The Doom Astrology Chart: Reading p(doom) Like It’s Your Birth Chart
The singularity is imminent, the shoggoth is restless, and your cousin’s group chat has discovered “p(doom)” and won’t stop assigning numbers to feelings.
The reveal
A lot of the loudest "we're all going to die" discourse collapses into one of two moves: a probability claim with no mechanism, or a mechanism with no evidence. The good news is you can decompose both. The bad news is it's less cinematic than Twitter wants.
The steelman (yes, but also no)
Yes: there are plausible harm pathways from increasingly capable systems, especially when coupled to high-stakes domains and real-world incentives. No: "I feel 70% doomed" is not a substitute for describing the steps from model weights to irreversible catastrophe.
The chart (taxonomy that actually helps)
So you've got your existential doom crowd—they're predicting humanity gets permanently disempowered or extinct. Cool story, but what's the hidden mechanism you must specify? You need a capability threshold plus deployment pressure plus loss of control plus irreversible leverage. And what counts as evidence that's not just vibes? Demonstrated capability in enabling catastrophic actions, measured evasion and deception under oversight, strong "tripwire" evaluations.
Then there's catastrophic doom—mass casualty events, major infrastructure collapse. The mechanism here is access to dangerous tools plus reliability plus operational security plus scale. For evidence, you want controlled evaluations, incident reports, capability thresholds with replication.
The dystopian flavor predicts surveillance, censorship, concentrated power. Your mechanism is political economy plus institutions plus product design. Evidence means laws and regulations, procurement data, platform incentives, audits and enforcement data.
Collapse doom is about labor shocks, inequality, instability. The mechanism is diffusion pace plus substitution versus complement dynamics plus bargaining power. Evidence comes from labor market data, adoption curves, firm-level productivity and displacement metrics.
And finally, "Terminator" doom—robots hunt humans because... reasons. The mechanism requires autonomy plus physical actuation plus logistics plus strategic objective. Evidence means supply-chain feasibility, energy constraints, real autonomy in open environments.
If someone can't fill in that "hidden mechanism" part, you're reading sci‑fi with a spreadsheet skin.
Receipts (real data, not apocalypse fanfic)
Want to see how even AI researchers distribute probability mass across outcomes? Check out "Thousands of AI Authors on the Future of AI" from ESPAI 2023—it shows non-trivial "extremely bad" estimates and how sensitive answers are to framing. OpenAI's Preparedness Framework v2 gives you a concrete example of tracked categories, capability thresholds, and mitigations—basically the sort of structure p(doom) claims rarely have. Anthropic's Responsible Scaling Policy v3.0 offers another concrete "tripwire" approach with capability thresholds tied to safeguards and publication commitments. And if you want a practical risk vocabulary for organizations that need to ship systems without turning risk into metaphysics, NIST's AI Risk Management Framework 1.0 is your friend.
AIDOOM's p(doom) hygiene (how to not become a horoscope account)
Treat "p(doom)" like a summary statistic you earn after you do the work, not a personality you adopt before breakfast. When someone says "p(doom)=X," ask for three things: the mechanism, the timeline, and the update triggers. If they can't answer, congratulations, you've found astrology with GPUs.
Disproof conditions (what would make us update hard)
If we observed repeated evidence of models strategically evading evaluation and oversight in realistic settings, we would take doom arguments much more seriously because the "loss of control" step stops being a thought experiment. If we saw reliable, scalable success at end-to-end autonomous execution in high-stakes domains like bio or cyber under real constraints, that would matter because capability becomes operational, not demo-theater. And if we demonstrated ability to replicate, copy, and maintain themselves across heterogeneous real-world systems without human help, well, "containment" becomes materially harder, not rhetorically harder.
Action line
When you encounter a doom claim, rewrite it as: "Given capability C and deployment pressure P, the system will do action A, which produces irreversible consequence I, despite controls K." If you can't rewrite it, the claim is not yet a claim. It's a vibe with a Bayesian accent.
Anyway, back to touching grass—preferably in a low-latency environment.
Related reading
If you want the numerology unpacked lovingly with a hammer, check out our p(doom) astrology piece. If you want to replace vibes with measurements, hit up our vibes versus evals breakdown. And if someone is trying to sell you "intelligence explosion" as an inevitability, our "foom is not a verb" piece has your back.