Meta Bans Claude Code and Codex Over Training Data Fears

Meta has temporarily halted certain internal work using Claude Code (Anthropic) and Codex (OpenAI), implementing policies that bar engineers from using these tools' outputs for test tasks or code analysis without mandatory human review. The company spends billions of dollars on internal AI use annually, according to internal documentation.

The core fear is distillation: if outputs from a competitor's AI model make their way into Meta's training data for MetaCode — its own coding assistant under development — the company risks violating the terms of service of its AI partners and potentially transferring competitor capabilities into its own models.

The legal and competitive stakes

Terms of service from OpenAI, Anthropic, and Google explicitly ban using model outputs to build competing AI systems. An internal memo warned of "serious escalations with partner companies" if these restrictions were violated. Meta is building MetaCode partly to reduce its dependence on external tools and their rising costs.

The restrictions signal a broader shift: as AI coding tools become embedded in engineering workflows, companies are realizing that every automatically generated code suggestion is potentially subject to IP restrictions. Manual review gates — expensive and slow — may become standard in organizations building their own AI systems on top of commercial ones.