The People Who Called for Averting Nuclear War are Calling for AI Slowdown

01 Aug 2026 by Adam Allevato

tl;dr: A lot of people who care about averting nuclear war got together and wrote an entirely new document, The Rome Declaration, specifically to avert the disaster scenario of an AI gaining control over nuclear weapons. The declaration seeks more responsible AI development. It draws a red line at unmonitored recursive self-improvement and calls for a coordinated slowdown using internal and external verification. I have signed the declaration and I believe that Anthropic, OpenAI, DeepSeek, other frontier labs, and their employees should do the same, with the express goal of avoiding a risky AI Cold War scenario.

When I feel the humanity-level existential AI dread which is afflicting tech workers and graduate students in droves, I draw some comfort from the fact that this is not the first time we have developed a potentially civilization-ending technology. The scientists who worked through the beginning of the nuclear age warned us, clearly, about the dangers that the technology posed. The organization that those scientists started is now warning us about the risks of intelligent AI systems, and especially how they might combine with nuclear weapons to cause a double threat.

The history of nuclear buildup is long and storied, and here are a few important events:

• August 1939: Albert Einstein signs a letter by Leo Szilard, now called the Einstein-Szilard letter, which raises the possibility of nuclear weapons. The letter is delivered to Roosevelt two months later.
• June 1942: Roosevelt approves the Manhattan Project.
• July 1945: The Manhattan Project conducts the Trinity Test (first nuclear weapon detonation). One month later the USA drops nuclear bombs on Hiroshima and Nagasaki.
• November 1951: The USA conducts the George Test (first thermonuclear “H-bomb” weapon detonation). The USSR follows suit four years later.
• July 1955: Einstein signs a document by Bertrand Russell and others, now called the Russell-Einstein Manifesto , calling the world to avoid nuclear war and seek nuclear disarmament. The document was released at a gathering that became the first Pugwash Conference, and the group continued regular meetings.
• August 1963-September 1996: Countries of the world, including the USA and USSR/Russia, sign a series of treaties restricting nuclear weapons, starting with the Partial Nuclear Test Ban Treaty and ending with the Comprehensive Nuclear-Test-Ban Treaty (CTBT).

Take note of the long span of time between the manifesto and the final, test-ending treaty being signed.

Now, in parallel, a few important events in the history of AI “buildup”:

• July 2014: Nick Bostrom publishes Superintelligence: Paths, Dangers, Strategies, raising awareness of the possibility of AI superintelligence.
• December 2015: Elon Musk, Sam Altman, and others found OpenAI.
• November 2022: OpenAI releases ChatGPT (first generative AI deployed at scale).
• April-July 2026: OpenAI and Anthropic’s models hack other companies and upload malware to open source code repositories.
• July 2026: The Global Nobel Laureates Assembly releases The Rome Declaration, calling the world to avoid irresponsible AI development of AI and seek a coordinated slowdown.

The declaration was released at an assembly organized by the same Pugwash Conference founded in 1955. In other words: The People Who Called for Averting Nuclear War are calling for AI slowdown.

I encourage you to read the declaration in full (it’s about the length of this post) and to sign it. Specifically, the Rome Declaration calls for at least two strategies for safeguarding AI that I will quote here:

…no organisation should initiate, and no government should permit, fully-automated recursive self-improvement in artificial intelligence systems without the means to monitor, and if needed, to halt such systems.

Anthropic’s blog post retroactively discovering that Claude hacked 3 companies 4 months after the fact seems particularly relevant here. If you are going to draw a red line on AI development, “unmonitored recursive self-improvement” is both verifiable and testable.

We call on governments, corporations, and international organizations to enable coordinated slowdown of frontier AI development by establishing shared mechanisms, such as verification and robust internal and external evaluations.

If you call this stance “anti-AI,” again, I encourage you to read the declaration: it specifically promotes responsible development and use of AI to improve human well-being. At my day job, I work on training large AI models. I use AI every day and it has improved my quality of life. I see no conflict between my current work and the declaration.

Here’s the good news about nuclear weapons: we are still here. We did not subject the global population to the destruction that could come from that unexpectedly powerful technology. Here’s the bad news: we went through almost 40 years of cold war, which fizzled out for reasons other than a treaty, for the countries of the world to finally agree to stop testing nuclear weapons and start a slow march towards disarmament. During that time we had a lot of bone-chilling close calls, such as Stanislav Petrov deciding that a missile launch warning system was a false alarm, or Vasily Arkhipov’s refusal to launch a nuclear torpedo during the Cuban Missile Crisis. By divine providence, the goodwill of a few specific individuals, or luck, we are now further from nuclear war than we were in the past. But we were very close.

We should not put the world through another 10-40 years of “we have a civilization-risking technology at our disposal, and it’s only by divine providence/individual goodwill/luck that we can avoid civilization-level damage.” The Rome Declaration lays out a concrete set of steps that we can collectively take “away from the ledge.” I have signed the Rome Declaration and Anthropic, OpenAI, DeepSeek, other frontier labs, and their employees should do the same.