"We're just trying to be safe here"

When Sam Altman, Demis Hassabis, Dario Amodei and Elon Musk agreed to slow AI development this weekend, they used the language of safety — third-party auditors, domestic regulation, global coordination. It was careful, measured, plausible. It was also nearly identical to what Intel said in 1994.

The chip industry's cross-licensing agreements deployed the same rhetoric: standards protection, consumer safety, preventing dangerous acceleration of untested technologies. No one was lying about the stated motives. But the actual mechanism was something else. When AMD or Cyrix needed to license x86 architecture to compete, the terms were cooperative in language and predatory in practice — access granted slowly, at costs that made entry expensive, with provisions that kept the incumbent's timeline and the challenger's timeline permanently misaligned. The difference between a genuine pact and a cartel is whether anyone has to slow down their own roadmap when it matters.

AI's current agreement lacks the one thing that would distinguish safety from anticompetition: independent audit authority with teeth. The proposal mentions third-party monitors. Not what happens when OpenAI or Google's internal review finds a frontier lab cutting corners. No enforcement mechanism. No penalty structure that costs the signatories more than it costs outsiders trying to enter the space. No asymmetry revealed.

History is precise here. Intel's cross-licensing pacts survived until they required the incumbent to actually cede competitive advantage. To slow their own roadmap or grant real-time architecture access at parity pricing. The moment that cost became real, the agreements fractured. What looked like industry coordination revealed itself as something else entirely. AI agreements structured identically will fail on the same timeline: roughly 18 months, when one of the major players realizes that genuine compliance costs them quarterly revenue or board pressure. Discovers that the oversight mechanisms lack teeth.

You see this pattern everywhere capital organizes. Pharma consortia that claim to share toxicology data but actually share only what competitors already know, standard-setting bodies where the voting structure ensures incumbents can veto innovations that threaten them. The stated motive and the actual structure can coexist because one is true and the other is also true. What matters is not which one someone believes. What matters is what happens when you make the safety provisions expensive. That is when you learn what you are actually in.