Meta restricts employee use of Claude and Codex to guard against 'Model Distillation' violation

According to Dynamic Beating monitoring, Meta is currently imposing strict limitations on its AI engineering team from using Anthropic's Claude Code and OpenAI's Codex, even going as far as requesting a pause on certain related tasks to enhance compliance reviews. Meta is concerned that employees are overly relying on external tools to develop internal alternatives, leading to competitor models' outputs seeping into their own training data. This "model distillation" practice may potentially violate competitors' terms of service and trigger severe conflicts with partners.

Currently, Meta only permits employees to use external AI for setting up workflows, organizing code, and building testing infrastructure as part of routine work, with all outputs requiring manual thorough review. However, the use of external model-generated programming challenges to assess their internal models is strictly prohibited, as is leveraging external AI to discover code vulnerabilities or brainstorm testing tasks.

This move is also part of Meta's efforts to reshape its internal toolchain, control its internal AI operations costs reaching up to tens of billions of dollars, by promoting its in-house coding assistant MetaCode (formerly known as DevMate).