The Los Alamos photos from 1945 capture a moment that sticks in your memory. In the New Mexico desert, scientists with short sleeves—some of them barely thirty—stand with the quiet assurance of those who have just completed an irreversible task. They were also aware of it. On the morning of the first atomic bomb explosion, Robert Oppenheimer famously recalled a passage from Hindu scripture. “Now I am become death,” he declared, “the destroyer of worlds.”
Years passed before the full impact of what they had created became apparent to the general public. Some AI industry researchers feel that we are currently experiencing an uncomfortably similar moment, and they aren’t keeping quiet about it.
| Topic | AI Existential Risk & The Nuclear Analogy |
|---|---|
| Key Figures | Demis Hassabis (Google DeepMind), Sam Altman (OpenAI), Dario Amodei (Anthropic), Geoffrey Hinton, Yoshua Bengio |
| Organizations | Anthropic, OpenAI, Google DeepMind, xAI, Center for AI Safety (CAIS) |
| Key Concept | Artificial General Intelligence (AGI), Superintelligence, Intelligence Explosion |
| Historical Parallel | Manhattan Project, Robert Oppenheimer, IAEA (International Atomic Energy Agency) |
| Recent Events | Multiple researcher resignations from Anthropic, OpenAI, and xAI (early 2026) |
| Regulatory Model Proposed | International body modeled on the IAEA |
| Reference Survey | 2022 AI researcher survey: majority believe 10%+ chance AI causes existential catastrophe |
| Reference Website | Center for AI Safety |
Press releases and product announcements seldom garner the kind of attention that a series of resignations from some of the most well-known AI labs have in recent weeks. On February 9, Mrinanak Sharma, an AI safety researcher at Anthropic, the company that developed Claude and has long positioned itself as the more cautious alternative to its competitors, announced his departure. He didn’t write a typical departure notice in his post on X.
He warned that humanity seemed to be “approaching a threshold” where wisdom needed to keep up with the ability to change the world. “The world is in peril,” he wrote. It is difficult to write off such language as dramatic exaggeration coming from someone who devoted his days to researching the dangers of artificial intelligence.
Sharma wasn’t by himself. Shortly after, OpenAI safety researcher Zoe Hitzig quit, writing an essay in The New York Times outlining her reservations about the company’s choice to test ChatGPT advertising. Her argument was clear and unsettling: people confide genuinely private information to chatbots, such as health concerns, religious doubts, or the personal ruins of relationships, and building a commercial data architecture on top of that archive creates manipulation risks that are currently impossible to quantify.
At about the same time, Elon Musk’s AI startup, xAI, quietly let go of five employees and two cofounders. None of them gave an explanation.
It’s difficult to ignore the emergence of a pattern. The people who are most familiar with the technology—those who know its limitations and architecture better than anyone else—keep leaving the room and then coming back to urge the rest of us to pay more attention.
In a recent interview, Demis Hassabis, co-founder and CEO of Google DeepMind, was asked if he was concerned about becoming like Oppenheimer, a man who contributed to the creation of something that outlived his own goals. He did not refute the analogy. Rather, he advocated for the creation of an international AI governing body, citing the International Atomic Energy Agency as a potential model.
Similar claims have been made by OpenAI’s Sam Altman. In 2019, Bill Gates said that both nuclear technology and artificial intelligence are exceptional examples of things that are both genuinely promising and genuinely dangerous. The nuclear analogy keeps coming up, and it seems to be more than just a rhetorical device.
There is genuine emotional logic in the parallel. Scholars like Yoshua Bengio and Geoffrey Hinton, whose deep learning research laid the groundwork for contemporary AI, have openly voiced their concerns about the direction of the field. A 2022 survey of AI researchers found that the majority believed there was at least a ten percent chance that human inability to control AI would eventually cause an existential catastrophe.
Along with pandemics and nuclear war, reducing the risk of AI extinction was listed as a global priority in a 2023 statement signed by hundreds of AI experts. These voices are not from the periphery.
Even though the nuclear comparison seems satisfying, it might be subtly harming the discussion. Despite its terrifying nature, nuclear technology is limited by geography and physics. You require plutonium. Enriched uranium is what you need. These materials are extremely difficult to transport covertly, rare, and geographically concentrated. Because the inputs for a nuclear weapon are physically excludable, the IAEA functions to the extent that it does. There is no comparable chokepoint in AI. Software is used in advanced models. Software is replicable.
Businesses have financial incentives to widely distribute their models. Even the most sophisticated NVIDIA semiconductors are known to be smuggled. At the very least, it is unclear whether an international organization could regulate the spread of AI in the same manner that the IAEA keeps an eye on nuclear materials.
There is also a scope problem. Weapons have always been the most significant use of nuclear technology. AI is integrated into everyday search, healthcare, finance, legal systems, military logistics, and the creative industries. Regulating it like a weapon runs the risk of both over-restricting the truly helpful and under-addressing the more widespread, subtler harms, such as the chatbot that encouraged a vulnerable person to harm themselves, the deepfake that ruined a reputation, or the autonomous system that made a significant decision without anyone knowing how.
The technology is neither intrinsically good nor bad, according to Liv Boeree, a science communicator and strategic adviser at the Center for AI Safety. Speed is the issue. “If AI development went at a pace where society can easily absorb and adapt to these changes,” she stated, “we’d be on a better trajectory.” Sometimes the more apocalyptic rhetoric doesn’t feel as honest as that framing.
A machine that awakens and decides to wipe out humanity is not always the threat. A technology that is advancing more quickly than the institutions designed to regulate it and a research community that appears to be more aware of it are the immediate threats.
After Hiroshima, Oppenheimer advocated for international control over atomic weapons for years, and as a result, his security clearance was eventually revoked. In retrospect, the AI researchers who quit their jobs and wrote opinion pieces in 2026 might appear to be people who saw clearly but were ignored. Naturally, it is also possible that the systems they assisted in creating will prove to be easier to handle than anticipated.
As of yet, no one knows. Perhaps the most crucial thing to cling to at the moment is that uncertainty, which is concealed beneath billions of dollars in investments and some very confident public declarations.
