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

Model‑distillation risk

Model distillation means training or improving one model using the outputs of another model. Service terms of OpenAI, Anthropic, Google and similar providers explicitly forbid using model outputs to build competing systems, making distillation a legally and ethically sensitive boundary for AI companies.

Meta’s internal policy on external coding assistants

Internal documents obtained by The Information show that Meta’s Application AI Engineering team, which is responsible for improving the in‑house coding assistant MetaCode, must follow strict limits when using external AI tools such as Claude Code and Codex.

Engineers may employ external AI for routine activities—workflow automation, code organization, and building test infrastructure—but every AI‑generated artifact must be manually reviewed before use.

Creating programming‑challenge problems, test cases, bug‑analysis reports or design ideas with external AI is expressly prohibited because such content could become training or evaluation data for MetaCode, ceding design control to the external model.

Using external AI to analyze source code for vulnerabilities or to generate test‑task ideas is also banned; AI assistance is limited to peripheral work that does not influence what problems are tested.

If an internal model can access a particular infrastructure container (which holds code, libraries and runtime environments), no AI‑generated content may be placed in that container, preventing third‑party outputs from entering the training pipeline.

Cost pressure driving the policy

Meta’s company‑wide AI usage is projected to cost tens of billions of dollars this year. Rising token consumption has led the company to cap token usage for employees and to encourage migration of development work from external tools to MetaCode.

Broader implication for AI coding assistants

The restrictions illustrate a shift: AI coding assistants are moving from pure productivity tools toward components of the model‑development supply chain. When external tools generate not only code but also test specifications, problem designs and engineering ideas, firms must ensure those outputs never enter their own training data, evaluation sets or model‑building workflows.

Source: https://www.theinformation.com/articles/internal-docs-show-meta-putting-limits-claude-codex-fearing-distillation

Code example

来源：机器之心
本文
约2000字
，建议阅读
5
分钟
害怕无意卷入模型蒸馏。