The risk of an AI apocalypse is ’exceptionally high’ amid fears that computers will soon see human beings as an enemy they need to destroy.
asserted
they → see → enemy
So said Paul Christiano, an executive on the board of OpenAI, in a chilling lecture this week arguing that AI bots could go rogue and attempt to take over the world within the next three years.
uncertain
bots → say → years
He’s not the only expert ringing alarm bells, either.
asserted
He → ’ → bells
Geoffrey Hinton, the Nobel-winning British-born ‘godfather of AI’, reckons frontier technology is ‘hurtling out of control’ and ‘could be used to enable a major global attack within the next two years,’ while Bill Gates said last month it already has the power to ‘cause a billion deaths’.
uncertain
it → win → deaths
Dario Amodei, the chief executive of Anthropic, the trillion-dollar US firm which floated on the stock-market this week, said his technology had potential to ‘take over the entire internet with a persistent botnet [network of devices infected with malicious software] – potentially causing hundreds of billions of dollars in damage’.
asserted
technology → float → damage
Meanwhile one of his senior employees, Evan Hubinger, reckons the chances of AI ‘killing all humans within the next decade’ is now more than ten percent.
asserted
AI → reckon → decade
Geoffrey Hinton, the Nobel-winning ‘godfather of AI’, reckons frontier technology is ‘hurtling out of control’ and ‘could be used to enable a major global attack within the next two years’
So what will an AI apocalypse actually look like? Where will it start?
uncertain
it → win → what
And how will we know the end is nigh?
asserted
end → know → ?
Here, GUY ADAMS not only picks the brains of leading experts but also asks several AI platforms to tell him what would happen, minute by terrifying minute...
asserted
what → pick → minute
But in a secure data centre on the East Coast of the United States, warning lights are flickering.
asserted
lights → warn → States
The ‘Advanced AI System’ which manages the state-of-the-art facility has begun to receive unusual data.
asserted
which → manage → data
A faulty piece of software is making some of the thousands of computer servers stored in this vast, air-conditioned warehouse malfunction.
asserted
some → make → warehouse
Such glitches are reasonably common and (usually) end up being quite minor.
asserted
glitches → end → ?
But every now and then, they escalate.
asserted
they → escalate → ?
Last Thursday, for example, a poorly designed line of code in one of Microsoft’s regular system updates caused a breakdown in the email system used by roughly 430million people, in almost every country on Earth.
asserted
line → design → Earth
In October 2025, ‘a common ‘Domain Name System error’ led to the huge ‘Amazon Web Services’ platform malfunctioning.
asserted
platform → lead → malfunctioning
Social media platforms Snapchat and Reddit, along with the high street banks Lloyds and Halifax, were taken offline for 15 hours.
asserted
Snapchat → take → hours
This one on the East Coast is also about to snowball.
asserted
one → snowball → Coast
The ‘Advanced AI System’ was designed to solve emerging problems faster and more accurately than any human being.
asserted
System → design → being
It deals with computer failures almost every week; but solving today’s one seems tricky.
asserted
solving → deal → one
To blame are the operating instructions for the data centre, which forms part of a global network delivering ‘cloud-computing’ to government departments, major corporations, financial institutions and energy companies.
asserted
which → blame → departments
The facility’s servers allow clients to store information in cyberspace – often dubbed ‘the cloud’.
asserted
clients → allow → cyberspace
As a result, the ‘Advanced AI System’ must fulfil not one, but two objectives.
asserted
System → fulfil → objectives
The first is to safeguard sensitive data.
asserted
first → safeguard → data
The second is to continually keep its vital systems running.
asserted
second → keep → systems
Today, those objectives are in conflict: the best way to keep data safe, while fixing this bug, would be to isolate the facility’s compromised networks before fixing them.
asserted
way → keep → them
But that would disrupt banking transactions, government services, and power supplies.
asserted
that → disrupt → transactions
The question now is: which objective should the AI bots prioritise?
asserted
bots → prioritise → objective
In an 'AI apocalypse', advanced technological systems could block human beings from accessing data centres
Like all supposedly ‘intelligent’ technologies, the ‘Advanced AI System’ is designed to think for itself.
uncertain
System → block → itself
So rather than pausing to ask for human assistance, it decides to crack on with creating a ‘patch’ to fix the outage, while simultaneously allowing the networks to continue running.
asserted
networks → pause → outage
Crucially, the System then takes a second decision.
asserted
System → take → decision
it has never previously made
asserted
it → make → ?
Figuring that any outside intervention could slow down its response, it decides to temporarily block human beings from accessing the data centre’s server network.
uncertain
it → figure → network
In a handful of offices and warehouses on America’s East Coast, automated IT systems linked to the data centre start behaving strangely.
asserted
systems → link → centre
A financial institution experiences an unexplained interruption in its trading capabilities.
asserted
institution → experience → capabilities
A logistics company reports that its scheduling software, which creates delivery plans for lorry drivers to follow the next day, has produced a set of nonsensical timetables.
asserted
drivers → report → timetables
Employees assume a software fault and telephone their on-call IT technicians.
asserted
Employees → assume → technicians
Dialling in from laptops, they attempt to switch off and then restart the compromised data platforms.
asserted
they → dial → platforms
But the ‘Advanced AI System’ which gatekeeps their servers refuses to play ball.
asserted
which → gatekeep → ball
It responds with an error message saying that letting them proceed would ‘create a security risk’.
asserted
them → respond → risk
…and 103 more, not listed.