Why Meta Is Restricting Claude Code and Codex Over Model Distillation Risks

As AI programming assistants become more capable, Meta finds itself in a paradox: the company relies on external tools like Claude Code and Codex for productivity, yet worries that these tools could compromise its own model development.

According to a report by The Information , Meta is limiting employee use of Claude Code and Codex because of concerns about model distillation – the practice of training a new model using the outputs of an existing one. OpenAI, Anthropic, Google and others explicitly forbid using model outputs to build competing systems, making distillation a legally and ethically sensitive boundary.

Earlier this year Meta created an Applied AI Engineering team tasked with improving its internal coding assistant, MetaCode. The team must build high‑quality datasets, design programming challenges, and use these tasks to train and evaluate code models.

Because Meta is also a major customer of Claude Code, engineers naturally use it in daily development. However, if code, test cases, bug analyses, or design ideas generated by Claude Code or Codex flow into MetaCode’s training or evaluation pipeline, the provenance of the model’s capabilities becomes ambiguous.

The internal guideline, as seen in the leaked memo, permits AI assistance for routine work such as workflow automation, code organization, and test‑infrastructure construction, but mandates thorough human review of any AI‑generated content. Crucially, the memo bans the use of external AI to generate programming‑challenge questions that are used to test Meta’s own models, stating that such practices would relinquish design control to the external model.

Additional restrictions forbid engineers from using AI‑generated analysis to discover code vulnerabilities or to create test‑task ideas. The policy also specifies that any AI‑generated artifacts must not be placed in infrastructure containers that internal models can access, preventing inadvertent contamination of the training data chain.

One internal memorandum even instructed teams to pause certain tasks that relied on Claude Code and Codex, citing the risk that third‑party model outputs could enter Meta’s training data. If a model vendor determines that distillation has occurred, Meta could lose access to the external model or face account‑level bans.

Beyond compliance, Meta faces mounting cost pressure. Company‑wide AI tool usage is projected to cost tens of billions of dollars this year, and token consumption has surged, prompting Meta to impose token‑usage limits on employees.

Consequently, reducing dependence on external AI coding tools and shifting more development work to MetaCode has become a strategic priority, but only if Meta can guarantee that its replacement does not absorb outputs from the restricted tools.

Overall, Meta’s restriction highlights a broader industry shift: AI coding assistants are evolving from mere productivity aids to integral components of the model‑development supply chain, raising new questions about the provenance and accountability of AI capabilities.