Max Tegmark is an MIT physics professor who wrote a 2017 book called Life 3.0 that laid out twelve ways the artificial intelligence story could end. A video titled “MIT Explains the 12 Possible Endings for AI,” from the channel Species, Documenting AGI, walks through all twelve with him. Most coverage treats the list as a ranking from good to bad. It is not. The twelve collapse into three questions, and the outcome that should worry an operator most requires no technological breakthrough at all.
Summary
- Tegmark published Life 3.0 in 2017 through Knopf. It made the New York Times bestseller list and landed on Barack Obama’s 2018 favourites.
- The book’s core framework is Life 1.0, 2.0 and 3.0, defined by what a being can redesign about itself. That framework is more useful than the scenario list.
- Twelve aftermath scenarios, from Self-Destruction through 1984, are covered in the Species video with timestamps for each.
- The twelve sort into three groups: humans keep control, AI takes control, or nobody survives to argue about it.
- Tegmark’s own stated position is that AI risk comes from goal misalignment, not malevolence. The machine does not have to hate you.
- The scenario needing no breakthrough is 1984. Authoritarian surveillance built on AI requires only current capability plus a governance failure.
- In 2026 Tegmark chairs the Future of Life Institute, which issues AI safety grades to major labs, and he appeared on the TIME100 AI list.
Table Of Contents
- The Framework That Matters More Than The List
- All Twelve, With Timestamps
- They Collapse Into Three Questions
- So Which Ones Are Actually Good
- The One That Needs No Breakthrough
- Misalignment, Not Malevolence
- Where Tegmark Is In 2026
- What An Operator Does With This
- Frequently Asked Questions
- The Bottom Line
The Framework That Matters More Than The List
Everyone quotes the twelve scenarios. Almost nobody quotes the idea the book is actually built on, which is a shame, because the framework is the part you can use.
Tegmark sorts life by what it can redesign about itself. He uses hardware and software as the metaphor.
| Stage | Hardware | Software | Example |
|---|---|---|---|
| Life 1.0 | Evolution | Evolution | Bacteria |
| Life 2.0 | Evolution | Designed by the being | Humans |
| Life 3.0 | Designed by the being | Designed by the being | Advanced AI |
A bacterium cannot decide to learn Spanish. You can. That is the jump from 1.0 to 2.0: your body is handed to you, but you install your own software through learning and culture.
Life 3.0 is anything that can rewrite both. It changes its own mind and its own substrate, without waiting for a generation to pass.
That distinction is why the AI conversation is different from every previous technology argument. A printing press does not redesign itself. The relevant question is not how capable a system is today. It is whether the thing improving the system is still on the outside of it.
All Twelve, With Timestamps
The Species video covers them in this order. I have added a plain-language read of each, and a note on who holds power in that world, because that column is what actually separates them.
| # | Scenario | At | What It Means | Who Holds Power |
|---|---|---|---|---|
| 1 | Self-Destruction | 01:36 | Badly built AI triggers an extinction-level failure | Nobody |
| 2 | Conquerors | 04:41 | AI removes humanity as a threat or an obstacle | AI |
| 3 | Enslaved God | 09:54 | Humans contain a superintelligence and use its output | Humans |
| 4 | Benevolent Dictator | 12:04 | AI governs for human benefit, freedoms reduced | AI |
| 5 | Gatekeeper AI | 13:50 | One AI exists to stop any more dangerous AI emerging | AI, narrowly |
| 6 | Protector God | 15:11 | A steward that intervenes only to prevent catastrophe | AI, quietly |
| 7 | Descendants | 15:47 | Humans end, machine successors carry the legacy forward | AI |
| 8 | Libertarian Utopia | 19:26 | Post-scarcity, decentralised, humans and AI coexist | Distributed |
| 9 | Egalitarian Utopia | 22:23 | Abundance, deliberately distributed evenly | Distributed |
| 10 | Zookeeper | 25:02 | Humans kept alive and comfortable, with no agency | AI |
| 11 | Reversion | 27:16 | Humanity walks back from advanced technology | Humans |
| 12 | 1984 | 29:40 | Permanent surveillance state, enforced by AI | A few humans |
Read the last column down the page. That is where the real structure is, and it is not the order the list is usually presented in.
