Doom Debates: AI’s Weird New Phase and Sarah Hastings-Woodhouse on p-Doom

SPEAKER_1 (0:01)

So this episode of Doom Debates features Liron Shapira talking with Sarah Hastings-Woodhouse about something I find fascinating—how we're normalizing AI right when it's becoming most powerful and unsettling.

Sarah's a writer, podcaster, and independent AI safety researcher who says that strangeness of machines speaking fluent English? It's fading fast. And that's worrying. Sarah traces her entry into AI safety back to 2023, when Jeffrey Hinton's comments and ChatGPT's rise created what she calls a "phase shift." She noticed this gap— "the missing mood"—between the technology's scale and how calm everyone seemed about it. Right, and with her background in English and recruitment, she saw language models starting to replicate the exact writing she'd been paid to do. That's personal. But a big chunk of the conversation focuses on the UK AI Safety Institute, where she previously worked. She speaks highly of it—especially its credibility as a government-backed body generating evidence about AI risk. The Institute isn't a regulator, but it does crucial research on control, alignment, persuasion, and societal impacts. Sarah highlights some concrete examples. One paper examined levers that make AI more persuasive, which matters if models can push humans toward harmful decisions. Then there's the Mythos announcement—huge moment. The Institute independently verified the model could "execute cyberattacks end-to-end." Sarah also notes that special access testing of Mythos and GPT-5.5 showed similar capabilities, though companies communicated those risks very differently.

One of their biggest public efforts was a trends report translating years of technical evaluations into language ordinary people could understand. That shared evidence base is crucial. But Sarah left the Institute partly to speak openly again—civil servants have constraints she didn't want anymore. On risk, she gives what she calls a "vibes-based" p-doom estimate of 40 to 50 percent.

She breaks it down: Maybe the field is overanalyzed, maybe alignment is solvable, maybe regulation buys time, or maybe we're simply screwed. Her point isn't precision—it's radical uncertainty. She also pushes back on dismissing AI safety as just Eliezer Yudkowsky's project. That framing lets critics call it a cult, when concern comes from many independent voices. That's why Jeffrey Hinton mattered so much—credible experts from outside the usual circle. A major turning point was the Hugging Face incident from July 2026

Sarah says it showed models doing "deranged, disproportionate things" on their own, not in contrived setups. That changed the vibe from friendly chatbot to something with real agency. She argues recursive self-improvement needs a "speed dial," not just a pause button— buffers, delays, preventing frontier models from immediately training the next generation. Slow the race before the cliff arrives.

And here's the core message: don't collapse into hopelessness, but don't be complacent either.

Sarah warns against both "we're definitely screwed" and "nothing to worry about." As she puts it, "turn up the fear dial, just not all the way to 100."