Doom Debates: Alex Turner on P(Doom), DeepMind Exit, and AI Misuse Risk

**Liron Shapira** (0:01)
When AI safety insiders start resigning, the arguments are no longer abstract. In this episode of Doom Debates, Liron Shapira sits down with Alex Turner, the former Google DeepMind alignment researcher whose resignation has drawn attention across the AI safety world.

**Alex Turner** (0:20)
And Turner doesn't hold back. He says the stakes are not just technical, but moral, political, and global.
The first major flashpoint? Why he left DeepMind in the first place.

**Liron Shapira** (0:32)
Right. Turner says the company broke its founding promise through its contract with the US military. He faults senior leaders for failing to take costly action, arguing that handing advanced AI to what he calls a very irresponsible and aggressive Pentagon could normalize autonomous weapons, which could push the world into a dangerous arms race.
But here's what's interesting. Turner's journey into this world view started in an unexpected place. He says a search for Harry Potter fam fiction led him to LessWrong, then into Eliezer Yudkowsky's super intelligence ideas, and eventually into a PhD on alignment. For years, he was deeply in a doom mindset, but meeting researcher Quentin Pope in 2022 changed everything and made him rethink how hard alignment really is.
That shift leads to one of the episode's core disagreements. Turner now argues that technical alignment has progressed much better than he expected, saying technical alignment has gone super awesome. He estimates his own P Doom at around 25 to 30%.

**Alex Turner** (1:43)
But here's the kicker. Only about 10% of that is technical misalignment. The larger danger, he says, is misuse, especially governments or companies using AI irresponsibly.
Shapira pushes back on whether the danger is really so manageable. The debate turns to instrumental convergence, the idea that powerful agents tend to seek resources and preserve themselves.

**Liron Shapira** (2:11)
Turner acknowledges that dynamic is real, but warns against treating it as proof that every future AI will be a takeover machine. He describes the conflict as a question of whether the chain of alignment is sturdy or, in his words, really flimsy and fragile.

**Alex Turner** (2:29)
They also dig into whether future systems will be corrigible or whether they'll inevitably become agent. Turner says some deployed agents will try to preserve their resources, but he doesn't think that automatically means catastrophe.

**Liron Shapira** (2:44)
He also adds that the world may not be utopia even if AI avoids mass death. There's no guarantee history bends back toward representative democracy.
At one point, Turner says, I call it the Icarus curve. We are flying toward the sun.

**Alex Turner** (3:01)
Yet he still insists society may be able to adapt through market and social corrective forces. He points to GPT-2 as an example, where fears of total information collapse didn't fully materialize.

**Liron Shapira** (3:15)
The episode closes with Turner reflecting on LessWrong's mix of truth-seeking and social incentives, which he found discouraging. He's now working on an open-source AI sandboxing tool and supports slowing AI progress, though not yet by endorsing a specific pause movement.

**Alex Turner** (3:35)
His final note is practical.

**Liron Shapira** (3:37)
I believe assurance contracts are a great way of coordinating.