Meta Platforms is actively working to reduce its reliance on high-cost AI programming tools from Anthropic and OpenAI, but it currently faces a significant challenge: preventing employee over-dependence on external tools from hindering the development of its own in-house alternatives.
Internal policy documents reveal that Meta Platforms, Inc. has imposed strict limitations on engineers within its Applied AI Engineering division regarding their use of Anthropic's Claude code tool and OpenAI's Codex model. An internal memo even directed several teams to halt certain work involving these two models, driven by corporate fears that content generated by third-party models could potentially seep into Meta's own model training data. The documents warn that such an occurrence could trigger a serious escalation of disputes with partner companies.
Meta Platforms, Inc. was once one of the largest customers for the Claude code tool. Earlier this year, the company established the Applied AI Engineering team with a core mission to iterate and improve its proprietary code assistant, MetaCode. Key tasks involve building high-quality datasets and designing various programming test questions for engineers to train and evaluate its own large language models for coding. While Meta permits this team to use external AI tools in certain scenarios, it explicitly requires engineers to independently design programming test questions, relying on their own professional expertise to conceptualize solutions, and strictly prohibits directly adopting creative ideas generated by AI.
According to informed sources, this internal policy, issued in May, remains in effect. The stringent restrictions were implemented due to Meta's fear of inadvertently engaging in model distillation—the practice of using the output from a competitor's large model to train its own AI model. Such behavior could potentially violate the user agreements for Claude and Codex.
Model distillation essentially allows a developing company to directly reuse a competitor's substantial investments in data reserves, computing power, and technical R&D, a practice that has been mired in industry controversy in recent years.
Meta Platforms, Inc.'s internal communication documents do not record any specific instances of employees improperly using third-party models. A company spokesperson stated, "We have established comprehensive guidelines that clearly define the boundaries for team use of AI tools, ensuring employees can focus on high-value work in a compliant manner."
Although no violations have been identified, the latest management guidelines indicate that, in the context of intense efforts to iterate AI products and seek commercial returns on massive computing infrastructure investments, company leadership believes it necessary to draw more detailed usage boundaries for employees.
A Major Push to Reduce AI Procurement Costs
As AI-related expenses soar, reducing dependence on external AI tools and shifting development work to the in-house tool MetaCode (formerly DevMate) has become increasingly critical for Meta. According to a recent internal memo, Meta's internal AI-related spending alone is projected to reach tens of billions of dollars this year. Following a company-wide push for AI tool adoption, Meta is now curbing skyrocketing AI usage costs by limiting employee token usage quotas.
Meta Platforms, Inc. allows the Applied AI Engineering team to use third-party AI tools for routine tasks, such as building workflows, organizing code and files, and setting up automated validation test environments for its own AI tools. These are referred to in the guidelines as "test scaffolding construction" and "solution calibration"—essentially building and debugging the entire system used to evaluate model performance.
Even in these compliant scenarios, all AI-generated content must undergo rigorous manual review before being put to use. The guidelines explicitly prohibit using output from external large models to design programming test questions for evaluating in-house models. The document states: "This behavior means engineers completely lose project ownership. We absolutely cannot allow reliance on third-party models to generate task materials."
Another restriction is that engineers may not use AI to search for source code vulnerabilities or to generate task ideas through code analysis. In short, models cannot be involved in determining which business problems need testing.
Furthermore, the guidelines stipulate that if the in-house model being tested has access to container resources, then no AI-generated content may be stored in application containers (which encapsulate the complete environment required for a program to run, including code and various dependency libraries).
Walking a Tightrope of Compliance
Industry experts note that technology companies using competitor models in their R&D processes are akin to constantly walking a compliance tightrope: they must enjoy the development conveniences offered by third-party models while absolutely preventing related outputs from entering their own proprietary systems.
Technology legal scholar and industry consultant Marc Lajeunesse commented that Meta's internal document almost completely maps out the boundaries of this compliance tightrope.
While current U.S. law does not explicitly prohibit model distillation, and AI-generated content is not protected by copyright, major AI labs commonly suspend API access for entities or individuals suspected of distillation practices.
Last year, Anthropic suspended OpenAI's API access to Claude. OpenAI explained at the time that the API calls were for benchmarking Claude's capabilities and safety testing performance, stating that such comparative benchmarking is a standard industry practice for technical evaluation.
An internal memo from Meta last month mentioned that it has become difficult to determine whether content created or modified by employees for assessment tasks was produced by humans or large models.
Meta Platforms, Inc.'s restrictions on Claude and Codex usage stem partly from concerns that third-party model output could indirectly mix into its own training datasets, potentially affecting model training quality and easily sparking disputes with partners. On the other hand, the temporary suspension of tool usage in certain scenarios is also aimed at refining internal control mechanisms, involving reviews of R&D processes by researchers from Meta's AI Research lab and Applied AI department management to ensure all operations meet compliance requirements.
Comments