We Built a Doomsday Clock for Nuclear War. Do We Need One for AI?

---

A Clock Born From the Bomb

In the years after Hiroshima and Nagasaki, the scientists who had helped build the atomic bomb faced a peculiar problem. They understood the danger they had unleashed better than anyone, but they had no way to make a distracted, complicated public feel it.

Their solution was elegant: the Doomsday Clock. First set at seven minutes to midnight in 1947, this symbolic timepiece—maintained by the Bulletin of the Atomic Scientists—translates a tangle of geopolitical and technological threats into a single, visceral image. Midnight is catastrophe. Every January, a board of experts decides whether humanity should be moved closer to, or further from, the abyss.

As of 2024, the clock sits at 90 seconds to midnight—the closest it has ever been. And notably, in recent years its keepers have begun citing a new factor in their deliberations: artificial intelligence.

This raises a provocative idea. If a clock could galvanize public understanding of nuclear annihilation for nearly eight decades, could a similar instrument help us reckon with the rise of machine intelligence?

The Case for an AI Singularity Clock

The "singularity"—the hypothetical moment when AI surpasses human intelligence and begins improving itself beyond our control—has long lived in the realm of science fiction and academic speculation. But the past three years have collapsed timelines that experts once measured in decades.

The argument for a public-facing Singularity Clock rests on three observations:

1. The pace is outstripping public comprehension. Most people experienced AI as a novelty chatbot in late 2022. By 2024, frontier models were scoring at the level of expert humans on graduate exams, generating photorealistic video, and writing functional code. The gap between expert understanding and public perception has rarely been wider for any technology.

2. The risk is genuinely existential—and contested. Leading researchers, including Geoffrey Hinton and Yoshua Bengio, have signed statements declaring that mitigating extinction risk from AI should be a global priority. Others dismiss this as hype. A clock would not resolve the debate, but it would force the disagreement into the open, with reasoning attached.

3. We lack shared situational awareness. Unlike nuclear weapons, whose development was concentrated in a handful of state programs, AI progress happens across competing corporate labs, academic institutions, and open-source communities worldwide. There is no single dashboard. A clock could become a focal point.

What Would the Clock Actually Measure?

Here is where the metaphor gets harder than it looks—and more interesting.

A nuclear Doomsday Clock measures the likelihood of catastrophe. An AI Singularity Clock would need to measure proximity to a capability threshold, which is far slipperier. But several concrete benchmarks could anchor it:

• Benchmark saturation. When AI systems max out the hardest tests humans can design—graduate-level reasoning (GPQA), competition mathematics, advanced coding (SWE-bench)—the clock ticks forward. We are watching these benchmarks fall faster than anyone predicted.

• Autonomous capability. Can a system pursue complex, multi-step goals over long time horizons without human intervention? The day an AI can independently conduct research, write and deploy code, and improve its own successors is the day "midnight" stops being theoretical.

• The compute curve. Training runs are growing roughly tenfold every two years. The clock could track this raw capability input, much as the nuclear clock once tracked warhead counts.

• Alignment lag. Crucially, the clock should not just measure capability but the gap between capability and our ability to control it. A leap in safety research might actually move the hands backward—a built-in incentive to invest in solutions, not just sound alarms.

This last point matters. A well-designed clock would reward governance, transparency, and safety breakthroughs, just as the nuclear clock has occasionally moved back during periods of arms-control progress.

The Obvious Objections

A skeptic might raise several fair points, and an honest proposal must address them.

"The singularity might never happen." True. Predictions of AGI have been wrong before, and current systems have profound limitations. But the Doomsday Clock was never a prediction that nuclear war would happen—it was a measure of risk. A Singularity Clock would track trajectory, not destiny.

"This is just fear-mongering that benefits the AI labs." A real danger. Hyping existential risk can distract from present-day harms—algorithmic bias, labor displacement, misinformation—while conveniently inflating the perceived power (and valuation) of the companies building these systems. A credible clock would need to be governed by an independent body with no financial stake, and arguably should track near-term harms alongside long-term risk.

"You can't quantify something this fuzzy." Also true to a degree. But "minutes to midnight" was always symbolic, not precise. Its power lies in being a conversation starter, a coordinating signal—not a scientific instrument with decimal-point accuracy.

More Than a Gimmick

The deepest value of the Doomsday Clock was never the clock itself. It was the annual ritual of serious people stopping to ask: Are we safer or more endangered than we were last year, and why?

That ritual is precisely what AI governance lacks. We have scattered papers, viral product launches, breathless headlines, and dismissive rebuttals—but no shared moment of collective stock-taking, no agreed-upon scoreboard, no institution whose explicit job is to translate technical reality into public meaning.

A Singularity Clock would not save us. No symbol can. But the atomic scientists understood something we are at risk of forgetting: technologies powerful enough to reshape the human future demand instruments that help ordinary people see the stakes clearly.

We built a clock for the bomb because we needed to look our own creation in the face. The question now is whether we have the wisdom to do it again—before the hands reach midnight on a future we can barely imagine.

What time is it on the AI clock? Right now, nobody can say with confidence. That uncertainty, more than any single benchmark, may be the most alarming reading of all.