Meta restricts its AI engineers from using Claude and Codex — to prevent model distillation

There’s an awkward tension inside Meta’s AI division. The company has quietly told its AI engineers to stop using Anthropic’s Claude Code and OpenAI’s Codex, two of the most popular AI coding tools in the industry. But the decision, according to internal documents reviewed by The Information, has nothing to do with cost or productivity.

The problem is something far more legally delicate: model distillation.

A company-wide policy from May, still in effect, reveals Meta’s concern that its engineers could inadvertently train the company’s own AI models on output generated by a competitor’s model. That practice — known as distillation — violates the user agreements for both Claude and Codex. If Meta’s training data became contaminated with third-party AI output, the company warned, it could trigger “serious disputes and escalation with partner companies.”

Meta has already asked some teams to pause tasks that rely on these tools. The policy is broad enough to cover any use case where a competitor’s AI output might wind up in Meta’s training pipeline.

A Meta spokesperson defended the move: “We have clear policies governing how teams use AI tools to ensure they focus on high-impact work in a responsible manner.”

Distillation is a growing flashpoint in the AI industry. Companies invest billions training models from scratch, and they view competitors piggybacking on their output — even accidentally — as a form of intellectual property theft. Meta’s internal memo suggests the company believes the risk is real enough to justify curbing its own engineers’ productivity tools.