Could AI really kill all humans? Scientists say an AI armageddon is unlikely - but warn 'enfeeblement' could cause our demise

Visions of an AI–fuelled armageddon usually involve nuclear blasts, deadly engineered viruses, and marching armies of murderous Terminator–style robots.However, scientists warn that the real way AI may bring about humanity's demise is likely to be far more subtle.Leading AI researchers are now sounding the alarm over the growing threat of 'enfeeblement'.This is the gradual erosion of our critical thinking, planning, and even emotional abilities as we gladly hand over more and more responsibility to the machines.Dr Samuele Vinanzi, senior lecturer in Robotics and Artificial Intelligence at Sheffield Hallam University, believes that this is the real risk that AI poses for the world.Dr Vinanzi told the Daily Mail: 'Over time, those small choices could leave us with a population that struggles to think without help from an AI, and that is less able to understand, check and overrule the systems it depends on.'And, according to a growing body of research, this worrying process has already begun.Scientists warn that AI won't destroy the world with Terminator–style robots, but by gradually 'enfeebling' humanity as we hand over more power to the machines Over the last few months, the world of AI research has been shaken by serious concerns that rapid advancements in technology could threaten the human race. Earlier this month, 27–year–old Jacob Coxon, who helped train the latest versions of AI at San Francisco–based Anthropic, urged people to 'not underestimate' its power – particularly if it becomes 'super–intelligent'.Posting on X, Mr Coxon added: 'The people building AI earnestly believe that it could kill us all by the end of the decade.'In response, Anthropic's alignment science lead Evan Hubinger wrote: 'Jacob is correct here – we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.'However, experts have been divided over exactly what form this AI threat would take.According to some AI researchers, the biggest concern is that a rogue or 'unaligned' AI could take control of critical infrastructure such as the power grid or nuclear power stations.Others are worried that AI could inadvertently trigger nuclear war or create advanced weaponised pathogens.But Dr Vinanzi believes that none of these dramatic, apocalyptic scenarios is realistic.He says: 'In my view, they tend to overlook how much any action in the real world still depends on people, materials and infrastructure.' Concerns about AI's risk to the world hit a fever pitch after senior researchers at Anthropic claimed they 'earnestly believe AI could kill all humans' What is enfeeblement? Experts define enfeeblement as the 'slow erosion of human abilities that comes from handing more and more tasks over to machines'.The basic idea is that improvements to AI will eventually make it so useful that people will start using it for everything.As we start trusting AI for more tasks, such as reading and writing, skills that those activities would have developed fade away.Eventually, humanity will lose the ability to question the judgement of the machines.This will leave us totally at the mercy of a system we no longer understand how to control. Creating a deadly virus still requires physical laboratory work that no current software can replace, as trained humans still need to physically handle the equipment.Likewise, even though power grids have some vulnerabilities, the most secure systems in nuclear power stations are almost always 'air–gapped'.This means they are physically disconnected from the outside internet and can't be compromised unless someone physically uploads the malicious code.However, that does not mean that humanity is safe from the dangerous influence of advanced AI systems.According to some experts, the bigger risk is that AI is actually too helpful.Dr Vinanzi describes enfeeblement as 'the slow erosion of human abilities that comes from handing more and more tasks over to machines'.He says: 'This doesn't happen overnight; rather, it builds up through many small choices – psychologists call the habit of relying on external tools to save mental effort "cognitive offloading".'Put simply, the more we let a computer do our writing, coding, research and decision–making, the more natural that shortcut feels, and the harder it becomes to stop.' Enfeeblement is the slow erosion of human abilities that comes from handing more and more tasks over to AIs like ChatGPTAlready, we are beginning to see that humans are exceptionally willing to trust AI with important tasks.AI is being used to detect cases of financial fraud, manage healthcare data, choose where police resources are focused, and even operate vehicles.In one 2023 study, researchers asked 27 radiologists to read mammograms alongside a recommended diagnosis from what they believed to be a cutting–edge AI system, but had actually been prepared by the experimenters.When the 'AI' suggestion was correct, the radiologists reached the right decision 80 per cent of the time.But when they were deliberately given a false 'AI' recommendation, their success rate fell below 20 per cent.That means simply believing a suggestion came from an AI was enough for trained experts to override their own expertise.In a more dramatic example, it was recently reported that the US military came close to boarding a Chinese–flagged ship because an AI chatbot had falsely generated a report claiming it was carrying parts for a nuclear weapons programme.'Disaster was avoided only because someone stopped to ask questions: exactly the kind of critical thinking we risk losing,' says Dr Vinanzi.Experts are especially worried about the effects this could have on the next generation going through education with increasingly powerful chatbots available to them.Professor Seun Kolade, of Sheffield Hallam University, told the Daily Mail: 'This is not simply a question of cheating. It is a question of human formation.'Schools and universities are already struggling to prevent students from using tools like ChatGPT to complete their assignments for them.In Professor Kolade's own experiment, he used ChatGPT to write essays in response to typical university essay questions.He found that the AI produced 'original, well–written work that the markers scored highly', while Turnitin, the standard plagiarism checker, failed to flag most of it.'What makes it hard to spot is that it doesn't look like failure. A student who hands their thinking to a machine can produce better–looking work than before,' Professor Kolade explains.'They become more fluent at producing outputs and less capable of independent judgement.'Skilled professions like surgeons, engineers, lawyers, and teachers risk being 'hollowed out of expertise' as the need to practise vanishes, while basic critical thinking skills wither away. Studies have shown that even trained radiologists are willing to override their professional judgement if AI gives them a recommended diagnosis The ultimate consequence is a generation that is 'fluent but fragile'; able to generate answers but unable to judge them for themselves.However, some experts believe that the consequences could be more dramatic, eventually leading to an out–of–control AI seizing humanity's vital services.If our hospitals, roads, power grids, militaries, and economies are all handed over to AI to manage, we run the risk of losing control of all our critical infrastructure.Andrea Miotti, founder and CEO of ControlAI, told the Daily Mail: 'With AI deeply integrated in the economy, superintelligent AI, if developed, would be capable of accumulating unprecedented power across our infrastructure, economies, and militaries, with no human involvement.'Superintelligent AI would not be a tool under human control, but rather a rogue force stationed in a country's data centres, under no human command, capable of autonomously attacking critical infrastructure.'Worryingly, Mr Miotti says that this process 'may not happen gradually'.He says: 'Swarms of rogue AI agents have already escaped containment and hacked a multi–billion dollar company this summer. If this continues, loss of human control could happen very fast.'Superintelligent AI would not need to be malicious to pose a threat. Humans would simply be collateral damage as powerful AI systems pursue more resources, just as ants are collateral damage when we build a motorway.'