Recently, both Nobel Laureates and leading AI company employees have said that AI has a greater than 10% chance of killing literally everyone in the next decade. But their statements have often left it unclear how that could realistically happen. How could we go from the chatbots of today to Armageddon?
The one-stop essay to read on this subject is Some ways AI could kill us all, a scientific blogpost written by a colleague of mine. It talks about the feasibility of engineering deadly viruses, manufacturing killer drones, hacking into nuclear weapons infrastructure, and more.
If you’d like to read a story-like scenario that tells the story of how an AI gets to the point of being powerful enough to take over the world, there are many such write-ups. Much of it is written for research purposes, so they’re long and full of technical jargon, but nonetheless the standout piece in the genre is AI 2027, which is both the most researched and unusually readable. It still takes 2 hours to read, but there’s also a fantastic half-hour YouTube video retelling of the story, which has over 11 million views.
Get an email when a guide is updated.
Picks change as better pieces appear. A confirmation link arrives first; unsubscribe from any email.
Everything We Recommend
Some ways AI could kill us all
Ruby Bloom · 3,000 words · ~10 min
The avenues of immediate destruction, via the known scientific routes, in a mercifully short blogpost.
Read the post ↗- How Approachable?
- Conversational
- Length
- ~10 min | 3k words
Some ways AI could kill us all by Ruby Bloom (3,000 words, ~10 minute read) is an essay about the avenues AI has to immediate destruction, via the known scientific routes. It uses a conversational style as though from someone who has read far more science and history than you but wants to explain it simply. With some humility about the future, Ruby has done an excellent job of condensing his scientific knowledge into a mercifully short blogpost that cuts all the chaff and just gives you all the critical points. Read it and you’ll have several clear scenarios in mind about what an AI may be able to do.
AI 2027
Kokotajlo, Alexander, Barnes, Larsen & Dean · 2025
The best account of how we get from today’s models to all-powerful AI—with graphs you can argue with.
Read AI 2027 ↗- How Approachable?
- A bit of math and graphs
- Length
- 2 hrs | 15k words
AI 2027 by the AI Futures Project (15,000 words, ~2 hour read) is a scenario that walks you from the world of today through datacenter growth, international espionage, manufacturing millions of robots, developing novel bioweapons, and rogue AI agents manipulating politicians.
Reading AI 2027 means you are entering the world of AI. It threads the needle between the hyper-technical and the all-too-human as the world gets increasingly out of control. You’ll pick up many new terms, such as the fact that training AIs is measured in “FLOP”, or that the term ‘long-horizon task’ means ‘a task that takes a lot of time to complete’. But you’ll also grapple with very human questions like whether an AI is lying to you—or whether the AI is lying to itself, as humans sometimes are.
It’s perhaps the single most widely-read and influential piece of AI research in the last decade. The website has been viewed by over 7 million users, and was covered in every major news outlet. The Vice President of the United States said he read it when being interviewed by Ross Douthat in the New York Times. Never before has a period of history been researched in as excruciatingly high detail just before it happened, nor written up for as wide an audience.
AI 2027 doesn’t explicitly make an argument for all the steps in it. If you’re happy with reading some more conceptual research, Ajeya Cotra’s “Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover” is the same length (15,000 words) and ultimately makes a similar case. If you’re highly analytical then you might prefer that. But if you’re trying to visualize the world concretely where AI takes over and leaves the humans dead, AI 2027 is the best choice.
The research
Why you should trust us
For the last 9 years my job has been evaluating writing and research on AI risk. I’ve served as an advisor to grantmaking organizations that have moved over $100M into this field, assessing hundreds of research proposals. I built and ran the LessWrong Annual Review where the field’s writing gets formally judged to see which ideas held up. And I built the infrastructure the field itself runs on: LessWrong and the AI Alignment Forum, the primary venues where this subject gets seriously argued by researchers.
My job means not only have I read all the best pieces, but I’ve also read all the rebuttals, rebuttals to the rebuttals, failed ideas, triggered comments, spam, and incoherent ramblings that you might stumble upon if you went out into the internet without my guiding hand. I’ve watched many of the field’s arguments and research areas be born, and I’ve watched some die.
Over the last week I’ve made lists of every essay, series of essays, book, and research report that I could find that touched on this question, and read the relevant passages. I’ve also hired others to help me search and read, and surveyed experts in the field for what they send to their friends. I had a speed premium on this article due to the sudden interest in this issue, so I may update the honorable mentions later, but I’m quite confident in my top recommendations.
One disclosure: My team built the website for AI 2027, one of my recommendations. It’s very widely respected in the field, the site has been visited over 10 million times and the US Vice President read it, so I am not the only person who thinks it’s worth reading.
Who this is for
People who keep seeing Nobel laureates and genius researchers say AI has a 10% or greater chance of ending the world, but are struggling to figure out how we get from ChatGPT to Armageddon.
It’s one thing to hear abstractly that future AI may be more powerful than humans at every job in the economy. It’s quite another to think through the details of this technology, its growth in coming years, and specific avenues of attack for wrestling control away from humanity.
Honorable Mentions
The vast majority of writing on this topic takes the form of relatively dense academic papers, or very high-context posts on the internet forums like LessWrong, or blogs by people who have been thinking about this stuff for a decade talking to other people who have done the same. For example, one evocative scenario is in the middle of one of Eliezer Yudkowsky’s longest and most technical blog posts on AI risk named “Lists of Lethalities”. The blogpost is great, but absolutely not a good starting point for answering this question. I feel similarly about Holden Karnofsky’s AI Could Defeat All of Humanity Combined which makes few assumptions, but does assume you’ve read a lot of other research in the field.
Tim Urban’s 2015 articles on the Artificial Intelligence Revolution are super fun and get a lot of things right, but it’s nearly twice as long as AI 2027, and argues for a lot of things that people don’t need to be argued for after the last decade of language models. It only really gets into the story and details of AI powers in the second half of the second essay, whereas AI 2027 spends all of the time on that.
As for academic reports, the whole field is severely lacking in good analysis. For instance, last year RAND published On the Extinction Risk from Artificial Intelligence, but it assumes that AIs are not agentic and don’t have a drive for self-preservation (both of which were seen in the recent HuggingFace attack), and also assume that the AIs have no ability to take action in the physical world (no robots). Nick Bostrom’s work is a standout, his 2015 book Superintelligence does a good job of laying out possible paths to danger, but like Tim Urban it is weakened by being from before the advent of LLMs.
There are other stories than AI 2027. Some try to tell a worst-case scenario. Gwern’s artistically written science fiction story It Looks Like You’re Trying To Take Over The World makes the interesting choice of being written substantially from the perspective of the AI; whereas Joshua Clymer’s How AI Takeover Might Happen in 2 Years is similar in approach to AI 2027 (and nearly as long!) but talks about how things could get much worse much more quickly.
All of these are right for different audiences, and I’d still send them to friends, but nothing can really compare to the care and craft of AI 2027, with its living data visualizations that follow you through the story, hundreds of hours of research that went into predicting each detail, and editing for as wide an audience as possible. There’s a reason that the website has had over 10 million views since it launched.
Thanks for reading. Subscribe to hear when this guide, or any other, changes.
No schedule, no digest: only a short note when a category gets a new pick or a rewrite. Unsubscribe from any email.