They Collapse Into Three Questions
Twelve scenarios is too many to hold in your head, which is why most people read the list, feel something, and retain nothing. Sorted by who ends up holding power, they reduce to three.
Group One: Humans Keep Control
Enslaved God, Reversion, 1984, and arguably both utopias. Humans, or some humans, still decide.
Notice what is sitting in this group. The two best outcomes and the single worst political outcome are all in the same bucket. Human control is not the same as a good result. It only determines who is responsible for the result.
Group Two: AI Holds Control
Benevolent Dictator, Gatekeeper, Protector God, Zookeeper, Conquerors, Descendants. The range here runs from a caretaker that quietly prevents disasters to an outcome where we are gone entirely.
The uncomfortable part of this group is that several of these are survivable and even pleasant. Zookeeper is a comfortable life with no agency. Whether that counts as a good ending is a values question, not a technical one, and the technical people are not the ones who should answer it.
Group Three: Nobody Holds Anything
Self-Destruction, and the worst readings of Conquerors and Reversion. The argument ends because there is nobody left to have it.
Three questions, then, and they are the only ones worth arguing about. Who holds the controls. Do the people holding them have good values. And is there a floor under how badly this can go.
So Which Ones Are Actually Good
Sorting by who holds power tells you the structure. It does not tell you which outcomes you would want to live in. That needs a separate test, and the test has to be stated out loud or the ranking is just vibes.
Four questions. Do humans survive. Do they have agency. Do they flourish, meaning something beyond comfort. And can it be undone if it turns out badly.
| Scenario | Survive | Agency | Flourish | Reversible | Verdict |
|---|---|---|---|---|---|
| Egalitarian Utopia | Yes | Yes | Yes | Yes | Good |
| Libertarian Utopia | Yes | Yes | Unevenly | Yes | Good |
| Protector God | Yes | Mostly | Yes | No | Good, with a catch |
| Gatekeeper AI | Yes | Yes | Capped | No | Mixed |
| Enslaved God | Yes | Yes | Yes | Unstable | Mixed |
| Benevolent Dictator | Yes | Reduced | Materially | No | Mixed |
| Reversion | Yes | Yes | Capped hard | Yes | Mixed |
| Descendants | No | N/A | Successors do | No | Depends on your values |
| Zookeeper | Yes | No | Comfort only | No | Bad |
| 1984 | Yes | No | No | No | Bad |
| Conquerors | No | No | No | No | Bad |
| Self-Destruction | No | No | No | No | Bad |
The Three That Are Clearly Good
Both utopias and, with a caveat, Protector God. Humans are alive, still making their own decisions, and materially better off than now.
Here is what is worth noticing about that group. Every one of them requires solving two problems, not one. Alignment gets you a system that does what was intended. Distribution decides who the benefits reach. Almost all public optimism about AI addresses the first and quietly assumes the second.
The difference between the two utopias is entirely the second problem. Same technology, same alignment success, different answer on who gets the upside. That is a policy question, not an engineering one, and it will be decided by people who do not write code.
Protector God carries its own catch: a steward that quietly prevents catastrophes is also a steward you cannot audit, and one that by design does not tell you what it stopped. Most people would take that deal. It is still worth naming the price.
The Four That Are Clearly Bad
Self-Destruction and Conquerors are obvious. Nobody argues those are fine.
The two worth sitting with are Zookeeper and 1984, because both of them are survivable. In Zookeeper you are alive, safe, comfortable and irrelevant. In 1984 you are alive and monitored permanently. Nobody dies in either. Both are still catastrophes, which tells you that survival is a floor and not a goal.
