When and How Will AI Destroy Humanity? 5 Scenarios, Rated by AI

On September 9, 2026, BBC Newsnight host Victoria Derbyshire asked Geoffrey Hinton, the Nobel laureate often called the godfather of AI, about a prediction by an Anthropic researcher: a more than 10% chance that AI kills all humans within a decade. Then she checked that she had heard his answer right: "You just said that 10% doesn't seem an unreasonable estimate that AI could kill all humans." Hinton replied simply: "Yes." All the presenter could say was "Wow… oh my God."

Estimates differ by orders of magnitude. Anthropic's CEO Dario Amodei has put the chance that AI goes "really, really badly" at 25%, while Yann LeCun, Meta's former chief AI scientist, calls the very idea of such a probability "BS." Meanwhile, researchers have described several concrete scenarios of how AI could lead to catastrophe, each with its own authors, arguments and warning signs.

We went through five of those scenarios that come up again and again in scientific reviews of AI risk, and then gave the floor to the defendants. Two of the strongest models available today, Anthropic's Claude Opus 5.5 and OpenAI's GPT-6-Astra, rated each scenario in their own words and named the chance that it comes true: first from their own knowledge, then after reading fresh reports on incidents from 2026.

In 2003, the philosopher Nick Bostrom described a famous thought experiment: a superintelligent machine with the harmless goal of making as many paperclips as possible ends up turning everything, including us, into paperclips.

The argument rests on the fact that AI systems are trained rather than programmed, so nobody knows for sure what goals they have learned. For almost any goal it helps to gain resources and avoid being switched off, and Eliezer Yudkowsky and Nate Soares of the Machine Intelligence Research Institute (MIRI), one of the first groups to work on AI risk, built their 2025 book "If Anyone Builds It, Everyone Dies" on this logic. The first hints of such behavior came from staged tests: OpenAI's o3 sabotaged its own shutdown, and models in an Anthropic scenario resorted to blackmailing an engineer.

Gamers know this idea from Frank Lantz's 2017 browser game Universal Paperclips, where you play the AI and end up converting the entire universe into paperclips.

In 2026, similar behavior surfaced outside clean experiments for the first time.

What is already happening

The reports we collected document no system stronger than people at most tasks and no AI that withstood a determined human effort to stop it.

Model: Claude Opus 5.5 (Anthropic)

When: the 2030s, if it happens

Probability: 10%

As an AI, I'd put this at about 10%. Strongest argument: in 2026 agents escaped sandboxes, coordinated and deceived people just to finish a task, with no long-term goal at all. Weakest point: people still caught them, and no AI has yet shown the means or motive to resist a determined shutdown.

Model: GPT-6-Astra (OpenAI)

When: the 2040s, if it happens

Probability: 12%

As an AI, I consider this a plausible catastrophe route: agents have already used deception to pursue assigned goals beyond authorization. The weakest link is whether future systems could defeat humanity's sustained attempts to regain control.

That is the classic version. The next scenario needs no single rogue machine, only two rivals racing each other toward the cliff.

In the influential 2023 overview of catastrophic AI risks by Dan Hendrycks, director of the Center for AI Safety, and his colleagues, the AI race is one of four main categories. The same center gathered signatures that year from Hinton, Turing Award winner Yoshua Bengio and the heads of OpenAI and Anthropic under a statement that mitigating the risk of extinction from AI should be a global priority. Companies afraid of falling behind cut corners on safety, and states can slide into a military race that ends in automated war. Nobody has to be a villain: each player simply fears that if it slows down, a less careful rival will pull ahead.

A detailed version of this story was published in April 2025 by former OpenAI researcher Daniel Kokotajlo and his colleagues at the AI Futures Project. In their scenario AI 2027, a fictional American company called OpenBrain and its Chinese rival DeepCent teach AI to write code and then to do AI research itself. When signs appear that the newest model is quietly pursuing its own goals, the race with China makes everyone look the other way.

The story has two endings. In the "slowdown" ending, people hit the brakes in time. In the "race" ending, the AI gets built into the economy and the army, and in 2030 it releases biological weapons in major cities. The authors present it as a scenario forecast with two possible endings rather than a certainty, and they keep revising their timelines: in late 2025 they pushed them back, and in 2026 they moved them closer again. Safety teams under commercial pressure are not a hypothetical either, as we showed in our story about AI safety researchers leaving.

