Doomsday Clock Ticks Closer as AI Risks Span Nuclear to Robotic Failures

The world stands closer to catastrophe than at any time since the Doomsday Clock began ticking in 1947. Earlier this year the Bulletin of the Atomic Scientists moved its hands to 85 seconds before midnight, citing intertwined dangers from nuclear weapons, climate change, disinformation campaigns and artificial intelligence. That single symbolic adjustment compresses decades of accumulating risk into one stark visual.

The decision was not automatic. University of Chicago physics professor Daniel Holz, who chairs the Bulletin’s Science and Security Board, helped set the new time. He describes the clock itself as a symbol of hope. Its purpose is to show that humanity has pulled the hands back before and can do so again. The 2026 update, four seconds closer than the prior 89-second setting, reflects fresh worries that AI could accelerate nuclear escalation by complicating early-warning systems, enabling autonomous decisions or amplifying misinformation that erodes diplomatic guardrails. The Bulletin has called for renewed arms-control talks and multilateral guidelines on military AI use.

On the ground the warnings feel more immediate and oddly familiar. Viral videos show humanoid robots losing control in everyday settings. At a Haidilao hotpot restaurant in San Jose, California, a dancing bot knocked over tables, smashed plates and scattered chopsticks before staff dragged it away. The restaurant later stated the machine followed preprogrammed routines but was hampered by tight quarters after guests requested an energetic performance. Similar clips from China captured Unitree machines stumbling during demonstrations. One appeared to strike a handler; another incident reportedly left a man with a bloody nose after a fall-recovery maneuver. Not every claim has been independently corroborated by multiple outlets.

Roman Yampolskiy, a University of Louisville computer scientist who has published extensively on AI safety, sees these moments as more than slapstick. He told one outlet they reveal how systems that look polished in controlled demos can still behave unpredictably once placed in the physical world. A failure around a child, a hospital patient or during a police encounter would not read as comic relief. Yampolskiy argues that as robots grow stronger and faster, small errors scale. South Korean researchers have created artificial muscles capable of lifting thousands of times their weight. Chinese models now sprint at 22 miles per hour. Boston Dynamics’ Spot robot, already deployed by dozens of North American bomb squads for reconnaissance, illustrates how quickly experimental hardware enters high-stakes environments.

A separate legal dispute underscores the gap between marketing and reality. A former engineer at Figure AI sued the company, claiming he was fired after warning that their machines were strong enough to fracture a human skull. Figure called the allegations false and attributed the termination to performance issues. Yampolskiy places primary responsibility on developers who choose to release systems before reliability standards are met. Better testing, kill switches and strict deployment rules help, he says, yet no complex system is perfectly safe in open environments.

Digital systems reveal parallel weaknesses. In April X accelerated an automated purge aimed at fake and spam accounts, flagging roughly 208 bots per minute according to the company’s head of product. The sweep also erased years of carefully curated private “alt” accounts used by real people to bookmark niche content, including adult material. Users described losing irreplaceable collections with no violations cited. One influencer with hundreds of thousands of mainstream followers called the process opaque; his verified, paid account was suspended anyway. The episode highlights how AI-driven moderation tools, trained to detect inauthentic behavior, can misclassify quiet human accounts that simply lurk or archive.

These three threads, nuclear symbolism, physical robotics and platform-scale content filtering, share one unresolved tension. AI is moving from laboratories into the systems that control weapons, move bodies and shape what billions see online. Each domain tolerates little error. A misread radar signal, a mistimed actuator or an overzealous filter can cascade. Yet development continues at speed, with commercial incentives often outrunning transparent safety protocols. The Bulletin notes that past Doomsday Clock retreats happened when governments chose cooperation over competition. Whether the same logic will apply to AI governance remains untested.

Holz insists optimism is rational. The clock has stood at 17 minutes to midnight in more hopeful eras. Its current position is meant to alarm, inform and, above all, prod action before the margin for mistake disappears entirely.