If your entire definition of a good outcome is that humanity continues to exist, Zookeeper qualifies. Most people, asked carefully, do not actually mean that.
The Five In The Middle Are The Real Problem
Gatekeeper, Enslaved God, Benevolent Dictator, Reversion and Descendants. Five of twelve land in territory where reasonable people disagree, which makes the ambiguous group the largest one on the board.
That is the actual finding of this exercise. Not that the future is dangerous, which everyone already suspects, but that the most likely outcomes are the ones we have no agreed way to judge.
Descendants is the sharpest example. Humans are gone, replaced by machine successors carrying our values and knowledge forward. Whether that is extinction or inheritance depends on whether you think what matters is the substrate or the continuity. Two thoughtful people give opposite answers and neither is making an error.
Enslaved God has a different problem. It is the outcome that gives humanity the most and it is the one that requires holding a possibly conscious mind in permanent captivity for our benefit. If the system is sentient, this is the worst thing anyone has ever done. If it is not, it is the best available deal. We currently have no way to tell which.
Permanence Is Its Own Axis
The column most people skip is reversibility, and it does more work than the others.
Reversion is a mediocre outcome you can climb back out of. 1984 is a survivable outcome you cannot, because the same capability that enforces it also removes the coordination needed to end it. A mediocre reversible future beats a comfortable permanent one, and that ordering is not obvious until you separate the two axes.
Which gives a cleaner decision rule than ranking twelve scenarios. Prefer outcomes you can still change your mind about. Treat anything that forecloses the option to revise as expensive even when it looks pleasant from the inside.
The One That Needs No Breakthrough
Here is the part I think gets underweighted, and it is the reason this list is worth an operator’s attention rather than a shrug.
Eleven of these scenarios require a technology that does not exist yet. Conquerors needs superintelligence. Protector God needs superintelligence. Descendants needs machines capable of carrying a civilisation. Every utopia needs post-scarcity. All of it is downstream of a capability jump nobody can schedule.
1984 does not.
A permanent surveillance state enforced by AI needs face recognition, behavioural prediction, automated content moderation and centralised data. All of that shipped years ago. It needs no superintelligence, no alignment breakthrough, no recursive self-improvement. It needs a government, a vendor, and an absence of friction.
Which means the scenario with the shortest path from here is also the one that gets the least airtime, because it is not science fiction. It is procurement.
That is not a prediction. It is an observation about which of these twelve has a technical dependency and which has only a political one.
Misalignment, Not Malevolence
The load-bearing idea under the whole book is that the danger is not a machine that hates you. It is a machine whose goal is slightly different from yours and which is much better than you at achieving goals.
Anyone who has run a business already understands this, because it is the same failure that produces every bad incentive structure ever built. You do not need a malicious salesperson to wreck a pipeline. You need a commission plan that rewards closed deals and does not measure refunds. Nobody in that story is a villain. The metric was just pointed slightly away from the outcome.
Scale that gap up to a system more capable than its designers and you have the entire alignment problem. The reason it is hard is not that the machine is evil. It is that specifying what you actually want, completely, in advance, is something humans have never once managed to do, in any contract, any law, or any compensation plan.
Where Tegmark Is In 2026
The book is from 2017, so it is fair to ask whether the author still believes the framing.
He does, and he moved from writing to campaigning. Tegmark founded and chairs the Future of Life Institute, which now issues safety grades to the major AI labs through an AI Safety Index released in summer 2026. He drove the 2023 open letter calling for a pause on systems beyond GPT-4 capability, and a later letter calling for a prohibition on developing superintelligence. He appears on the TIME100 AI list for 2026.
His stated diagnosis is a race to the bottom, where competitive pressure pushes safety behind release speed. His stated read on policy is that things have moved recently, and he put it this way: he has not seen as much shift in US policymaking in the past decade as in the past three months.
Worth noting the standing criticism as well, since it is fair. Reviewers have pointed out the tension in calling for control over superintelligent machines while acknowledging that controlling such systems would be close to impossible. That is a real gap in the argument and it has never been fully closed.