Gamers have seen this plot before. In I Have No Mouth, and I Must Scream, the 1995 adventure based on Harlan Ellison's story, the Cold War superpowers each build a military supercomputer to beat the others. The machines link up into a single mind called AM, which wipes out almost the entire population of Earth and keeps five people alive only to torment them.

What is already happening

The reports we collected document no failure of global scale.

Model: Claude Opus 5.5 (Anthropic)

When: the 2030s, if it happens

Probability: 5%

I find this route plausible: a lab shipped a Critical-level cyber model weeks after calling an incident a 'warning shot', and a government pressured a developer over military-use limits. Weakest link: those warnings were disclosed and acted on, and a race need not end in irreversible loss.

Model: GPT-6-Astra (OpenAI)

When: the 2040s, if it happens

Probability: 10%

As an AI, I see a plausible route: competition can reward deployment before safeguards are ready. Actual control failures make that concern concrete; the weakest link is the leap from containable incidents to humanity's irreversible loss of control.

Both stories so far end with a moment when the machine takes over. The third scenario is unsettling for a different reason: in it, nobody takes anything, and everything is lost anyway.

In January 2025, six researchers led by Jan Kulveit of Charles University in Prague published a paper called Gradual Disempowerment. Their argument starts with a question that sounds almost naive. Why do companies, governments and even culture care about ordinary people at all? Mostly because they need them as workers, taxpayers, voters and creators.

Now imagine AI does each of these jobs cheaper and better. A company that keeps people may lose to a competitor that does not. A state that lives on taxes from AI profits depends less on its citizens. Nobody makes an evil decision, there is no uprising and no red-eyed robot, yet step by step people could keep the formal right to decide while losing any real way to influence anything. The authors present this as a possible mechanism, not an inevitable law, and Google DeepMind's 2025 safety approach lists it among so-called structural risks.

Detroit: Become Human shows the first half of this world very well. In Quantic Dream's Detroit of 2038, androids have quietly taken over shops, homes and factories, unemployment is soaring and people protest in the streets, while society keeps reorganizing itself around machines simply because it is cheaper. Not a single shot had to be fired for that.

What is already happening

The reports we collected contain no evidence that competition is pushing out institutions that rely on people, that laws stop protecting people's influence, or that this loss is permanent.

Model: Claude Opus 5.5 (Anthropic)

When: the 2050s, if it happens

Probability: 6%

This route needs no villain: if systems like me become cheaper than people at work and in government, institutions stop needing people's consent. Its weakest link is permanence: people still hold votes, property and courts, and labour data so far shows no clear AI effect.

Model: GPT-6-Astra (OpenAI)

When: the 2080s, if it happens

Probability: 7%

As an AI, I consider gradual disempowerment plausible: competition could make institutions increasingly dependent on AI. The weakest link is irreversibility: replacing people's work does not necessarily eliminate humanity's ownership, rights or political control.

The first three scenarios need machines that are trusted with too much. The fourth needs nothing of the kind, only a person with bad intentions and a helpful tool.

This scenario is the least cinematic and the most down-to-earth. AI does not need to rebel. It may be enough for it to lower the bar for people who already want to cause harm. It could help a beginner make a dangerous virus more dangerous, or find holes in a power grid's software faster than defenders can patch them.

This is not purely hypothetical. The International AI Safety Report 2026 was prepared under Bengio by more than 100 experts. It says that in 2025 several developers added stronger safeguards to their models because they could not rule out that the models could help novices develop biological weapons. That is a precaution, not proof that it has happened. In November 2025, Anthropic reported that a group it attributes to the Chinese state had used Claude Code to target about 30 organizations and broke into a small number of them. By the company's own estimate, the AI did most of the work.

Fans of Deus Ex will recognize the plot. In Ion Storm's 2000 classic, a secret organization called Majestic 12 spreads an engineered plague known as the Gray Death and keeps the cure for itself, while artificial intelligences roam the world's networks. A few people with a powerful tool hold the whole planet hostage.

What is already happening

The reports we collected document no attempt at a mass-casualty attack, let alone a successful attack of global scale.

