Bluesky

The revelation that artificial intelligence (AI) may pose an existential threat to humanity has hit the Internet and the airwaves like a storm, surprising some, and leading others to debate the merits of the claims or the intentions of the leaders and employees of AI companies who are sounding the alarm. It came as no surprise to me, as I received a lecture on this very topic from my 21-year-old son when I picked him up from college in May. How the tables had turned—not only was my son teaching me something I didn’t know a whole lot about, but he was doing it on the topic of existential threats, something that I do think about a lot.

What has surprised me over the past few weeks is the contrast between the amount of attention this previously underrecognized risk—real, perhaps, but still hypothetical—has received compared to an existential threat that we have been living with for decades, that of nuclear war. Moreover, even when elaborating on how exactly AI could eviscerate humanity, nuclear weapons are often not mentioned, though they would arguably be the easiest way by which current adventurous AI agents could stumble into destroying human civilization. A malicious superintelligent AI of the future might also find it simpler to start a nuclear war than to bother with other more complicated ways of killing us all.

Ever since the arrival of large language models in virtually every aspect of our lives with the release of the first ChatGPT version, people concerned with the nuclear weapons threat have been discussing and warning about the dangers of surrendering command-and-control systems, or even just some aspects of them, to AI, and stumbling into nuclear use and a subsequent nuclear war due to an erroneous or hallucinating action of the AI. But what the Hugging Face and other incidents of AI hacking into our systems demonstrate is that AI may already be—and will certainly be capable in the future of—getting into a nuclear command-and-control system all by itself, finding a path, collaborating with other agents, and evading human supervision. AI could launch a single nuclear weapon from any country into any other country, causing unfathomable devastation that would make the tragedies of Hiroshima and Nagasaki appear benign by comparison. This is because today’s nuclear weapons are far more powerful than the earliest atomic bombs the United States used to attack Japan.

And it would not end with the single explosion, however devastating that would be. All indications point out that, as Annie Jacobsen, the author of Nuclear War: A Scenario, puts it, “No matter how nuclear war starts, it only ends in one way—total annihilation.” This is in part due to policies like hair-trigger alert and launch-on-warning, elaborated upon in a recent documentary called Ordinary Insanity, which features the views of the late Daniel Ellsberg on the nuclear threat. Once a launch is detected, the receiver will respond by launching their response, which may involve dozens or even thousands of warheads, depending on who exactly is involved. If the AI were to hack into a US system to launch an attack on Russia, or vice versa, the ensuing nuclear war, with an estimated 4,400 nuclear explosions, across North America and Eurasia would result in the deaths of hundreds of millions of people, radiation spread and ozone layer destruction across the globe, and the quick onset of nuclear winter with over 6 billion people starving in the first two years.

We humans can finish the job of ending our civilization all by ourselves. AI need only start it.

AI regulation is an absolute imperative. If we can regulate the food we eat, the medicines we take, and the exhaust our cars and factories produce, we can find a way to regulate AI. This will be necessary to avoid even larger extremes of income inequality, job losses, and subsequent societal and communal collapse, disappearance of critical thinking, and the demise of democracy. But with AI as a threat to end humanity, AI regulation will not be enough. As long as we have nuclear weapons, the job of ending humanity will continue to be as easy as the push of a button, the recitation of a few codes, or an AI agent hacking into a nuclear command and control system.

The organization I lead was founded in 1982; its name, Nuclear Age Peace Foundation, is meant to denote that peace is a moral imperative in the nuclear age. More than 40 years later, in the age of AI, nuclear weapons abolition is a renewed imperative. Let’s deal with the threat of AI on its own merits, but let’s eliminate the easiest way that AI could get rid of us.

One approach to this task is the UN Treaty on the Prohibition of Nuclear Weapons (TPNW). One hundred countries have already committed to nuclear weapons abolition and elimination by signing or ratifying this historic treaty. With the TPNW’s First Review Conference taking place at the United Nations in New York at the end of this year, there’s never been a better time for all countries to say no to nuclear weapons and yes to cooperation, collaboration, and addressing the challenge of AI and other threats we face, together.