Meta bans Claude Code & Codex over training data concerns.

Internal documents reveal that Meta has restricted some internal projects from using Claude Code (Anthropic) and Codex (OpenAI). These measures include prohibiting engineers from testing or analyzing their code using the AI-generated code without a human review. Meta invests billions of dollars each year in internal uses of AI.

Meta’s concern is that competitors’ AI code will be distilled into its own MetaCode training data. This could cause them to breach their contracts with their AI vendors and even incorporate the capabilities of other firms into their own algorithms.

1The legal and competitive stakes

Google, OpenAI, and Anthropic all have policies forbidding users from using the output of their AI models to create other AI models. In an internal email, Meta said that “there will be serious escalations with partner companies” if these policies are breached. One reason why Meta is developing its own MetaCode is to decrease its reliance on third party tools and the escalating costs associated with them.

As AI code generation becomes more integrated into software development, companies will need to consider how they can avoid infringing on intellectual property rights for any auto-generated code. Organizations developing their own AI tools on top of commercial offerings might find themselves having to implement manual review processes to mitigate this risk, which can be time-consuming and costly.