Meta is cracking down on employees using rival AI coding tools — specifically Anthropic's Claude Code and OpenAI's Codex — over fears that the company's own models could be tainted by competitor outputs. At one point, Meta even ordered a pause on certain tasks to reinforce compliance reviews. The concern? 'Model distillation,' the practice of using outputs from one AI model to train another, which may violate service terms and strain relationships with partners.
Currently, Meta only allows employees to use external AI tools for routine tasks like setting up workflows, cleaning up code, and building test infrastructure — as long as a human carefully reviews the output. But it explicitly bans using those models to generate coding challenges for evaluating its own models, or to probe for software vulnerabilities or brainstorm test scenarios.
The move is part of Meta's broader push to reshape its internal toolchain around its homegrown coding assistant, MetaCode (formerly DevMate), and to rein in its multi-billion-dollar internal AI operations costs.