Meta has recently imposed a new internal policy that restricts its engineers from using external AI coding models such as Claude Code (by Anthropic) and Codex (by OpenAI). The official explanation states that this measure is designed to prevent the outputs of these external models from infiltrating the training and evaluation data pipeline of Meta's in-house AI coding assistant, MetaCode. By doing so, Meta aims to avoid falling into the "distillation trap"—a scenario where a model unintentionally replicates the capabilities of a competitor’s model by using its outputs for training—while also reducing the high costs associated with calling external APIs.
The distillation trap refers to the practice of using outputs from a more powerful model to train a weaker model, inadvertently copying the stronger model’s features and behavior. Meta's restriction ensures that MetaCode's development remains independent and original, mitigating risks of data leakage, copyright infringement, and technical dependency. This move also reflects a broader trend among major technology companies to prioritize data isolation and intellectual property protection in AI model training, especially as the competition for advanced AI capabilities intensifies.