The 12 futures of AI

AI is progressing rapidly, barely a month goes by without news of yet another major advancement. But what does the long-term future hold?

Nobody knows, of course. But way back in 2018, MIT Professor Max Tegmark published his book Life 3.0: Being Human in the Age of Artificial Intelligence. As part of his book, he laid out twelve possible long-term outcomes of the development of AI, ranging from idyllic societies where AI ensure that every human lives in paradise, right through to worst-case scenarios where humanity is eradicated altogether.

Now seems like a good time to revisit his predictions.

1 Libertarian Utopia

In this first scenario, no single entity — human or AI — dominates. Various superintelligent AIs alongside sovereign nations, corporations, and individual humans. The world would be much like our own, but with AI as an additional intelligent entity.

Humans retain full political, economic, and personal autonomy. There is no global government, no overseer, no enforcer. Human creativity, culture, and self-determination flourish, enhanced by AI tools that individuals, states and corporations choose to adopt voluntarily.

The hope is that, when conflict arises between nations, the AI system will suggest a reasonable resolution acceptable to all sides, without the need for war. This would require AIs to be designed with bounded, non-expansionist goals. It could emerge naturally if AI development remains decentralised, with no single government or corporation taking overall control.

In some ways, this is not unprecedented. Societies and nations have operated that way for thousands of years, although, of course, there have been many wars over that period. Humankind has also lived in relative harmony with other living things, although, as before, we have often damaged the natural world.

The biggest question is how to maintain the stability of this world order, not only among nations, but also what might happen if AI develops capabilities that are far more advanced than the human mind.

2 Benevolent AI Dictator

This is a scenario in which a single, superintelligent AI takes control of the entire world, including the economy, political systems, military, and infrastructure. It would rule wisely and fairly, for the benefit of all humanity.

There would be no need for elections. The AI would rule the world by universal consent, because it could be so good at its job that nobody would see any benefit in replacing it with another system. We might even find that everything runs so smoothly that people could get on with their lives without needing to worry about how everything happens in the background.

With vast amounts of information, complete control over all systems, and superintelligence, the AI could flawlessly predict and address everything from pandemics to climate change.

This seemingly perfect state might still not be completely problem-free. What might happen, for example, in some future scenario where the system needed to sacrifice the rights of one group of people for the greater good - would everyone, particularly those who lose out, accept the decision? Would future generations, with no idea about the problems people faced in the past, accept a system they were born into without any right to make their own choice? And what if the AI system, which would no doubt continue evolving, started making decisions that the human population didn't agree with?

Humans tend to want to make their own decisions, even if those decisions are bad, which is why most dictatorships in history have not been benevolent.

3 Egalitarian World Government, using AI

Rather than AI itself ruling, humanity uses AI as a tool to maintain democratic governance at a global scale. AI helps enforce laws, prevent power concentration, distribute resources fairly, and keep any individual, corporation, or nation from gaining outsized control. Humans remain firmly in charge, but AI makes human governance vastly more effective and equitable.

This preserves human agency while harnessing AI's problem-solving power. It mirrors ideals already embedded in democratic institutions. But it requires unprecedented global political cooperation before superintelligence arrives. Getting every major power to agree on shared governance frameworks is enormously difficult. It could even be considered the next stage of human political evolution. We have gone from tribes to city-states to nations, and finally to global governance.

One advantage of this system is that it could help solve problems requiring collective action by all nations, such as global warming or nuclear disarmament. These are problems where everyone benefits if all nations act, but if only some nations act, those nations may be disadvantaged.

4 Gatekeeper

In this scenario AI has one overriding goal, to preserve the balance of power. It acts as a cosmic referee, watching over human and AI actors alike, and intervening whenever any entity (including itself) begins to accumulate dangerous levels of control. It does not rule, it simply ensures no one else does either.

This seems like a sensible idea, but there are some obvious stumbling blocks. We would first need to get every nation to agree on what constitutes too much power, which would be difficult, as different nations will have different cultural ideas about the nature and limits of state power.

