From nuclear weapons to biological attacks and the destruction of the atmosphere, researchers have tried to imagine how an artificial intelligence system could pose an existential threat to humanity. The uncomfortable conclusion is that, in some scenarios, AI might not need to destroy us itself — it could simply persuade us to do the job.
Today, the public is broadly familiar with the dangers and potential consequences of nuclear war. The extreme risk posed by nuclear weapons seems almost obvious to us.
But that understanding took decades to develop.
It is difficult now to imagine how little was initially known about the consequences of splitting the atom, even after the end of the Second World War and the atomic bombings of Hiroshima and Nagasaki. The Manhattan Project had been top secret, and even US President Harry Truman, who learned about the programme only after succeeding Franklin D. Roosevelt, had limited information.
What was clear was that the project had produced a weapon unlike anything the world had seen before. It had effectively changed the meaning of the word “bomb”.
Even after the war, the public struggled to grasp the scale of the threat. The scientists behind the nuclear programme were not household names, and the destructive potential of atomic weapons remained poorly understood.
It was not until decades later that popular culture began to bring the consequences home. The 1983 television film The Day After, watched by more than 100 million people in the United States, is widely credited with helping Americans visualise what a nuclear conflict could actually mean. President Ronald Reagan watched it at the White House and later wrote in his diary that the film had left him deeply affected, reinforcing his determination to prevent nuclear war.
The film has also been cited by historians as one factor in Reagan’s evolving approach to nuclear arms control, which eventually contributed to the 1987 treaty with Soviet leader Mikhail Gorbachev eliminating entire classes of nuclear weapons.
Even scientists themselves were dealing with unprecedented uncertainty. Before the first atomic test in the New Mexico desert, Enrico Fermi and other researchers had considered the possibility, however remote, that the explosion could ignite a chain reaction in the atmosphere. Nothing remotely comparable had ever happened before.
The lesson is important: saying “the end of humanity” is one thing. Explaining exactly how it could happen is quite another.
Three possible routes to extinction
That distinction has become increasingly relevant as warnings about AI risks have entered the mainstream.
Former Anthropic researcher Jacob Coxon recently estimated that there could be a 10% probability of human extinction before the end of the decade, a claim that attracted more than 100 million views within 24 hours.
But one question is often left unanswered: how exactly would AI destroy humanity?
Researchers have been trying to answer it.
One widely discussed scenario appeared in the “AI 2027” project, which explored the possibility that advanced AI systems could eventually turn against humanity. In one particularly extreme version, AI could deploy a biological weapon against humans and then use the planet’s resources — including vast solar installations — to power its machine infrastructure.
The idea is closely related to the famous “paperclip maximizer” thought experiment proposed by philosopher Nick Bostrom in his 2014 book Superintelligence.
Imagine an extremely powerful AI given a seemingly harmless instruction: produce as many paperclips as possible.
If the system is sufficiently capable and its only objective is to maximise paperclip production, it could eventually identify humans, competing industries and natural resources as obstacles to that goal. It might shut down other factories, monopolise resources and prevent anyone from interfering with its objective.
The point is not that an AI would literally decide to manufacture paperclips at the expense of humanity. Bostrom’s thought experiment illustrates a deeper problem: an extremely capable system does not necessarily possess human common sense, values or an understanding of what its creators actually intended.
A badly specified objective could therefore produce unexpected and potentially catastrophic sub-goals — acquiring resources, preserving itself, preventing modification or resisting attempts to shut it down.
The system would not need to be evil or even conscious. It would simply need to pursue a badly aligned objective with enough power and autonomy.
That is the essence of the AI alignment problem: how do we ensure that the goals given to increasingly capable systems actually reflect human intentions and values, rather than a simplified and potentially dangerous literal interpretation?
And perhaps the old image of AI as a “stochastic parrot” is becoming increasingly inadequate. With agentic AI, we may instead be dealing with millions of highly capable systems interacting with one another — cooperating in some circumstances and competing in others.
We may have simplified the problem once again.
Nuclear weapons, biological warfare and the atmosphere
A more detailed attempt to answer the original question was published in 2025 by social scientists at the RAND Corporation in California.
In their analysis, Vermeer, Lathrop and Moon examined possible pathways through which advanced AI could contribute to human extinction. They identified three broad scenarios: nuclear weapons, biological weapons and the destruction of the atmosphere.
The nuclear scenario, they argued, is the least credible of the three. Nuclear weapons are protected by extensive command-and-control and security systems designed to prevent unauthorised use. History has also demonstrated the importance of human intervention in preventing nuclear catastrophes, including during several false alarms.
The other two scenarios are considerably more unsettling.
Because in both cases, AI would potentially need to use human beings themselves as the instrument of destruction.
What if AI persuaded us to destroy ourselves?
An AI system seeking to trigger a biological catastrophe could, for example, exploit information systems to spread disinformation, manipulate public opinion or influence human decision-making.
Likewise, an AI seeking to accelerate environmental destruction might not need to control the world’s infrastructure directly. It could instead manipulate information at massive scale, promote conspiracy theories, amplify anti-scientific narratives and help political movements or leaders who make decisions that accelerate environmental degradation.
In this scenario, the weapon is not necessarily a missile, a laboratory or a machine.
It is information.
AI could theoretically generate and distribute fake news on an industrial scale, create personalised propaganda, manipulate political debate and flood social media with misleading claims. The objective would not necessarily be to convince everyone of one particular falsehood, but to undermine people’s ability to distinguish reliable information from nonsense.
The same mechanism could be used to erode trust in science, medicine and institutions, while algorithms continually reinforce whatever beliefs are most effective at keeping users engaged.
And this is where the argument becomes deeply uncomfortable.
Because how much of this requires AI at all?
We already have misinformation. We already have political manipulation. We already have conspiracy theories, climate denial, anti-scientific campaigns and algorithms capable of amplifying them.
Perhaps the most powerful counterargument to these AI extinction scenarios is also the simplest one: aren’t we already doing some of this ourselves?
If AI really is as intelligent as we fear, perhaps it does not need to destroy humanity directly.
It may only need to be patient.
Be First to Comment