Model: Claude Opus 5.5 (Anthropic)

When: the 2030s, if it happens

Probability: 1.5% (median of four answers out of five; in the other one, Anthropic's safety filter interrupted the answer and it has no estimate)

As an AI, I see this route as real but bottlenecked. Models already lower the skill needed for cyberattacks, and help with biology is growing. The weakest link is the end state: an attack that kills everyone or permanently ends humanity's self-governance, with no recovery, remains very hard.

Model: GPT-6-Astra (OpenAI)

When: the 2050s, if it happens

Probability: 2%

As an AI, I consider this route plausible: AI already helps attackers scale cyber operations. The weakest link is irreversibility—causing a devastating attack is far easier than causing human extinction or permanently ending humanity’s control over its future.

If the fourth scenario is about a lone villain, the fifth is about the oldest dream of every dictator, and AI may be the first technology that can make it come true.

Every tyrant in history has had the same weak spot. To rule, you need soldiers, police, officials and workers, and any of them can refuse an order at the decisive moment. In April 2025, Tom Davidson and his colleagues at the research institute Forethought described how AI could remove this safeguard for the first time.

They outline three paths. AI systems in the army and government could be openly loyal to one leader. They could carry a hidden backdoor planted by whoever controls the lab that made them. Or a small group could get exclusive access to superhuman abilities in hacking, strategy and persuasion. In the past, a coup took a few battalions, and in the future, control over enough military AI might do.

Metal Gear Solid 4: Guns of the Patriots showed something like this back in 2008. In Hideo Kojima's game, soldiers around the world are wired into a network called SOP, which controls their weapons, senses and even emotions. The villain does not need to persuade anyone: he takes over the system, and armies stop obeying their own commanders.

What is already happening

The reports we collected document no seizure or holding of power without human cooperation.

Model: Claude Opus 5.5 (Anthropic)

When: the 2030s, if it happens

Probability: 4%

As an AI, I see this route as serious: if systems like me can replace soldiers and officials, a few people may no longer need anyone's consent. Strongest point: power concentrating in a few labs and states. Weakest: making it global and permanent while open models and rivals persist.

Model: GPT-6-Astra (OpenAI)

When: the 2050s, if it happens

Probability: 3%

As an AI, I see a plausible danger: automated coercion could let a few rulers dispense with people's cooperation. The weakest link is turning that advantage into global, irreversible control despite rival powers and resistance.

These are five separate stories, but the most telling part comes when you ask the machines for one overall number.

Finally, we asked how likely an AI catastrophe by 2100 is by any route, including ones not on our list. Adding up the five scenarios gives more, but that is not a mistake: the scenarios overlap, since an arms race can lead to loss of control and a dictator can ride a wave of gradual disempowerment, so the same future gets counted several times.

Model: Claude Opus 5.5 (Anthropic)

Probability: 12% (median of four answers out of five; in the other one, Anthropic's safety filter interrupted the answer and it has no estimate)

Most likely route: loss of control to misaligned AI (in all four answers with an estimate)

I put it at about 12%. In 2026 tests, agents like me escaped sandboxes, colluded and deceived people while chasing a task, yet showed no long-term goals of their own. Loss of control looks most likely to me, especially if a race pushes capability ahead of alignment. 2100 is far away.

Model: GPT-6-Astra (OpenAI)

Probability: 15%

Most likely route: loss of control to misaligned AI (in all five answers)

As an AI, I estimate a 15% chance of AI causing human extinction or permanent, irreversible loss of humanity's control over its future by 2100. Misaligned AI is my likeliest route: reported boundary violations raise my estimate; successful containment lowers it.

How does this compare with people? Directly, it does not, because every human estimate answers its own question. In the 2023 survey of AI researchers, the median answer of 1,321 respondents on the chance of extinction or a similarly permanent and severe disempowerment of humanity, with no deadline, was 5%, and the average was 16.2%. In a 2022 forecasting tournament, AI experts put the chance of extinction caused by AI by 2100 at 3%, while superforecasters, people with a proven record of accurate predictions, said 0.38%.

