Internal Documents Reveal Meta Restricts Claude and Codex Usage to Guard Against Model Distillation Risks
Claire Weston
Internal documents obtained by The Information reveal Meta has sharply restricted its AI teams from using Anthropic's Claude Code and OpenAI's Codex, driven by fears that model distillation — outputs leaking into Meta's own training data — could trigger contract violations. The move signals that IP tensions among top AI firms have escalated from policy language to operational lockdown.
What is distillation, and why does Meta fear it?
Distillation — training your own model on a competitor's outputs — lets developers free-ride on others' massive spending in data, compute, and research.
OpenAI, Anthropic, and Google all explicitly prohibit using model outputs to build competing systems in their terms of service.
This means → if code generated by Claude or Codex seeps into Meta's training pipeline, Meta could breach those contracts directly. Internal documents warn this risks "serious escalation with partner companies."
Where exactly is the red line?
Meta still allows engineers to use external AI tools for routine tasks: configuring workflows, organizing code and documentation, building test infrastructure for internal tools.
But anything touching training-data generation is off limits — engineers must design coding challenges themselves, relying on their own technical judgment, not AI-generated concepts.
In plain terms = you can use outside AI for housekeeping, but anything that might feed into Meta's own models must be "handmade."
How big is the cost pressure?
A recent internal memo shows Meta's spending on internal AI usage alone will reach billions of dollars this year.
The company is imposing per-employee token-usage caps and accelerating migration to its in-house coding assistant MetaCode (formerly DevMate).
This means → restricting external AI tools is not just a compliance call — it is also about cutting a massive procurement bill. Meta was previously one of the largest enterprise customers of Claude Code.
How heated is the distillation debate industry-wide?
Anthropic this month accused Alibaba of carrying out the largest distillation attack to date against Claude, targeting agentic reasoning, software engineering, and long-horizon task capabilities.
Elon Musk acknowledged in April courtroom testimony against OpenAI that his xAI had "partially" distilled OpenAI models.
This reflects a shift: IP disputes among top AI companies have moved from "the ToS says no" to active litigation and internal isolation measures, raising the compliance bar for enterprise procurement of external AI tools.
Content is for reference only, not financial advice.