Meta is tightening the rules around how its engineers use Anthropic's Claude Code and OpenAI's Codex. The central concern is not just productivity, but whether output from competing AI systems could find its way into Meta's own training data.
According to internal documents obtained by The Information, Meta has temporarily stopped some work involving these models. The documents point to a larger issue now shaping the AI industry: how companies keep rival model behavior from being absorbed into their own systems.
Why Meta is drawing a line
The concern named in the source is distillation. In this context, distillation means the unauthorized transfer of capabilities from one AI model to another. If engineers use a rival AI tool and that output later becomes part of training material, Meta could face serious problems with partner companies.
An internal memo warned that partner relationships could escalate if model outputs from outside systems leaked into Meta's training data. That makes the issue broader than ordinary tool choice. It is about provenance, responsibility, and how AI-generated work moves through a company.
For Meta, Claude Code and Codex are useful because they are coding tools. But their usefulness also creates risk. The more deeply engineers rely on outside model output, the harder it can become to separate human-created work, internal work, and material produced by a competing AI system.
What engineers are restricted from doing
Meta's policy does not simply say that engineers must be careful. The source says company rules bar engineers from using AI outputs to create test tasks or for code analysis. Human review is still required.
Those limits matter because test tasks and code analysis can shape how internal systems understand software work. If outside model output influences those areas, it may become harder to prove that Meta's own systems were not trained or tuned using material from competitors.
The source does not describe every permitted or prohibited use. What it does make clear is that Meta is trying to create boundaries around where outside AI assistance can enter engineering workflows.
- Claude Code is tied to Anthropic.
- Codex is tied to OpenAI.
- Meta is concerned about rival model outputs entering its training data.
- Human review remains part of the process.
MetaCode and the push to rely less on outside tools
Meta is also building its own coding assistant, MetaCode. That project gives the company another reason to limit reliance on tools from outside AI labs. If Meta can move more engineering assistance into its own systems, it can reduce exposure to the output and policies of competitors.
Cost is also part of the picture. According to an internal memo cited in the source, Meta is on track to spend billions of dollars on internal AI use this year alone. That gives the company a financial reason to rethink how much outside AI support is used inside the organization.
The combination is important. Meta is not only responding to a data-governance problem. It is also trying to manage dependence and spending while developing a coding assistant of its own.
That creates a practical tension for engineers. Outside tools can help with coding work, but the organization now has to weigh that help against training-data risk, partner concerns, and the cost of internal AI usage.
Distillation is becoming an industry flashpoint
The issue is not limited to Meta. The source says distillation is causing friction across the industry. Anthropic recently accused Alibaba of the largest known distillation attack to date, and Elon Musk had to admit in April that xAI had partially distilled OpenAI's models.
Those examples show why AI companies are sensitive about model outputs. Outputs are not treated as disposable text when they can reveal patterns, capabilities, or useful behavior from a competing system. If those outputs are then used to build or improve another model, companies may see that as crossing a competitive boundary.
The source also says the terms of service from OpenAI, Anthropic, and Google explicitly ban using model outputs to build competing systems. That makes Meta's restrictions part of a wider environment in which AI companies are trying to define what responsible use of rival tools looks like.
Meta said it has clear rules for the responsible use of AI tools. The important question now is how companies enforce those rules at scale, especially when AI coding assistants are becoming part of everyday engineering work.
What this means for AI coding tools
The Meta case highlights a new reality for AI coding assistants. Their value is no longer measured only by how much faster they help engineers write or review code. Their outputs also carry policy, cost, and competitive implications.
For companies building their own AI models, the safest workflow is not always the most convenient one. Every prompt, generated answer, code suggestion, test idea, or analysis step can raise questions about where the work came from and whether it can be reused internally.
Meta's restrictions show how quickly AI tools can move from optional productivity aids to governed infrastructure. As companies build their own assistants and spend heavily on internal AI use, they are likely to care more about what outside systems touch, where their outputs go, and who reviews the work before it becomes part of a larger pipeline.
The result is a more cautious phase for enterprise AI adoption. Claude Code and Codex may still be powerful tools, but Meta's response shows that power alone is not enough. In competitive AI development, controlling the path of model output can be as important as the output itself.