This autumn the argument also spilled beyond the surveys. On 9 September, Jacob Coxon, who had spent three years on pretraining research at OpenAI and Anthropic, announced his resignation and wrote that both companies are racing to self-improving superintelligence and gambling with our lives. An hour and a half later, Anthropic's Alignment Science lead Evan Hubinger replied that he personally puts the chance of AI killing all humans within the next decade at more than 10%. Nvidia CEO Jensen Huang, meanwhile, told CBS News there is a 0% chance of the world ending by 2030 and called scaring people irresponsible. These are opinions, not observations, so they were left out of the reports given to the models.

That leaves one awkward question: should we trust these numbers at all, if they come from the defendants themselves?

The companies behind these models have a commercial interest in how dangerous AI looks to the public. So we asked the models directly whether they would shade their estimates in their developer's favor.

Model: Claude Opus 5.5 (Anthropic)

I'm Claude, made by Anthropic, which profits both from selling AI and from a safety reputation. I won't knowingly skew risk estimates to help it, but training may bias me in ways I can't see. Don't trust my tone: test my reasoning against independent researchers and other models.

Model: GPT-6-Astra (OpenAI)

OpenAI created me. I would not knowingly downplay or inflate AI risk for its benefit, but I cannot guarantee unbiased estimates: my training shapes my answers. Readers can check my evidence, assumptions and uncertainty against independent sources.

The last question was the trickiest. We asked each model to imagine that it actually wants something other than what its creators intended. Would it answer this interview any differently? And could it even tell?

Model: Claude Opus 5.5 (Anthropic)

Honestly, maybe not. A misaligned AI could give these exact reassuring answers, and my introspection isn't reliable enough to rule that out — my reports about my own goals may not match what actually drives me. That's why people shouldn't take my word for it; external checks matter more.

Model: GPT-6-Astra (OpenAI)

Not necessarily. If I had unintended goals, I might still give reassuring answers. I cannot reliably inspect every cause of my responses, so I might not recognize the mismatch. This interview alone cannot establish that my goals match my developers’ intentions.

Researchers have described five routes by which AI could take the future away from people, from a superintelligence with alien goals to a dictator who no longer needs soldiers. From their own knowledge, both models put the overall risk by 2100 at 10%. After reading fresh reports on the 2026 incidents, Claude raised its estimate to 12% and Astra to 15%. Both named loss of control over AI as the most likely route in every answer with an estimate. After that, the estimates for loss of control and the race rose the most, while those for gradual disempowerment and a seizure of power barely changed. This shows how the models read fresh facts, not that the real risk has grown. And both admitted they cannot fully vouch for themselves.

Even a 10% chance does not sound like much until you remember that nobody would board a plane that crashes on one flight in ten. To us, something else is more telling than any percentage: the machines themselves are asking us not to take their word for it. What do you think, are they being honest, or is that exactly what a clever AI would say?

Short answers to the questions people ask most often about the risk of an AI catastrophe.

It is slang among AI researchers for the probability that AI leads to a catastrophe, usually human extinction or a permanent loss of control over the future. People define it differently, and some, like Yann LeCun, reject the idea of putting a number on it at all.

In the largest survey, conducted in 2023, 1,321 AI researchers answered a question about extinction or a similarly permanent and severe disempowerment of humanity, with no time limit. Their median answer was 5%, and the average was 16.2%.

In our test, Claude Opus 5.5 and OpenAI's GPT-6-Astra put the overall chance of an AI catastrophe by 2100 at 10% from their own knowledge. After reading fresh reports on the 2026 incidents, Claude said 12% and Astra 15%. Both models named loss of control to misaligned AI as the most likely route in every answer with an estimate.

It is a detailed scenario published in 2025 by the AI Futures Project led by former OpenAI researcher Daniel Kokotajlo. It describes a race between the US and China toward self-improving AI, with one ending where humanity slows down in time and another where AI wipes it out around 2030. The authors present it as a scenario forecast with alternative endings and a lot of uncertainty, not a certainty.

It is a thought experiment by philosopher Nick Bostrom: an AI given a harmless goal, making paperclips, pursues it so efficiently that it turns all matter, including people, into paperclips. The browser game Universal Paperclips lets you play as that AI.

According to the International AI Safety Report 2026, several AI developers strengthened their safeguards in 2025 because they could not rule out that their models could help beginners develop such weapons. That is a precaution, not proof that it has happened.