The technology landscape fundamentally shifted in November 2022 when ChatGPT woke the world up to the power of generative AI. As artificial intelligence rapidly integrates into every facet of digital infrastructure, the open source community faces a critical question: what does "open source" actually mean in the age of AI?
Defining Open Source AI is not a simple translation of existing software rules. It presents unique, practical challenges that require a fundamental rethink of how we evaluate digital sovereignty, transparency, and collaboration.
The Paradigm Shift: Code vs. Weights
To understand the challenge, we must first recognize the key differences between "traditional" software and AI.
● Traditional Programming: Developers write explicit, step-by-step instructions telling the computer exactly how to perform a task. The logic is entirely predefined.
● AI Model Training: Developers provide data and a model architecture, allowing the system to learn patterns directly from the data. The logic dynamically emerges from the encoded knowledge rather than being explicitly programmed for every specific rule.
Traditional Open Source Software (OSS) focuses heavily on human-readable code and guarantees four essential freedoms: the freedom to run the program for any purpose, study and change how it works, redistribute copies, and distribute modified versions. Applying this paradigm directly to AI systems presents unique challenges because AI relies on vast datasets and opaque model weights, which are components that fall outside the traditional scope of OSS.
Furthermore, "modification" means something entirely different in the AI space. In traditional OSS, you modify a program by editing its code. In AI, meaningful modification often requires retraining or fine-tuning the model, which necessitates access to the model weights, the training code, information about the data, and substantial computational resources.
The Data Dilemma
Training data is critical for understanding AI capabilities, evaluating bias, and ensuring reproducibility. However, sharing raw data faces significant legal constraints, such as privacy laws and copyright, along with ethical and competitive hurdles that simply do not exist for standard software code.
This creates a massive hurdle for reproducibility. While OSS compilation is usually deterministic, training large AI models is often incredibly difficult to reproduce exactly due to algorithmic randomness, hardware variations, and data ordering.
Frameworks for Openness
As the industry grapples with these complexities, three main approaches have emerged to establish a definition for Open Source AI:
1. Criteria Scoring Approach: This method evaluates an AI release by assigning a numerical score based on how well it meets a predefined set of criteria. A prominent example is Stanford’s Foundation Model Transparency Index, which provides 100 fine-grained indicators to systematically define transparency across the model's lifecycle.
2. Spectrum Approach: This classifies an AI release into one of several categories along a gradient ranging from "fully closed" to "fully open." An example is the Model Openness Framework (MOF), which categorizes openness across different classes.
3. Binary Approach: This method classifies an AI release into one of only two mutually exclusive categories: strictly "open" or "closed." The Open Source Initiative’s (OSI) Open Source AI Definition (OSAID) is the leading example of this.
OSAID attempts a pragmatic compromise by balancing open principles with legal realities regarding training data. It mandates full code availability, accessible model parameters, and detailed information about the training data (such as scope, characteristics, and processing), but it does not strictly mandate the release of the complete original training dataset in all cases.
The French Intervention: Categorizing "Open Weights"
While frameworks like OSAID attempt a pragmatic binary compromise, European institutions, particularly in France, reject the strict division of "fully open" versus "fully closed" altogether. Addressing the distinct differences between traditional code and model weights discussed earlier, they propose a tiered, safety-anchored approach.
The French data protection authority (CNIL) and the Pôle d'Expertise de la Régulation Numérique (PEReN) assert that making model parameters available does not automatically equate to "open source." Instead, they formally recognize the concept of "open-weight" models: systems whose numerical parameters are freely available for public download and adaptation, explicitly acknowledging that underlying training data or processing code may remain withheld.
This nuanced French perspective heavily influenced the May 2026 G7 framework, which categorizes AI system releases based on specific criteria to mandate accurate labeling:
| AI System Designation | Licensing and Asset Availability Criteria | Regulatory and Technical Implications |
| Weights Available | Proprietary licensing applied to model parameters | Denotes models where weights can be downloaded or accessed, but usage, modification, or commercial deployment is restricted by vendor terms |
| Open Weights | Model parameters distributed strictly under an approved Open Source license | Allows unrestricted fine-tuning and commercial use of the weights, through training data may remain proprietary. |
| Open Source AI with Open Data | All assets (weights, deployment code, training code, and full training data) are released free of charge under an open source license. | Represents the highest tier of transparency, aligning fully with traditional OSS freedoms and allowing deterministic reproducibility. |
By clearly delineating "open weights" from true "open source AI," this framework provides a vital regulatory tool to combat a growing problem in the tech industry: "Openwashing".
The Danger of "Openwashing"
Because definitions of Open Source AI are still evolving and complex, the industry is currently experiencing a wave of "Openwashing." This refers to deceptively marketing AI systems as "open source" when they only release model weights and do not meet genuine openness criteria.
The motivations behind Openwashing are multifaceted:
● Marketing and PR: Companies seek to capitalize on the positive, collaborative image of open source.
● Regulatory Arbitrage: The EU AI Act provides exemptions for open-source models, allowing them to bypass some administrative requirements, such as maintaining extensive internal technical documentation. This reduction in compliance overhead creates a powerful incentive for companies to use Openwashing to exploit these legal loopholes.
● Competitive Strategy: Companies can undercut rivals and gain community feedback without true reciprocity.
Openwashing often relies on a "release-by-blogpost" strategy, utilizing cherry-picked wording to create the appearance of openness without enduring rigorous peer review. This ultimately erodes public trust, misleads stakeholders, obscures potential risks, and significantly complicates the regulatory landscape. Currently, only a few systems top true openness rankings, with all major commercial players occupying the lower tiers.
Moving Forward
We must be vigilant and advocate for clarity in the AI ecosystem. It is essential to critically evaluate any "openness" claims by jointly reviewing an AI model's license, underlying components, and governance structures to avoid hidden risks. True open source AI enhances flexibility and improves security through community oversight. By translating technical transparency into business value through demanding clear model information, data origins, and documented methods, we can better manage risk, meet evolving compliance standards like the EU AI Act, and build enduring trust in digital solutions.
About the Author
Alfonso Cancellara is a Senior Technical Account Manager for OpenShift at Red Hat, specializing in enterprise open source software solutions. He focuses on bridging the gap between foundational technology concepts and practical implementation, guiding organizations through the complexities of cloud-native architecture, Open Source ecosystems, and the rapidly evolving landscape of Artificial Intelligence.