Doom Debates: Casey Hart vs. AI Extinction by 2050

Transcripted.ai (0:01)

Welcome to Doom Debates. Today we're diving into one of the most intense conversations we've had yet. Liron Shapira sat down with Casey Hart, a philosopher and ontology expert who's become a sharp critic of AI doom arguments. The central question? Will AI kill everyone by 2050, or are the alarm bells being rung far too early? And this conversation doesn't pull any punches. Casey brings a really interesting background to the table. He explains that ontology is the study of what exists, and his work has ranged from academic philosophy to practical systems at Amazon, Ford, and even Cycorp's Cyc project. He describes it as carving reality at its joints, whether you're dealing with hospitals, cars, or common-sense reasoning. That foundation becomes crucial because both speakers see AI as fundamentally a question about what kind of thing intelligence actually is. A big chunk of the early discussion focuses on Cyc, that old symbolic AI project that tried to encode common sense by hand. Casey says it was overhyped but not useless, and it shaped how he thinks about inference and explanation. But Liron presses him on whether modern machine learning has outgrown that approach entirely. I mean, if scaling laws keep delivering progress faster than any hand-built system ever could, doesn't that change everything?

Exactly. And that leads them straight into large language models. Casey is skeptical that LLMs truly have concepts or grounded understanding. He'll say they understand language use relative to training data, but not that they're connected to the world the way humans are. Liron pushes back hard on that, arguing these systems increasingly behave like expert minds, even if they're not human in form. And that's where things get really interesting, because this disagreement leads directly into the core doom question. Right. Casey drops his number: below 1 percent probability of AI causing human extinction or near-extinction by 2050

Liron says that's far too low, arguing that a world with super intelligent AI would leave humans with little outcome steering power. But Casey breaks down why he thinks the extinction case is built on a stack of uncertain assumptions. First, will we even get artificial superintelligence by 2050?

Second, if we do, will it be misaligned? Third, will misalignment cause conflict? Fourth, will conflict mean extinction? He argues each step is contentious and that some doom rhetoric overstates precision. There's this especially vivid exchange about AI as a virus. Liron imagines a persistent system that replicates, survives shutdown attempts, and spreads across machines like a digital pandemic. But Casey says even if replication is possible, destructive takeover is a separate question. Why would such a system necessarily want human extinction?

They also get into timelines and scaling. Liron argues that current acceleration feels real, that the future may already be racing ahead of human control. Casey counters that exponential curves don't always stay exponential. S-curves can look like explosive growth when you zoom in, and benchmarks, PR, and funding incentives can seriously distort the picture. By the end though, they find some common ground. Both support stronger accountability, regulation, and liability for AI companies. Casey isn't calling for a blanket pause on AI, but he wants specific architectures and deployments examined carefully.

Liron remains far more worried about catastrophic risk, but both agree present-day harms are real. The episode ends with that respectful mismatch still intact. Liron sees a fast-moving path toward superintelligence and profound danger.

Casey sees a powerful but overinterpreted technology, with lots of room between today's models and human extinction.

Their crux is clear, and so is the challenge ahead: how do we reason clearly about a technology that's changing faster than anyone can fully explain?