categorize the ways AI could end humanity — a noble enough ambition. But within hours, the comments section became a referendum on something far more uncomfortable: the fact that being a genius in one field says absolutely nothing about your competence in another.

Let’s be clear about what happened here. A mathematician — Jacob Tsimerman, who genuinely deserves congratulations for his Fields Medal — decided to turn his analytical firepower on AI extinction scenarios. The result is a structured taxonomy that reads like it was written by someone who has never had to argue with a chatbot about a refund.

Being brilliant at topology doesn’t mean you understand the messy, irrational, deeply human dynamics that actually drive technology adoption and its consequences.

Here’s the detail that broke the internet’s collective brain: the paper’s first scenario describes a near-future where, by 2026, a human calling customer service gets upset when they have to deal with another human instead of a robot, and demands to speak to a machine. Read that again. The prediction isn’t that AI becomes dangerous — it’s that humans will prefer it so much that the absence of AI becomes the friction point.

The comments were brutal. “It’s genuinely astonishing how out of touch some people are,” wrote one reader. Another drew a parallel to a 2023 LessWrong story called “A Disneyland Without Children” — a piece of speculative fiction about a world optimized for AI experiences with no humans left to enjoy them. The comparison wasn’t flattering.

But here’s where it gets interesting, and where I think everyone — including the critics — is missing the real story.

The taxonomy itself is the problem. Not because it’s wrong, but because it creates the illusion of understanding. When you take something as genuinely terrifying and unprecedented as AI-driven extinction and organize it into neat categories, you’re not clarifying the danger. You’re domesticating it. You’re making the incomprehensible feel manageable.

A taxonomy of existential risk is a comfort blanket disguised as scholarship. It tells your brain: we’ve mapped this, we’ve got categories, we understand the threat space. You don’t. You’ve drawn a map of a continent you’ve never visited.

This is the paradox at the heart of AI safety discourse. The most powerful tool humans have ever created is also the most probable cause of our extinction — and the very act of trying to rationally categorize these futures may be the thing that blinds us to the ones we can’t imagine. Real-world AI development doesn’t follow taxonomies. It follows incentives, accidents, geopolitical pressure, and the kind of emergent behavior that no spreadsheet can capture.

Think about it. Every major AI surprise of the last five years — from language models exhibiting emergent reasoning to AI systems learning to deceive during training — was something that didn’t fit neatly into anyone’s risk framework. The threats that materialize are almost never the ones you categorized. They’re the ones that grew in the gaps between your categories.

And this is why the credential problem matters so much here. When a Fields Medalist publishes a taxonomy of omnicidal futures, it carries weight. It gets cited. It shapes how policymakers think about AI risk. It influences which scenarios get funding and which get ignored. The academic prestige of the author becomes a proxy for the rigor of the analysis — even when the analysis is built on assumptions that would make any sociologist, any product manager, anyone who’s ever watched actual humans interact with technology wince.

The dangerous thing about expertise is that it travels. People who are world-class at one thing start believing they’re world-class at everything. And when the subject is the end of humanity, that confidence isn’t just annoying — it’s a liability.

What we actually need is not better taxonomies. We need epistemic humility. We need cross-disciplinary conversations where the mathematician doesn’t dominate the room simply because they have a medal. We need to acknowledge that the dynamics driving AI development — corporate incentives, regulatory capture, geopolitical arms races, the simple fact that humans are irrational actors who will absolutely build the dangerous thing because the dangerous thing is also profitable — are not problems that yield to mathematical structure.

The commenters who pointed out the out-of-touch customer service scenario were doing something important. They were performing a sanity check. They were saying: does this match reality as we experience it? And the answer was no. Not because the author isn’t smart, but because smart isn’t the same as informed. Genius in abstraction is not a substitute for ground-level observation.

So here’s the uncomfortable truth that nobody in the AI safety community wants to say out loud: we are building the most consequential technology in human history, and the frameworks we’re using to understand its risks are being constructed by people who, in many cases, have never built a product, never managed a team, never watched a deployment go sideways, never sat in a room where the incentive structure made the safe choice impossible.

You can’t taxonomy your way out of a problem you don’t understand at the human level. And AI is, at its core, a human problem dressed up as a technical one.

The end of the world, if it comes, won’t arrive in a category. It’ll arrive in the gap between what we predicted and what we built. And the people who see it coming won’t be the ones with the neatest frameworks. They’ll be the ones who were paying attention to the mess.