OpenAI, Anthropic, and Google are sharing intelligence through the Frontier Model Forum — a body they jointly control — to detect and counter what they term "adversarial distillation" of their models by Chinese AI laboratories. [1] Three companies that compete for talent, customers, and compute are cooperating when the threat is Chinese. The arrangement raises a question they would prefer not to answer: is this geopolitical responsibility or the world's most expensive trade association?

Model distillation is the practice of training a smaller model on the outputs of a larger one, effectively transferring capabilities without access to training data or weights. Several Chinese AI laboratories have produced models that independent researchers concluded incorporated significant distillation from American frontier systems. [1] DeepSeek's rapid progress, announced earlier this year, accelerated these concerns.

The Frontier Model Forum was originally established as a safety body — a venue for AI labs to coordinate on red-teaming, evaluations, and deployment standards. [2] Its expansion into anti-distillation intelligence sharing represents a significant scope change. The forum is now functioning, at least partly, as a competitive intelligence operation.

The cartel question is real. Three companies with combined market positioning that shapes the global AI landscape coordinating to identify and respond to a competitor class is the structure of a cartel, whatever the geopolitical framing. [1] The companies would argue that protecting model integrity is a safety matter. That argument would be more persuasive if the beneficiaries were not identical to the participants.

-- DAVID CHEN, Beijing