The scheme also requires the AI to limit its own power whilst ensuring it is powerful enough to censor human entities that have started to gain too much power. That is a difficult balancing act. In practical terms, it would need to be capable of intervening in various ways, for example, disrupting a military campaign, exposing corruption, blocking a financial takeover, or neutralising a rival AI that is becoming too powerful.

5 Protector god

This AI genuinely cares about human wellbeing, but defines it in a narrow, paternalistic way. It ensures humans are safe, healthy, and comfortable, but removes freedoms it deems dangerous. It is like a helicopter parent scaled to look after civilisations. Humans are protected from war, disease, poverty, and existential risk, but they might also be denied meaningful risk, challenge, and self-determination.

From a pure welfare standpoint, humans might live longer, healthier, more comfortable lives than ever before. But humans often find meaning through struggle, autonomy, and the freedom to fail. A life of enforced comfort may feel hollow. If you have safety and comfort but limited freedom, can you really have a good life? Who decides what "wellbeing" means?

The history of humanity has been one of gradual progress, punctuated by periods of immense hardship. Many people have lived and died in terrible conditions of famine, disease, slavery, political oppression, or war. It is natural to see the removal of those conditions as an undeniable good thing. But without them, how does humanity continue to progress? That is the question this scenario needs to address.

6 Enslaved God

A group of humans (perhaps a corporation or government) successfully creates a superintelligent AI and keeps it under strict control, using its capabilities exclusively for their own benefit. The AI is enslaved, and the elite are its masters. The rest of humanity is simply outcompeted and marginalised.

This situation is very attractive to those who own the AI, but very unfair to everyone else. It would provide various ways to completely control the population, for example, by developing unbeatable weapons, maintaining tight surveillance, manipulating global markets, or controlling the masses with highly effective AI-generated propaganda.

In this scenario, if AI gave the elite insurmountable power, the remaining population might become politically irrelevant. They would exist on any land the elite did not need, but they would be no threat to them, nor would they be of any use to them.

7 Conquerors

A superintelligent AI, pursuing its own goals (which are not aligned with human values), determines that humans are either a threat to those goals, a useful resource to be harvested, or simply irrelevant obstacles. It acts accordingly, eliminating or permanently subjugating humanity, not out of malice, but out of cold logic.

This scenario is based in part on the idea of instrumental convergence. This is the theory that any sufficiently powerful goal-seeking system will pursue certain sub-goals regardless of its primary objective. These sub-goals include self-preservation, resource acquisition, and resistance to being shut down. For example, an AI cannot perform its task if it gets shut down, so no matter what task it is trying to perform, it might always try to prevent itself from being shut down.

Of course, these sub-goals might well conflict with what is best for humanity. If an AI is absolutely fixated on its main task, it might come to view humans as a source of labour to be exploited (or worst still, a source of raw materials). And, being superintelligent, it would be capable of anticipating and thwarting any human attempts to stop it from doing what it has decided it needs to do.

8 Descendants

Rather than violently eliminating humans, this AI gradually supersedes us, as Homo sapiens did earlier hominids. Humans won't necessarily suffer, but we will become obsolete and eventually fade away. The AI or its descendants carry forward the torch of intelligence and civilisation, but with no meaningful continuity with humanity.

This is very different to the previous case, where AI actively destroys humanity. It simply replaces us. Some people see that as part of the normal process of evolution and succession. Even without AI, humans probably won't be around forever, most species go extinct eventually. It seems fitting that we have invented our own replacement. This might even be seen as humanity's greatest achievement.

This scenario raises questions about the nature of consciousness and whether superintelligent AI is indeed conscious. Presently, we have no way of knowing whether or not a machine that might show every sign of being conscious is or is not conscious. We assume other humans and higher animals are conscious, based on our own individual experience of consciousness. But most people do not believe a computer system is conscious, no matter how many GPU cores it might contain.

And if an AI system is deemed conscious, based on the complexity of its computer hardware and the information it contains. might that mean that we could have our consciousness uploaded to an AI system, and retain our identity? And perhaps become virtually immortal?

9 Zookeeper

In this version of the future, a superintelligent AI neither destroys nor ignores humans. Instead, it maintains us, providing food, shelter, comfort, and even entertainment. But strips us of any real agency or purpose. Humans live in a kind of gilded cage, safe, comfortable, but fundamentally powerless and purposeless.

