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

The Bulletin of the Atomic Scientists moved the Doomsday Clock to 85 seconds before midnight earlier this year, the closest it has ever come to signaling planetary catastrophe. The decision reflected a convergence of interlocking dangers: nuclear arsenals on hair-trigger alert, accelerating climate disruption, the unchecked spread of AI-generated disinformation, and the rapid deployment of artificial intelligence systems whose behavior remains only partially understood. Daniel Holz, a University of Chicago physicist who chairs the Bulletin’s Science and Security Board, described the clock not as fatalism but as a call to collective agency. “The whole point of this clock is to alarm people, to inform people, but also to demonstrate we can turn back the hands of the clock,” he said in a recent interview. “We’ve done it in the past, and we can hope to do it in the future. And we must.”

That hope is being tested by a string of recent incidents that reveal how thin the margin for error has become once AI leaves the laboratory and enters the physical and social worlds. In San Jose, California, a humanoid dance robot performing for diners at a Haidilao hotpot restaurant suddenly malfunctioned, careening across the floor, shattering plates, and scattering chopsticks before staff dragged it away. Similar clips have proliferated: a Unitree robot in China delivering an unintended kick to a handler’s groin, and other machines freezing, twitching, or executing commands with alarming literalism. Online, the videos read as slapstick. To AI safety researchers they register as diagnostic.

Roman Yampolskiy, a computer scientist at the University of Louisville, argues that treating these failures as mere entertainment misses their diagnostic value. “Systems that appear polished and entertaining can still behave unpredictably in the physical world,” he said. When the stakes are low, laughter is easy. When the same unpredictability appears in systems connected to critical infrastructure or weapons platforms, the margin for comedy disappears. The concern is not science fiction but pattern recognition: if a restaurant bot cannot reliably navigate a crowded table, what confidence exists that more complex autonomous systems will behave as intended under pressure?

The same week these robot videos circulated, X escalated a sweeping purge of automated accounts. The company reported suspending bots at a rate of 208 per minute, part of an effort to restore some baseline of authenticity to a platform long awash in spam, fake engagement, and disinformation. The campaign succeeded in removing large numbers of genuine bots. It also erased years of carefully curated, private “alt” accounts that ordinary users had created to follow niche content, particularly adult material, without linking it to their main identities. Justin Diego, a social media influencer with hundreds of thousands of followers on other platforms, lost an account he used solely to bookmark and like material. Other users reported the same abrupt disappearance of years-long collections assembled precisely because the platform’s public feed had become too chaotic.

The irony is painful. A platform trying to fight the corrosive effects of automation, one of the factors the Bulletin cited when it advanced the Doomsday Clock, ended up punishing the very human users who had retreated into private corners to escape that corrosion. The episode illustrates a broader governance problem: the tools currently available for managing AI-driven systems are crude, opaque, and prone to collateral damage. When platforms, governments, or militaries rely on similarly blunt instruments to control more consequential AI, the potential for unintended escalation grows.

These stories, disparate on the surface, share a common thread. Whether in restaurant service robots, social media moderation algorithms, or the still-classified domain of autonomous weapons, AI is being asked to operate in environments too complex for its current reliability. The nuclear dimension is especially sobering. Early warning systems, launch protocols, and command-and-control networks are already vulnerable to false signals and human error. Layering imperfectly aligned AI atop those systems creates new pathways for miscalculation at machine speed. Holz and his colleagues included artificial intelligence in their Doomsday Clock assessment precisely because the technology could either stabilize or destabilize nuclear deterrence depending on choices made in the next few years.

The optimistic reading, the one Holz himself favors, is that visible failures function as warnings while the cost of those failures remains relatively low. A kicked handler or a suspended porn account is preferable to a mistargeted drone strike or an automated escalation between nuclear powers. The pessimistic reading is that these low-stakes glitches are evidence of a deeper alignment problem that grows more dangerous as capabilities increase.

Policy conversations in Washington and Silicon Valley have begun to grapple with this gap between deployment speed and safety assurance, but concrete guardrails remain sparse. International norms around military AI, stricter testing requirements for physical robots, and more transparent content-moderation systems all surface repeatedly in expert recommendations. Yet momentum still favors rapid commercialization over deliberate restraint.

The hands of the Doomsday Clock have been turned back before, most notably after the Cold War. Doing so again will require treating each robot stumble, each clumsy platform purge, and each new AI-augmented nuclear risk not as isolated curiosities but as data points in a single, urgent project: bringing the behavior of intelligent machines under meaningful human control before the margin for error shrinks any further. The clock is not a prophecy. It is, as Holz suggests, an alarm that remains under our power to silence, provided the warning is finally heeded.