AI Industry Clashes Over 'Open Source' Definition: Weights vs. Data

A fundamental conflict has emerged in the artificial intelligence industry over whether models that release only their weights can legitimately be called "open source." The Open Source Initiative (OSI) and academic experts argue that this practice is a misnomer, creating a divide between true transparency and what they term "open distribution."

At the center of the dispute are "open weights," the final numerical parameters of a trained neural network. These parameters allow users to run and fine-tune models on their own hardware without relying on a proprietary API. However, the OSI maintains that weights alone provide only a fraction of the information required for full accountability. To address this, the OSI released the Open Source AI Definition (OSAID 1.0), which establishes the specific requirements a model must meet to be considered truly open source.

The Transparency Gap

Historically, open source referred to the accessibility of software source code. In the era of Large Language Models (LLMs), the "source" is more complex, consisting of the code, the model architecture, and the weights derived from massive datasets. While many companies release the weights, they frequently keep the training data and the curation process secret. This has led to accusations of "open-washing," where companies use the prestige of the open-source label for marketing purposes without providing the actual transparency the term implies.

James Landay, director of the Stanford Institute for Human-Centered AI (HAI), argues that without access to the training data or a thoroughly documented, auditable account of it, the community cannot see how a model was built or why it behaves as it does. "Open weights are progress," Landay stated, "But... That's not an open model. That's open distribution."

Why the Distinction Matters

This semantic battle has significant implications for trust, safety, and scientific progress. Without knowing the training data, users and researchers cannot independently identify inherent biases, copyright infringements, or potential data leaks. If "open source" becomes a hollow marketing term, it undermines the ability of the community to verify and improve AI systems. Critics warn that this could effectively lock the future of AI into a few opaque corporate silos, even if those systems appear open on the surface.

The Path Toward a Standard

To bridge the gap between corporate interests and open-source purists, the Linux Foundation's Mike Dolan submitted the Open Model, Data, and Weights (OpenMDW) license to the OSI. This proposed license aims to cover the architecture, data, and weights under a single agreement, potentially providing a legal and technical framework for true openness.

What remains to be seen is whether the industry's largest players will adopt these stricter standards. For now, the tension persists between those who view the release of weights as a sufficient contribution to the commons and those who believe that without the data, the "source" remains closed.