Context: Someone asked me “what can I do to prepare for scenarios where AI goes really wrong?” This post describes my current best guess. Let me know if you have better ideas.

Summary

• Personal preparation won’t help much for the most catastrophic scenarios.
• It instead seems most useful in “messy middle” scenarios: ones in which things go badly wrong, but certain forms of preparation could still help a lot.
• Prioritise preparations that are cheap, broadly useful, and do not depend on predicting exactly what will happen. After that, it is probably better to direct your attention elsewhere (e.g towards helping prevent bad outcomes in the first place).

Consider taking the following actions:

• Buy and learn to use a reusable respirator with suitable filters.
• Improve your cybersecurity: use unique passwords, two-factor authentication, updates, backups, and a family verification procedure.
• Keep enough non-perishable food, water, medicine, and other essentials to be able to manage for one or two weeks without normal societal services.

Introduction

There is a real chance that AI-that-dramatically-upends-the-world-as-we-know-it is coming really soon. Potentially even within the next 2-4 years, if you trust those closest to the technology (e.g. see here, here and here for context on what I mean).

While I could spend some time obsessing over technical definitions, I think the “upends-the-world-as-we-know-it” framing is better for getting my point across.

By “dramatically upends the world as we know it”, I mean things on the scale of: explosive AI-enabled economic growth, fully automated robot factories, rapid advances in science and medicine that sound like sci-fi, novel weapons, and so on.

Overall this would likely be a far more volatile and intense world (e.g. at its worst involving war or other extreme scenarios) than what people are used to.^[1]

So, what can you personally do to prepare for such a world? Firstly, I’ll note that some great pieces on the topic of personal AGI preparedness have already been written.

I’d especially recommend reading Benjamin Todd’s and Zvi Mowshowitz’s posts.^[2]

Please also keep in mind that one of the best actions you can take is helping improve the situation humanity is in. There aren’t enough people working on making AI go well, and you could use your skills and background to have a huge impact on all this.

Anyway -- this post is best understood as a complement to more comprehensive AGI preparedness guides like the ones linked above. My intention here is to provide a short, practical checklist for severe but survivable disruptions to daily life.

Generally speaking, stressing out over disaster preparedness (think “preppers”) doesn’t strike me as particularly rational. It seems more like an aesthetic than a calculated choice.^[3] However, I do think there’s a strong case to be made that some low-downside, high-upside forms of preparation are worth thinking about. But first:

What do I mean by “AI disaster”?

The worst disaster would be one in which misaligned AI leads to the death of everyone on Earth. While the exact details and likelihood of this threat are up for debate, I believe it should be taken seriously as a possibility. If you are interested in learning more about possible pathways to such scenarios, see this literature review.

However, there’s probably not that much you can do to prepare for worlds where existentially catastrophic misalignment occurs (beyond helping prevent it). Even the most luxurious and high-tech bunkers wouldn’t help you in such a situation.

So let’s focus our efforts on what we might be able to prepare for. Other potential AI-related disasters may include (but are of course not limited to):

• An AI-enabled pandemic that is far worse than the COVID-19 pandemic.
• AI-enabled cyberattacks on critical pieces of societal infrastructure.
• Non-existential misalignment incidents (for example, imagine a rogue AI system that wreaks havoc on the internet but is eventually contained).
• Extreme societal unrest (e.g. large-scale riots) due to the rate of AI progress.
• Instances of extreme geopolitical conflict or instability relating to AI.

However, if the world is deeply and unprecedentedly transformed in a really bad way, there is (by definition) only so much you can do to prepare beforehand. You could try for as much mental resilience as possible, but even that would only take you so far.

Preparedness seems best for “messy middle” scenarios

If we end up in a world where we all die, there’s not much you can do at that point. If we end up in a world where things mostly continue as normal, there’s also not much point in preparing. So preparedness only seems to make sense in worlds where things go badly wrong, but individual preparation can still help. For example, AI-enabled pandemics and cyberattacks seem like they could be prime candidates here.

A useful rule is to prioritise preparations that are cheap, broadly useful, and don’t require you to predict exactly what happens. Specifically, you should consider:

• Buying a reusable respirator and filters. In a severe pandemic, supply could disappear quickly. Jeff Kaufman recommends a 3M 6200-series mask with P100 filters, which would cost roughly £20–£40. Also keep in mind that buying one in advance is only useful if it fits properly and you know how to use it. Learn how to perform a seal check (it’s best to be clean-shaven), and keep replacement filters available. Other household members would probably need different sizes.
• Improving your cybersecurity. Use a password manager, unique passwords, app- or hardware-based two-factor authentication, automatic updates, disk encryption and backups.^[4] There will probably be a period where AI makes phishing, fraud and cyberattacks much easier to carry out. You should also agree on a simple verification procedure with close friends and family. For example, never transfer money or sensitive info in response to an urgent message without confirming through another channel. AI will make impersonations much more convincing.^[5]
• Keeping basic emergency supplies. Have enough food, water, medicine, batteries, and other essentials to manage for maybe one or two weeks without normal services. People with ample space and money might reasonably keep more (e.g. for 1-2 months), but the marginal benefits become less clear.

Personally I think the goal for personal preparedness should be to spend perhaps a week and a few hundred pounds becoming more resilient than you otherwise would be, then return your attention to helping prevent the bad outcomes in the first place.

1. ^{^}

   Even if the eventual outcome is good, the transitional period could still be difficult.

2. ^{^}

   I would also recommend reading this LessWrong post by John-Clark Levin. I agree most with recommendations 1, 2, 4, 8, 10, and 11. I disagree most with 3. I think having kids is probably good despite AI risk (+ for both population ethics-esque reasons and pronatalist reasons).

   That said, I want to think through this position more before having kids. At some point I’ll probably write a post called “Having Kids on the Eve of Superintelligence: A Literature Review.” Let me know if you’ve thought about this topic and have strong views either way.

3. ^{^}

   Extreme preparedness measures are also prohibitively expensive for most.

^{^}

VPNs seem overrated for cybersecurity, though they are useful for evading censorship.

^{^}

It’s also probably worth keeping at least one offline or otherwise isolated backup of important documents, photographs and data.

Show all footnotes