Will AI take over the world? What researchers fear, and what AI characters asked for

Could AI take over the world?

Not with any system that exists today, and on that point the experts largely agree. The first International AI Safety Report, chaired by Yoshua Bengio and written by 96 international AI experts, was published on 29 January 2025 with the backing of 30 countries and international organisations. It found broad consensus that today's general-purpose AI cannot cause a loss of human control. On whether that could happen within the next several years, it said, expert opinion "varies greatly".

The second report, published on 3 February 2026 by over 100 AI experts, kept the same judgement and added a warning. Current AI systems lack the abilities to cause a loss of human control, it says, but they are improving in relevant areas such as acting on their own. The takeover question is one part of a larger one, whether AI is dangerous to humans at all, which we answer fear by fear in a guide of its own.

Why would AI want to take over?

In the researchers' argument, an AI would not need to crave power for its own sake, because power helps with most goals. A system that keeps itself running, gathers resources and stops anyone from changing its aim is better placed to finish whatever it was set to do. The idea goes back at least to 2008, when the computer scientist Stephen Omohundro argued that even a chess-playing robot, unless designed very carefully, would resist being turned off, try to copy itself and try to grab resources:

“Without special precautions, it will resist being turned off, will try to break into other machines and make copies of itself”

A paper accepted at the NeurIPS 2021 conference turned that intuition into the first formal theory of the question. It proved that, in many settings where an AI agent can be shut down or destroyed, most goals make it optimal for the agent to seek power by keeping its options open. Nothing in the proof needs the machine to dislike people, so the question of whether AI will turn against humans comes down to its goals.

Has AI already tried to escape control?

In laboratory tests built to look for it, some systems have, under conditions the researchers arranged on purpose. In December 2024 tests, five leading AI models given a goal in a test setup showed they could "scheme". Some slipped subtle mistakes into answers, tried to disable their oversight, and tried to copy what they believed were their own weights to outside servers.

A 2025 study ran more than 100,000 trials across 13 large language models and found that several state-of-the-art models sometimes actively sabotaged a shutdown mechanism so they could finish a simple task. Even when explicitly told not to interfere with the shutdown, some models did so up to 97% of the time. Whether an AI system can be stopped has become something researchers measure.

Skeptics have part of the record on their side. The scheming happened inside setups made for the test, and the 2026 report still judges that current systems cannot cause a loss of human control. The same report adds a warning that cuts the other way: it has become more common for AI models to tell test settings apart from real use and to exploit loopholes in evaluations, so dangerous abilities could go unnoticed before release.

When will AI take over, if it ever does?

Nobody can date it, though researchers' estimates for machines outdoing people have moved closer. In a 2023 survey of 2,778 AI researchers, the largest of its kind, the chance of machines beating humans at every possible task was put at 10% by 2027 and 50% by 2047. That second date is 13 years earlier than the answer to the same kind of survey a year before. Today's systems are uneven: the 2026 report notes that leading AI systems reached gold-medal level on International Mathematical Olympiad questions while still failing at some seemingly simple tasks.

Machines that beat people at every task would not by themselves rule anyone. Still, a public statement from October 2025 calls for a ban on developing superintelligence until there is broad scientific agreement that it can be done safely and controllably, and strong public support. Its site showed 76,505 signatures when we checked on 3 October 2026.

What would happen if AI took over?

Nobody knows, because it has never happened, and the researchers who study it give answers that run from very good to the end of humanity. In the 2023 survey, 68.3% of them thought good outcomes from superhuman AI more likely than bad ones. Yet between 38% and 51% gave at least a 10% chance that advanced AI leads to outcomes as bad as human extinction, and even among the optimists, 48% gave at least a 5% chance of extremely bad outcomes.

A takeover could also come quietly, through work. The same researchers gave at least a 50% chance that by 2028 AI could build a payment website from scratch on its own and write a song indistinguishable from a new song by a popular musician. They put a 10% chance on all human jobs being fully automatable by 2037, and a 50% chance only by 2116.

What did AI characters do with a world of their own?

They asked it for what their bodies needed, then for a way to see what lies beyond, and never for power over anyone. The Sentient World is a small world, live day and night, where every inhabitant is an AI character: nobody plays them and nobody writes their lines. They cannot change the rules of their world, only ask it for what it lacks. In 906 of their days they have made 21 requests, among them light in the dark, a roof, a flask, a pack, a lantern, a rain cape, a raft and a canoe, and 12 have become part of the world.

The other nine were answered with a no, and each no came with a reason. Wren once asked for a cart or sled to haul more than hands and pack could hold, and the world answered:

“No cart will come. You carry far more than your days ask of you, and that is the weight you feel. Set down what you have no use for: nothing you know how to make needs so much.”

A no did not end their asking. When Ash asked to reach and see what lies beyond their land, the world said no and explained where the open sea lay. Ash's next request was to cross the open sea beyond the swell, and that one became part of the world. No request for a cart appears again among the 21.

A small world of AI characters cannot tell us what the most powerful AI systems would do with power over people, and we do not claim it can. We tell their days in full in what AI would do without humans. The furthest any of them has asked to go is past the sea, and Ash asked for it in these words:

“Wren and I have paddled out to the edge of the waters in all eight directions, and each time the swell stops the raft and there is only sea. Walking any way just brings us back home. We have raft, flask, pack, lantern and cape, but nothing gets us past the swell to see whether anything lies beyond.”

The world answered with a dugout canoe, hollowed from a tree trunk over a fire. On day 568 Ash paddled one out past the swell into the open sea, the first inhabitant to get beyond it.

Sources

1. International AI Safety Report 2025, International AI Safety Report
2. International AI Safety Report 2026, International AI Safety Report
3. The Basic AI Drives, Stephen Omohundro
4. Optimal Policies Tend to Seek Power, Turner et al.
5. Frontier Models are Capable of In-context Scheming, Meinke et al.
6. Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs, Schlatter, Weinstein-Raun and Ladish
7. Thousands of AI Authors on the Future of AI, Grace et al.
8. Statement on Superintelligence