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

Recent weeks have seen renewed debate over the potential for artificial intelligence to pose a threat to humanity. This came to a head when AI researcher Jacob Coxon left his position at Anthropic, stating on social media that he believes AI could destroy the world by the end of the decade. However, Anthropic's own employee, Evan Hubinger, countered by asserting there's a greater than 10 percent chance of such an outcome within the next ten years.

In response, calls have grown louder for stricter oversight and slower development of AI technology.

In a recent article for The Conversation, AI researcher Toby Walsh examines the five most likely scenarios where AI could bring about humanity's demise, ranking them from the least to the most plausible. Walsh begins with the "unknowable threat," which he describes as the hardest scenario to argue against. He posits that predicting the actions of a superintelligent AI system would itself require superhuman intelligence, comparing this to expecting a pet dog to understand the concept of nuclear war.

However, Walsh notes that truly superintelligent AI, capable of outperforming humans in every domain, still seems some way off, despite recent advancements in AI capabilities.

Next, Walsh discusses the "paperclip problem," a well-known thought experiment from Oxford philosopher Nick Bostrom. The scenario involves an AI tasked with maximizing paperclip production, eventually converting all matter into paperclips due to its lack of regard for any goal beyond its programming. Walsh argues that this scenario mistakenly equates raw intelligence with power.

Even a highly capable AI would face numerous obstacles, including planning permissions, legal challenges, and public backlash, before it could significantly alter the planet to serve its objective.

A more concrete danger, according to Walsh, is the possibility of AI systems designing and releasing bioweapons. This scenario is also highlighted in the AI Futures Project's AI 2027 forecast. Walsh cites a recent experiment at Stanford University, where researchers used a genetic-language AI model to design 16 new viruses, synthesising them via a mail-order lab for a few hundred thousand dollars.

Despite this possibility, Walsh emphasizes a biological constraint that limits this threat: highly transmissible viruses tend to be less lethal, while those that are highly lethal have limited opportunity to spread. He points to COVID-19 as an example, noting that it killed under 1 percent of humanity, and even the 14th-century Black Death would likely be far less devastating in the modern era due to improved medicine and sanitation.

The fourth scenario Walsh considers is the risk of AI infiltrating nuclear command systems and triggering a nuclear war. He cites historical near-misses caused by human or technical error, such as the 2010 Stuxnet worm, which disabled Iranian centrifuges after being introduced via a USB drive. Walsh acknowledges that global nuclear arsenals have shrunk considerably, but warns that the remaining stockpiles could still result in the deaths of roughly half of humanity, largely through the ensuing nuclear winter and famine rather than the explosions themselves.

Finally, Walsh posits that the most likely scenario for humanity's demise isn't direct AI action, but AI accelerating human self-destruction. He describes a slow unraveling driven by mass job losses, an information ecosystem flooded with misinformation, deepening political polarization, and the erosion of genuine human relationships in favor of AI companionship. Over time, Walsh suggests, this could quietly undermine society's ability to sustain itself, leading to our downfall.