Could AI Wipe Out Humanity? Expert Lays Out Five Most Likely Scenarios

The debate over whether AI could eventually destroy humanity has gained fresh urgency in recent weeks. It was reignited when AI researcher Jacob Coxon quit Anthropic after only four months on the job, declaring on X that the people building AI genuinely believe it could kill everyone by the end of the decade. His departure drew an unusual endorsement from within the company itself: Anthropic staffer Evan Hubinger said he personally puts the odds of such an outcome within the next ten years at above 10 percent. Since then, calls to slow AI development and tighten human oversight of the technology have grown louder. Writing in The Conversation, AI researcher Toby Walsh examines the most commonly cited doomsday scenarios, ranking them from the least to the most plausible.

1. The unknowable threat

Walsh opens with what he calls the hardest scenario to argue against. The idea that humans simply cannot foresee how a superintelligent AI might destroy them, since predicting such a mind's actions would itself require superhuman intelligence. He compares it to expecting a pet dog to grasp the concept of nuclear war. His reassurance: truly superintelligent AI systems that outperform humans across every domain, not just narrow tasks, still appears to be some way off, even though he notes that AI has recently cracked one of the hardest known mathematical problems and appears to be making headway on others.

2. The paperclip problem

Next is a well-known thought experiment from Oxford philosopher Nick Bostrom: an AI tasked with maximising paperclip production that eventually converts all matter into paperclips, not out of malice but sheer indifference to any goal beyond its programming. Walsh argues this scenario mistakes raw intelligence for actual power. Even a highly capable AI, he writes, would run into planning permissions, legal challenges and public backlash long before it could remake the planet in service of its objective.

3. Engineered pandemics

A more concrete danger, according to Walsh, is the possibility of an AI system designing and releasing a bioweapon - a scenario also featured in the AI Futures Project's 'AI 2027' forecast. He cites a Stanford University experiment last month in which researchers used a genetic-language AI model to design 16 new viruses, synthesised via a mail-order lab for just a couple of hundred thousand dollars. Still, Walsh points to a biological constraint that limits this threat: viruses that are highly transmissible tend to be less lethal, and vice versa, since a pathogen that kills too fast has less opportunity to spread. He notes that COVID-19 killed under 1 percent of humanity, and even the 14th-century Black Death, would likely prove far less devastating today given modern medicine and sanitation.

4. Nuclear miscalculation

Walsh also raises the risk of AI infiltrating nuclear command systems and triggering war, noting several historical near-misses caused by human or technical error. While nuclear networks are meant to be fully isolated from the internet, he cites the 2010 Stuxnet worm, which disabled Iranian centrifuges after being introduced via a USB drive, as evidence that such isolation isn't foolproof. AI-generated false intelligence, he adds, could also push militaries toward catastrophic decisions. His caveat: global nuclear arsenals have shrunk considerably, though he warns the remaining stockpiles could still kill roughly half of humanity, largely through the resulting nuclear winter and famine rather than the blasts themselves.

5. Humanity turning on itself

The scenario Walsh considers most probable isn't AI acting directly, but AI accelerating human self-destruction. He describes a slow unravelling driven by mass job losses, an information ecosystem flooded with misinformation, deepening political polarisation, and the erosion of real human relationships in favour of synthetic AI companionship. Over time, he suggests, this could quietly undermine society's ability to sustain itself.