Every honest threat list eventually hits the one it can’t cleanly model. For us that’s AI catastrophe. It’s the newest entry, the most speculative, and the one where good analysis has to admit where analysis runs out. A genuine near-term AI catastrophe is very unlikely — bottom of the scale. But the worst case is civilization-scale and genuinely hard to put a ceiling on. That’s an uncomfortable pairing, because the usual method — name the threat, then build the matching response — falls apart when the threat is this fuzzy. The move that feels responsible here, building a special “AI collapse kit,” is actually the mistake. When you can’t say what’s coming, the right answer is resilience that doesn’t need to know.
If you already prep for concrete hazards, this is about doing the abstract one without fooling yourself.
Why this threat is different
For a hurricane or a grid failure, you can describe how it works, put a rough number on the odds, and build a specific response. AI catastrophe fails all three tests:
- You can’t describe one mechanism. “AI catastrophe” covers a huge spread of scenarios with different timelines and effects. There’s no single failure mode to build against.
- The odds aren’t just low — they’re genuinely unknown. Informed people disagree wildly. The biggest survey of the field — the 2023 Expert Survey on Progress in AI, with 2,778 AI researchers — found a median estimate of at least a 5% chance of extinction-or-similar from advanced AI, but the individual answers ran from near-zero to near-certain. The same year, hundreds of leading researchers signed the Center for AI Safety statement calling extinction risk a global priority alongside pandemics and nuclear war. We rate it low for the near term, and that’s defensible — but the honest label is deep uncertainty, not a confident little number.
- The worst case has no clean ceiling. By its nature the tail is severe and hard to bound, which is exactly why you can’t just wave it off.
Here’s what that adds up to: a detailed “AI plan” is a guess dressed up as readiness. Any specific collapse checklist is you inventing particulars nobody can actually predict — it gives you the feeling of being ready without the substance. That’s not preparedness; that’s fortune-telling.
So how do you prepare for the unknown?
RAND developed an approach called Robust Decision Making for exactly these problems: instead of optimizing for the one future you predicted, you look for strategies that work acceptably across the widest range of futures you can’t rule out. You aren’t trying to be right about what happens. You’re trying to be okay no matter what happens. And the way you get there is generality, redundancy, and the ability to adapt — which, conveniently, is the same thing that hedges every other threat you face. The less you can specify the threat, the stronger that argument gets.
The disciplined move is to prepare for the effects such a shock would plausibly cause, using measures that pay off across many threats, so you’re covered whichever way it breaks — or if it never comes at all.
Map the realistic part onto things you already prep for
The far tail is unplannable. But the nearer, more realistic end of AI risk lands squarely on disruptions you already understand:
- Digital and infrastructure disruption. AI could make it easier to knock over important computer systems and critical infrastructure. That’s a grid-and-infrastructure problem — same territory you already cover.
- Fakes, scams, and lost trust. Convincing AI-generated misinformation and fraud could make it hard to trust what you see and hear. That’s a fraud-and-information problem — one you can build habits against today.
- Economic disruption. Fast AI-driven change could destabilize parts of the economy or the job market. That’s an economic-shock problem — the territory of a cash buffer and job-loss readiness.
- The far tail. The science-fiction end resists specific prep entirely, and honesty means saying so out loud. Broad resilience is the only hedge, and even it is partial.
The point: the parts you can actually act on aren’t new categories — they’re bigger versions of shocks already on your list. Prep those well and you’ve covered the realistic surface of this threat while staying honest about the part nobody can cover.
Fix the one thing everything hangs on
Here’s the concrete version, and it’s worth working through. Picture how much of your life runs through a single point. One email account is the password-reset path for your bank, your investments, your photos, your medical records, your work. One cloud login holds the only copy of your documents. If a wave of AI-boosted scams or a bad outage takes down that one thing, it doesn’t take one thing — it takes all of them at once. That’s the fragile setup, and it’s fragile against a dozen threats, not just this one.
So make it less fragile, deliberately:
- Strong, unique passwords and two-factor everywhere. A different password per account means one leaked login is one problem, not twenty. Two-factor means a stolen password alone isn’t enough to get in.
- Offline backups of what matters. Essential documents, records, and photos copied somewhere that doesn’t need a live internet connection or a company staying in business — a drive in a drawer, printed copies in your binder.
- No single account you can’t route around. A second way to reach your money, your contacts, your key records if one service goes dark.
- Default skepticism toward messages and media. Convincing fakes and scams are among the most likely near-term effects. Slow down, verify, and assume a too-perfect message might not be real.
Notice the whole list protects you from ordinary fraud and outages right now, AI or not. That’s what makes it a high-return move regardless of whether this threat ever materializes.
The rest of the plan is the plan you already have
Because this threat is undefined, your general depth isn’t a fallback — it is the primary preparation. Measure it in how long your household could keep running if the systems you depend on went unreliable — not in gear owned. Stored water and no-cook food sized in weeks. A way to get through an extended outage. A cash cushion against an economic shock. Practical skills. And the deepest hedge of all against a future you can’t forecast: adaptability — broadly useful skills, strong local relationships, and the ability to get by without leaning on any single fragile system. Against a threat you can’t specify, being able to adapt to whatever actually shows up beats any specific stockpile, and it compounds your readiness for everything else.
Stay informed, but keep it level. Follow the field enough to update as it moves, and skip both the breathless doom and the dismissive hand-waving. Your estimate here should be provisional — held loosely, revisited as things change — not locked in either direction.
Bottom Line
- AI catastrophe is the most speculative entry — very low near-term odds under deep uncertainty, with a worst case that has no clean ceiling.
- A detailed “AI plan” is a guess in a costume — you can’t prep specifics for an unspecifiable threat, so prep the effects, with measures that pay off across many shocks.
- The realistic part maps onto shocks you already cover — cyber/infrastructure, fraud/information, and economic — so covering those covers the realistic surface.
- Fix the one thing everything hangs on — unique passwords, two-factor, offline backups, and scam skepticism protect you from everyday trouble today, whatever AI does.
- Broad resilience and adaptability are the real hedge — against a threat you can’t forecast, the ability to adapt beats any specific stockpile.
When you can’t know what’s coming, don’t pretend to. Broad resilience isn’t the consolation prize for an unpredictable threat — it’s the correct answer to one.
Prepare for the undefined threat the way that actually works — rate what you face and build the plan that covers the most ground — with Threat Planning.