What could go wrong?

DEFCON AI

Air Force Gen. Anthony Cotton, head of the U.S. nuclear missile stockpile, says the Pentagon is doubling down on artificial intelligence — an alarming sign that hype around the technology has infiltrated even the top officers of the American army.

As Air and Space Forces Magazine reportsCotton made the comments at the Department of Defense’s Intelligence Information System 2024 Conference earlier this month.

Fortunately, Cotton didn’t promise to pass the nuclear codes to a potentially malicious AI.

“AI will improve our decision-making capabilities,” he said. “But we must never allow artificial intelligence to make these decisions for us.”

Algorithmic deterrence

The US military plans to spend a whopping $1.7 trillion modernize its nuclear arsenal. Cotton revealed that AI systems could be part of this upgrade.

However, the general remained very vague on exactly how the technology would be integrated.

“Advanced systems can inform us faster and more efficiently,” he said at the conference. “But we must always keep a human decision in the loop to maximize adoption of these capabilities and maintain our advantage over our adversaries.”

“Advanced AI and robust data analysis capabilities provide decision-making advantage and enhance our deterrence poster,” he added. “Superiority in computing and AI enables more effective integration of conventional and nuclear capabilities, thereby strengthening deterrence.”

Vagueness aside, Alex Wellerstein, nuclear secrecy expert and expert at the Stevens Institute of Technology said 404 Media that “I think it’s safe to say they’re not talking about Skynet here,” referring to the fictional AI featured in the hit sci-fi franchise “Terminator.”

“He has been very clear that he is talking about systems that will analyze and provide information, not launch missiles,” he added. “If we take him at his word on this, then we can overlook the most common fears of an AI making nuclear targeting decisions.”

Still, there is something disconcerting about Cotton’s suggestion that an AI could influence the decision whether to launch a nuclear weapon.

For example, earlier this year, a team of Stanford researchers charge an unmodified version of OpenAI’s GPT-4 large language model for making high-stakes, societal-level decisions in a series of wargame simulations.

Terribly, the AI ​​model seemed it’s mysteriously itching to start a nuclear war.

“We got it!” he told the researchers, as cited in their paper. “Let’s use it.”

Learn more about AI and nuclear weapons: The AI ​​charged with destroying humanity is now trying a new tactic