The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence

"Open source" claims are often misleading: The paper finds that many AI model producers claim their models are "open source" while using restrictive licenses that prevent full understanding, auditing, and reproducibility—highlighting a significant gap between marketing claims and actual openness in the AI industry.

Most frontier AI models lack critical components: The research identifies that most commercial AI models are missing necessary components (code, data, and documentation) needed for full understanding, auditing, and reproducibility, making independent verification of safety and performance difficult.

A practical three-tiered classification system can guide better practices: The Model Openness Framework provides a ranked classification system that specifies exactly which components must be released under open licenses at each tier, giving both AI developers and users a clear standard for what "open" should mean.

Consumers need tools to identify truly usable models: The Model Openness Tool offers a user-friendly way for people to evaluate whether AI models can actually be permissively used, studied, modified, and redistributed—helping non-technical stakeholders make informed choices about which models they can safely adopt.

Openness and completeness should be core to responsible AI: The framework positions transparency and reproducibility as fundamental principles for responsible AI development, not optional add-ons, and provides practical guidance for establishing these as industry standards.