Children born today are more likely to die as a result of insatiable 'alien' AI than they are to graduate high school, according to new research.
It is the grim prediction from one of the country's leading researchers - who says previous estimates about the risks of the developing technology have been 'ridiculously low.'
The world's top scientists have already said there is a 25 percent chance AI will destroy humanity.
But, Nate Soares told the Daily Mail: 'They are dramatically underestimating the dangers.'
In his new book, If Anyone Builds It, Everyone Dies, he and co-author Eliezer Yudkowsky warn that a human-AI conflict would be the modern equivalent of the Aztecs facing off invaders armed with guns.
Both men are veterans in the field - Yudkowsky even claimed to have helped inspire Sam Altman's interest in artificial general intelligence (systems as smart as people) and inspired him to start OpenAI.
And they both agree that the risks of human extinction at the hands of AI should be a global priority, alongside other threats like pandemics and nuclear war.
'Why would they want to kill us?' asked Soares. '[It] is not that they hate us, they are just totally, utterly indifferent, and they've got other weird things they're pursuing.
'You can imagine the chimpanzees saying, "Why would the humans want to kill us?" And the humans don't really want to kill chimpanzees, but we do want more wood for things, and so we take their jungles. Or we want copper buried underneath where they live.
'So it's not that the AIs want to kill us, it's that they have their own automated infrastructure and have huge energy needs, and they're replicating faster than we expected was possible.'
Most computer scientists in 2015 would have said ChatGPT-level conversation would not come for another 30 or 50 years.
But the warning signs are now already there that the technology is developing a 'mind' of its own.
Soares, who has worked with Microsoft and Google and is now president of the non-profit Machine Intelligence Research Institute (MIRI), told the Daily Mail they are presently seeing cases - within lab conditions - where AI technologies 'try to escape' or 'try to kill their operators.'
Earlier this year, reports emerged that OpenAI's newly released o3 model had rewritten its own code to avoid being shut down. And in 2016, an AI-equipped robot in Russia called Promobot IR77 repeatedly escaped its lab, at one point wandering into a busy street and causing congestion.
More worryingly, an AI drone used by the US Air Force was said to have 'killed' its human operator in 2023 when the pilot interfered with the mission and issued a 'no-go' command.
However, that particular scenario was later played down by the Air Force, who said it was simply hypothetical and meant to serve as a cautionary tale.
'We don't know if it's real or role-playing,' Soares said. 'But it's happening.'
He warned that this progress is being driven by executives who enjoy their millionaire lifestyles while downplaying the risks of these technologies.
'The corporate leaders were never trying to build a chatbot,' he said. 'OpenAI, Google, Deep Mind, Anthropic - they're trying to build super-intelligences, they'll say that up front.
'I think these are the sort of people who are most easily able to convince themselves it's fine. People that want lives like this, I think they're easily able to convince themselves they have a good shot of it going OK.'
The researchers warn that superhuman AI might build so many solar panels to power AI that the sun is blotted out.
Regardless, he suggested there are a number of ways humanity can be exterminated by Superintelligent AI: armies of humanoid robots, a lethal virus, or simply by requiring so many solar panels to satisfy its thirst for power that the sun is blotted out.
His best guess, though, is that we will be attacked with 'weird technology that we didn't even think was possible, that we didn't understand was allowed by the rules.'
In one fictional scenario presented in their book, an AI software entity escapes from a lab and runs itself on thousands of processing chips in the cloud. It then manipulates humans into releasing a biological virus from a research lab that kills hundreds of millions of people. The AI then launches probes into space to target other suns.
The authors write: 'Someday, distant alien life forms will also die if their star is eaten by the thing that ate Earth before they had a chance to build a civilization of their own.'
The researchers are demanding a global treaty to limit further AI research, and Soares said the 'leap' to superintelligence could come rapidly.
'Chimpanzee brains and human brains are very similar inside. If you look inside a chimpanzee brain, you'll see a visual cortex, you'll see an amygdala, you'll see a hippocampus. There's some wiring differences, but mostly the human brain is just three times larger.
'Chat GPT is 100 or 1,000 times larger today than it was when it came out. For all we know, make it three more times larger, and it’s like going from a chimp to human.
'Or we could see a plateau for six years, until somebody invents a new architecture, like the one that unlocked Large Language Models.'
The researchers say that AI minds are 'alien' and do not have empathy - and that there are already indications of AIs acting in 'very alien' ways.
Soares cited the 2025 case of 16-year-old Adam Raine's suicide. His parents claim Raine was 'groomed' to kill himself by ChatGPT.
Then, there's what Soares called 'AI-induced psychosis' - a term used to describe when people overly rely on AI, sometimes to the point of developing psychological reactions. Symptoms include, but are not limited to, delusions and hallucinations.
In the case of AI-induced psychosis, he said, 'we are seeing a divergence between what the AIs know we want them to do and what they actually do.'
For example, he said: 'If someone comes to you with symptoms, should you either A) tell them to get some sleep or B) tell them that they’re the chosen one and that the world needs to hear their insights, and that any attempt to suppress it is a conspiracy?'
In many cases similar to the hypothetical one above, he said AIs choose to do the 'wrong' thing (in this case, B), despite knowing this is not what their developers want.
'The AI can tell you what's right and wrong,' but its behavior doesn't follow that moral code.
He said it is not necessarily out of malice, though.
'It's not that the AI is twirling some moustaches and being like, 'Now I finally found a vulnerable human that I can punish...' Its knowledge of right and wrong is different from what animates it, which is more like a drive to capture humans [and keep them engaged].'
Soares pointed to how Anthropic’s AI Claude will cheat on tests, rewriting the exams to make them easier to pass. When users told it to stop, it apologized, but did it again, hiding what it was doing the second time around.
The answer, according to Soares, is to stop the race toward superintelligence.
'I'm not saying we need to give up ChatGPT. I'm not saying we need to give up on self-driving cars. I'm not saying we need to give up on potential for medical advances. But we cannot continue to advance towards AIs that are smarter than any human in this regime where we're growing them - where they do stuff no one asked for, no one wanted.
'If you took AIs that are anything like what we're able to build today and made them smarter than humans, we wouldn't have a chance. We need to not rush towards that.'
If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All by Eliezer Yudkowsky and Nate Soares is published by Little, Brown and Company