p(doom) Is Just Tech Bro Astrology
If you’ve ever watched someone say “p(doom)=0.7” with the calm certainty of a weather report, you already understand the main problem: they’re using a number to launder a story.
The reveal
"p(doom)" is an informal shorthand for the probability of catastrophic or existential outcomes from advanced AI. Sometimes it's a serious attempt to aggregate beliefs under uncertainty. Sometimes it's a vibe in a lab coat. The hard part is telling which one you're looking at.
The steelman (yes, but also no)
Yes: putting a number on uncertainty can force clarity, reveal disagreement, and make update triggers explicit. No: the number is not a substitute for the mechanism. A precise probability attached to an underspecified chain of events is just numerology with citations.
What the numbers look like when you actually ask experts
A large 2023 survey of AI researchers asked about "extremely bad" outcomes (including "human extinction") and also asked variants directly about extinction or severe disempowerment. Results included a median of 5% for "extremely bad" outcomes (mean ~9%), and substantial fractions assigning 10%+ in different framings. That doesn't prove doom. It does prove the discourse is not settled, and that framing matters.
Receipts (what people actually wrote down, not what you heard on Twitter)
The ESPAI 2023 survey of thousands of AI authors shows the distribution: median 5% (mean ~9%) for "extremely bad" outcomes, with large variance across framings for extinction/disempowerment questions. Meanwhile, Carlsmith's "Is Power-Seeking AI an Existential Risk?" from 2022 demonstrates the right approach—explicit premises with explicit credences, leading to an overall estimate that can actually be updated. For historical context, the Bostrom & Müller survey from 2013 provides older expert survey data that's useful for comparison and shows how question framing shapes answers. The CAIS "Statement on AI Risk" from 2023 demonstrates how "extinction framing" became mainstream messaging, often without the accompanying mechanism detail that would make it meaningful.
The astrology failure mode (how p(doom) becomes identity)
When the number becomes a social signal, it stops being an estimate and starts being a membership badge. That's astrology: the purpose is not to predict accurately, it's to communicate "who I am" and "who you should listen to." The AI version adds a Bayesian veneer and a GPU pun, but the dynamics rhyme.
So what's a non-embarrassing way to use p(doom)?
A p(doom) estimate is only as good as the decomposition behind it. If you can't break it into major premises (capability, deployment incentives, loss of control, irreversibility) and assign rough uncertainty to each, you're not estimating. You're gesturing.
Disproof conditions (what would make a p(doom) estimate look more like science)
If we observed public, pre-registered forecasts with measurable triggers and track records over time, we'd update because accountability would replace vibes. If we had standardized definitions of outcome classes—catastrophic versus existential—with consistent elicitation methods, comparability would improve and framing effects would shrink. If evals and incident data meaningfully constrained the probability mass, the number would become anchored to observed reality instead of floating in theoretical space.
Action line
If you want to say "p(doom)=X," you owe three things: a claim graph, evidence per node, and your update triggers. Without that, you're doing horoscope discourse with better math fonts.
Anyway, back to touching grass—where the only acceptable p(doom) is "probability of doomscrolling if you open the app again."
Related reading
Check out the doom-astrology-chart for the mechanism-first taxonomy that p(doom) should be downstream of. Then read vibes-vs-evals, because measurements are how we stop arguing about feelings. Finally, see how-to-raise-100m-with-doomer-fanfiction, because doom probabilities interact with incentives, not just beliefs.