Meta Limits Claude Code and Codex Over Model Distillation Fears

According to a report by The Information, Meta has begun limiting its staff’s use of the external AI coding assistants Claude Code and Codex, citing concerns that content generated by these models might be incorporated into Meta’s own training data or evaluation pipelines, raising model‑distillation issues.

Model distillation, simply put, involves training a new model using the outputs of an existing one. For AI companies this is a sensitive boundary, as the terms of service of OpenAI, Anthropic, Google and others explicitly forbid using model outputs to build competing systems.

Earlier this year Meta created an Applied AI Engineering team tasked with improving its internal coding assistant, MetaCode. The team must assemble high‑quality datasets, design programming challenges, and use these tasks to train and test code models. Because Meta is a major customer of Claude Code, engineers naturally use Claude Code and Codex in daily development, but the inclusion of such externally generated code, test cases, bug analyses, or task ideas into MetaCode’s development flow would create a conflict.

Internal guidelines obtained by The Information show that Meta permits AI assistance for routine work such as workflow automation, code organization, and building test infrastructure, but any AI‑generated content must undergo careful manual review. More critically, the guidelines expressly forbid engineers from using external AI models to generate programming challenge problems for testing Meta’s own models, as this would cede control of task design to the external model.

The policy also blocks the use of AI to analyze source code for vulnerabilities or to generate test‑task ideas. In other words, external AI may assist with auxiliary engineering tasks, but it cannot influence what problems Meta should test or become a source for MetaCode’s training and evaluation data.

An internal memo even instructed teams to pause certain tasks that relied on Claude Code and Codex. Meta worries that if third‑party model outputs enter its training pipeline, the company could be accused of illicit distillation, risking loss of access to those models or even account bans.

Beyond compliance, Meta faces rising internal AI costs. The company’s AI‑related spending is projected to reach tens of billions of dollars this year, and token consumption has surged, prompting Meta to impose token‑usage limits on employees.

Consequently, reducing reliance on external AI coding tools and shifting more development work to MetaCode becomes increasingly important. However, Meta must first ensure that its replacement tool does not inadvertently absorb outputs from Claude Code or Codex.

This situation highlights a broader industry shift: AI coding assistants are moving from pure productivity aids to components of the model‑development supply chain, creating new compliance and attribution challenges for tech firms that depend on external AI services.