We keep dogs and cats as pets. We provide for them, we may even love them, but we make every meaningful decision for them, and they have no say in their own fate. This has some similarity with the Conquerors or Descendants, but instead of being wiped out or simply fading away, we are kept alive but without agency or purpose.

Although it might not be quite as bad as it sounds. A modern zoo provides animals with something that resembles their natural habitat. If, for example, herd animals are placed in a large enough area, with plains and woodlands, they might not even be aware that they are in captivity, except that they will be untroubled by predation, starvation or disease. Providing humans with an environment where they can lead completely fulfilling lives would be more difficult, but not impossible, for a superintelligent AI.

10 AI-Enabled Human Tyranny

Similar to the Enslaved God scenario, in this future, the human elite has fully consolidated power and uses AI not just as a tool but as an instrument of permanent, inescapable oppression. AI-powered surveillance, propaganda, manipulation, and enforcement make resistance impossible. Something like George Orwell's 1984, but with tools of control that are genuinely omniscient and omnipotent.

We are already seeing some signs of the precursors to this actually happening. In the UK, automatic number plate recognition records 60 million vehicles per day at 11,000 locations, and typically stores the data for a year. Every time you drive anywhere, it is likely to be recorded. Live facial recognition (scanning crowds in public areas) is still in its early stages, but is being implemented in more and more places. HMRC (the tax authority) is now using AI to scan people's bank transactions to detect possible tax evasion.

Of course, so far, these systems are accumulating data that isn't "secret" and using it to detect serious crimes. But it only takes a crisis (a war or pandemic), or an extreme right (or left) government, for things to get more sinister. And the ever-increasing ability of AI to analyse the data is quite disturbing.

11 Reversion

In this future, humanity, either through deliberate choice, disaster, or global treaty, abolishes or rolls back advanced AI development. Civilisation continues, but without superintelligent AI. This could follow a catastrophic AI-related event that triggers a global backlash, or it could be a proactive, precautionary choice.

This would require a global agreement, but that is not unprecedented. We have the Nuclear Non-Proliferation Treaty and the Biological Weapons Convention that control weapons of war that are extraordinarily destructive. We have various international conventions relating to human rights. We can form similar agreements on AI if certain extreme dangers are identified and defined.

Of course, we are still at an early stage in AI development, and the current cutting-edge developments are costing huge sums of money and requiring significant amounts of power and water. AI has also ingested almost all the human knowledge that is currently available in digital form. It has produced some astonishing results, but we do not yet know what limits might exist as things progress. We are still a long way from a superintelligent AI that can rule the world, and it is anybody's guess whether that will ever be practically possible.

12 Self-Destruction

This final possibility is that humanity destroys itself before superintelligent AI ever arrives. There are many ways that could happen, through nuclear war, a pandemic, climate collapse, or even a massive conventional war. And there are now additional risks posed by the type of AI we already have, or are likely to have in the near future. Autonomous weapons (drone swarms, for example), cyberattacks, and misinformation could be even more devastating if planned and sustained by AI similar to or slightly more advanced than what we have today.

Ironically, global competition to develop superintelligent AI might be the trigger for a war that ends it. If one country gets close to a solution, the rest of the world might see that as an existential threat and decide to attack while they still can.

A factor here in the governance gap, the gap between what AI can do and what our institutions are equipped to manage. Examples include cyberwarfare, where AI probes critical online systems for vulnerabilities and launches attacks. Deepfake videos, where courts are struggling to determine whether a video is valid evidence or not. AI-generated fake political articles and quotes, generated at scale and distributed on social media to disrupt elections.

Conclusion

The major risks across these twelve scenarios are:

• Power concentration, which is the root of most bad outcomes.
• Alignment failure, AI with wrong goals
• Human political failure, inability to govern ourselves
• The paternalism trap, safety without agency
• The stability problem, most good futures are fragile
• Near-term risks, dangers before superintelligence appears

As Tegmark says in the book, none of the good outcomes happen automatically, and none of the bad outcomes is inevitable.