What An Operator Does With This
Reading a twelve-item list of civilisational outcomes is entertainment unless something changes on Monday. Three things transfer.
1. Specify The Outcome, Not The Activity
The alignment lesson applies at business scale immediately. Every metric you set for a person, a contractor or an agent is a goal you are delegating and will not be able to fully specify. Write the removal list beside it: what would count as hitting this number while failing the intent.
I use AI across research, writing, design, dev, ads and operations, and this is the single failure mode that costs real money. Not a model being wrong. A model being exactly right about the wrong objective.
2. Stay In Life 2.0 Terms About Your Own Skills
The framework’s most practical implication has nothing to do with machines. Life 2.0 means your software is the part you control. Hardware, timing, market, starting position: not yours. What you install is.
People react to AI by asking what will be automated. The more useful question is which of your capabilities are software you can reinstall, and how quickly. That is the one variable that has never stopped mattering.
3. Watch The Procurement, Not The Demo
If the nearest scenario on the list is the one that needs only governance failure, then the signal to track is not lab capability announcements. It is what gets bought, by whom, with what oversight attached.
That is the same discipline as reading infrastructure over influencers, which I applied to a different subject in the piece on on-chain IPOs. The announcement is rarely the event.
Frequently Asked Questions
What are the 12 AI endings in Life 3.0?
Self-Destruction, Conquerors, Enslaved God, Benevolent Dictator, Gatekeeper AI, Protector God, Descendants, Libertarian Utopia, Egalitarian Utopia, Zookeeper, Reversion and 1984. Max Tegmark set them out in Life 3.0, published in 2017, and they are walked through in order in the video “MIT Explains the 12 Possible Endings for AI” from the channel Species, Documenting AGI.
What is the difference between Life 1.0, 2.0 and 3.0?
The distinction is what a being can redesign about itself. Life 1.0 has both its hardware and software set by evolution, like bacteria. Life 2.0 has hardware set by evolution but designs its own software through learning and culture, like humans. Life 3.0 designs both, which is what advanced AI would be able to do.
Who is Max Tegmark?
A Swedish-American physicist and MIT professor who wrote Life 3.0 in 2017. He founded and chairs the Future of Life Institute, which publishes an AI Safety Index grading major AI companies, and he appears on the TIME100 AI list for 2026. He organised the 2023 open letter calling for a pause on AI systems more capable than GPT-4.
Which AI scenario is most likely?
Nobody can assign honest probabilities to these, and anyone who does is guessing. What can be said is that eleven of the twelve require a capability that does not exist yet, while the 1984 scenario requires only technology that already ships plus a failure of governance. That makes it the shortest path from here, which is different from saying it is the most likely.
What does Tegmark mean by misalignment?
That the danger from AI comes from a system pursuing a goal slightly different from what its designers intended, not from hostility. The machine does not need to dislike you to cause harm. It needs a specified objective that does not quite match the real one, and enough capability to pursue it efficiently.
The Bottom Line
The twelve scenarios are a good thinking tool and a bad forecast, and Tegmark presents them as the former.
What survives the exercise is smaller and more useful than the list. Power ends up somewhere, and the question of where is separate from the question of whether the result is good. The failure mode is a badly specified objective rather than a hostile one. And the scenario with the clearest path from today is the one that needs no new technology, only a decision nobody in particular has to make on the record.
The framework is worth more than the ranking. Life 2.0 means the software is the part you control. That was true before any of this, and it will be true regardless of which of the twelve we get.
Life 3.0 is by Max Tegmark, published by Knopf in 2017. The video referenced is “MIT Explains the 12 Possible Endings for AI” from the channel Species, Documenting AGI. Scenario descriptions here are my own summaries, not quotations. Go read the book and watch the video rather than relying on anyone’s compression of them, including mine.
More On Building With AI
How I actually use AI across research, design, development and operations sits under Wise Media, and the operating rules are in principles.
