Meta restricts the use of Anthropic’s Claude Code and OpenAI’s Codex by its own engineers, according to internal documents. The reason is the concern that outputs from these external AI tools could unintentionally flow into Meta’s own training data – a process known in the industry as distillation.
What Meta restricts according to the documents
According to The Information, which was able to view internal guidelines, the restrictions mainly affect engineers in Meta’s “Applied AI” department, which is responsible for building Meta’s own coding assistant, MetaCode. An internal memo reportedly instructed individual teams to temporarily suspend certain tasks using Claude Code and Codex. As reported by TheNextWeb, the memo warned that outputs from competing models could “leak” into Meta’s training data, potentially triggering “serious escalations with partner companies.” Specifically, AI outputs are not allowed to be used for creating test tasks or for code analysis; human review remains mandatory in any case.
Why distillation becomes a problem for Meta
Distillation refers to a training method where a smaller or cheaper model learns from the outputs of a stronger, usually external model – without directly copying the source code, weights, or training data of the original model. For a company that is simultaneously developing its own models and using competitors’ tools in everyday work, a structural risk arises: If code suggestions, architectural decisions, or debugging logic from Claude Code or Codex inadvertently flow into internal training pipelines, documentation, or synthetic training data, the capabilities of competing models could indirectly transfer into Meta’s own Llama model family. The terms of use of OpenAI, Anthropic, and Google explicitly prohibit using model outputs to train competing systems.
Cost pressure as a second driver
In addition to the distillation concern, the reports also highlight the cost factor: Meta wants to reduce its dependence on the relatively expensive external tools and is reportedly on track to spend billions of US dollars on internal AI usage alone this year. Meta is not alone in this – other large tech companies are also exploring cheaper alternatives in light of rising prices for AI coding tools. Meta stated to The Information that it has established clear rules for the responsible use of AI tools.
Distillation as an industry-wide contentious issue
The topic of distillation is currently causing significant tensions in the industry beyond Meta’s individual case. According to a letter viewed by CNBC to the US Senate, Anthropic accused the Chinese technology company Alibaba of conducting over 28.8 million interactions with Claude through around 25,000 fraudulently created accounts between April 22 and June 5, 2026, to specifically extract its advanced capabilities in software development and agentic reasoning. Anthropic described this as the largest known distillation attack on its company to date. Alibaba did not comment in detail on the allegations, according to CNBC.
The issue also recently attracted attention at a US competitor: In the legal dispute between Elon Musk and OpenAI, Musk reportedly admitted under oath during his testimony in April 2026, according to Forbes, that his company xAI had “partially” distilled OpenAI’s models to train its own systems. When directly asked by an OpenAI lawyer, Musk initially gave a vague answer, stating that AI companies generally distill each other’s models, before specifically confirming the practice for xAI.
Assessment
The Meta case illustrates a dilemma that an increasing number of AI companies must face: In order to develop competitive tools in-house, engineers often rely on the already mature products of competitors – thereby creating the very dependency that is supposed to be reduced. At the same time, the series of allegations surrounding Alibaba and xAI makes clear that distillation is no longer just a theoretical risk, but leads to tangible diplomatic, legal, and political disputes between the major AI providers.
