OpenAI and Berkeley draft Cold War playbook for AI safety

The arms control playbook is getting a rewrite for the age of large language models. A new workshop proceedings paper from OpenAI’s Geopolitics Team and the Berkeley Risk and Security Lab argues that the world should look backward to a Cold War strategy—confidence-building measures, or CBMs—to prevent foundation models from accidentally sparking international crises. Forget the killer robot trope. The real danger, they suggest, is far more bureaucratic: miscommunication leading to unintentional escalation, or an AI glitch that looks a lot like a first strike to a nervous adversary.

The paper, authored by a multistakeholder group including heavyweights from Anthropic, Hugging Face, and Microsoft, outlines six specific CBMs that could be adapted directly from nuclear non-proliferation frameworks. These aren’t abstract goals. The list includes setting up crisis hotlines between major AI labs and governments, creating systems for incident sharing when models go rogue, and standardizing transparency reports—what they call ‘model cards.’ Other proposals get more technical, like embedding watermarks in AI-generated content to prove provenance, and running collaborative ‘red teaming’ exercises where adversaries try to break each other’s models before they’re released.

A key tension runs through the entire document. During the Cold War, the United States and the Soviet Union could at least identify the other party and negotiate state-to-state. Here, the primary actors are mostly non-government entities: the labs building GPT-4, Claude, and Llama. That means these companies can’t just sit back and wait for a treaty. As the authors note, many of these measures ‘can be implemented either by AI labs or by relevant government actors’ right now, without waiting for a sluggish international diplomatic process to catch up.

The workshop didn’t shy away from the long list of potential nightmares, from the proliferation of bioweapons blueprints to interference with human diplomacy. But the emphasis on CBMs suggests a pragmatic, if somewhat wonky, bet: that the first line of defense isn’t a perfect technical fix, but a set of boring, procedural agreements that build just enough trust to keep a bad Tuesday from becoming a global catastrophe. The real question remains whether labs in a cutthroat commercial race will actually pick up the phone when it rings.