Meta has placed internal restrictions on employee use of Claude and OpenAI's Codex, according to internal documents reviewed by The Information. The move reflects growing anxiety inside major AI labs about model distillation, a practice where outputs from a powerful model are used to train a competing, often cheaper, system. Meta, which both builds and openly releases its own AI models, appears concerned that its own engineers could inadvertently feed a competitor's advantage.
What the Documents Reveal
The restrictions reportedly limit how Meta employees can interact with Claude and Codex in their day-to-day work. Rather than an outright ban, the internal policies appear designed to prevent large-scale, systematic querying that could generate training datasets. This kind of guardrail is specifically aimed at stopping distillation pipelines from being built using a competitor's frontier model as the teacher. The timing is notable: it comes as Claude's capabilities have expanded significantly across coding, reasoning, and analysis tasks, making its outputs increasingly attractive as training signal.
Key Facts
- Meta's internal documents show restrictions placed on Claude and OpenAI Codex usage by employees.
- The primary concern cited is AI model distillation, where outputs train rival models.
- The policy stops short of a full ban, focusing on limiting systematic or bulk use.
- The restrictions apply internally, not to Meta's external products or open-source releases.
- The report was published by The Information, citing documents it reviewed directly.
Distillation has become one of the more contentious issues in the AI industry over the past year. The concern is straightforward: if engineers at one company can query a competitor's model thousands of times and use those outputs as labeled training data, they gain a meaningful shortcut. This is not a hypothetical threat. Earlier this year, reports surfaced that Alibaba used tens of thousands of fake accounts to extract Claude AI data in what appeared to be a coordinated distillation effort, highlighting just how seriously the risk is taken at Anthropic.
The restrictions reflect a broader industry recognition that model outputs themselves carry intellectual property value, and that competitors with access to enough of them can close capability gaps without building equivalent research infrastructure.The Information
A Competitive Landscape Shifting Fast
Meta's decision to restrict access to Claude in particular speaks to how prominent Claude's model family has become as a professional and coding tool. Anthropic has been aggressively expanding Claude's role in developer workflows, and the competitive pressure is visible in how both sides are responding. Anthropic recently moved to increase capacity for coding use cases, with Claude Code's weekly limits raised 50% in a direct response to OpenAI Codex, signaling that the coding assistant market is a key battleground.
For Meta, the irony is hard to miss. The company releases its Llama models openly, and critics have long argued that open releases themselves enable distillation of other labs' proprietary models at scale. Meta's internal policy draws a line between what it permits externally and what it considers acceptable within its own walls. Whether that line holds as AI tooling becomes more deeply embedded in everyday engineering workflows remains an open question. What the documents make clear is that the industry's biggest players are watching each other's model outputs as carefully as they watch each other's research papers.
