The term “open-source AI” has become one of the most frequently used, and frequently misused, phrases in the artificial intelligence industry, with companies routinely applying the label to products that fall well short of a formal definition established by the organization that has governed open-source standards for more than two decades.
According to the Open Source Initiative, the nonprofit widely recognized as the authoritative steward of open-source principles, an AI system genuinely qualifies as open source only when it grants users four specific freedoms: the freedom to use the system for any purpose without seeking permission, to study how it works and inspect its components, to modify it to suit specific needs, and to share it, with or without changes, for any purpose.
Why AI complicates a decades-old definition
Open-source software has a comparatively simple test: anyone must be able to access, understand, modify, and redistribute the underlying code. AI systems are structurally more complicated, because a functioning AI model depends on far more than code alone. IBM’s technical explainer notes that AI systems encompass the model itself, the datasets used during training, the model’s weights and parameters, and multiple layers of supporting source code, including code for filtering training data, training and testing the model, and running inference. All of these components must be openly available for a system to satisfy the open-source standard.
That complexity led the Open Source Initiative to publish a dedicated standard, the Open Source AI Definition, on October 28, 2024, following a lengthy multiyear process involving monthly meetings with open-source developers worldwide. According to a practical guide published by Moesif, OSAID 1.0 requires three categories of disclosure working together: the source code used to train and run the system under an OSI-approved license, the model’s parameters, meaning its trained weights, released under terms permitting free use and modification, and detailed information about the training data sufficient for a technically skilled person to substantially recreate the system, even if the raw dataset itself isn’t distributed.
The gap between marketing and the standard
Despite that formal definition, most models publicly described as “open source” satisfy only a fraction of its requirements. A recent analysis published by Tech Policy Press argued that “open-source” has become a marketing register rather than a rigorous standard, noting that when researchers apply the strict criteria, requiring architecture, training code, model weights, and training data to all be publicly available under licenses permitting unrestricted use, “very few models qualify.”
That gap has a name within the industry: “open-washing,” a term describing the practice of labeling a product as open source when it does not meet the accepted definition. Tech Policy Press’s analysis pointed to this practice as increasingly common across the AI sector, as companies capitalize on the reputational and marketing benefits associated with open-source branding without accepting the transparency obligations that traditionally accompany it.
Open weights: A related but distinct concept
A lot of the confusion arises from mistaking “open-source AI” with a more limited, more widespread approach called “open weights. The New York Times, in a recent explainer, described weights as the calculations and internal guidelines that govern how an AI system processes information and generates outputs. When a company publishes those weights publicly, users can download and modify how the system behaves, tailoring it toward specialized applications such as medical research or cybersecurity or emphasizing particular types of information.
Critically, however, open weights do not require full disclosure of the code that operates the system, nor do they necessarily reveal anything about the data used to train the model in the first place. A guide published by Tech Insider described open weights as meaning “the trained parameters of a model are published for anyone to download and run,” typically through platforms like Hugging Face, while explicitly distinguishing this from full open source in the traditional software sense, “since training data and code aren’t always released alongside the weights.”
A geopolitical dimension
The distinction between open weights and true open source has taken on added significance amid a broader industry shift toward openly released models, much of it driven by competition between American and Chinese AI developers. An analysis published by Simply Wall St noted that American large language models have largely been built around closed, subscription-based business models, with companies such as OpenAI, Anthropic and Google retaining control over their models’ foundational inner workings.
Chinese firms, by contrast, pivoted toward open-weighted models following the emergence of DeepSeek in January 2025, a strategy the analysis says allows Chinese companies to avoid some of the costs of building a full commercial AI business while helping offset restrictions they face in accessing advanced computer chips. That dynamic has made open-weight releases a recurring flashpoint in discussions of both AI competitiveness and national security.
Academic and regulatory scrutiny
The ambiguity surrounding open-source AI has also drawn sustained attention from computer science researchers and regulators. A paper published in Communications of the ACM examined what the authors describe as competing perspectives on openness in foundation models, noting that some view AI transparency as central to innovation while others see it as introducing distinct security risks not present in traditional open-source software.
Wikipedia’s overview of the topic similarly frames open-source AI as promoting “a collaborative and transparent approach to AI development,” specifically because full disclosure of data, code and parameters allows outside parties to create a substantially similar result independently, the same replicability standard long central to open-source software more broadly.
Real-world examples illustrate the spectrum
Recent product launches illustrate how varied “open” AI releases can be in practice. Nvidia’s PersonaPlex voice model, for instance, ships its code under the MIT license while distributing its trained weights under Nvidia’s own Open Model License, an arrangement that permits commercial use without royalties but does not necessarily disclose full training data details, according to a review published by tbreak.com.
Separately, Tech Insider’s coverage of a wave of nine open-weight model launches within a 12-day span in July 2026, led by a 975-billion-parameter model from Thinking Machines, noted that even models released under permissive licenses like Apache 2.0, which allow commercial use, fine-tuning and redistribution without royalty payments, still frequently withhold their original training data, placing them squarely in the “open weights” category rather than satisfying the OSI’s full open-source AI standard.
Why the distinction matters
For developers, businesses and policymakers, the difference between genuine open-source AI and merely open-weight releases carries practical consequences. A model that discloses only its weights allows fine-tuning and self-hosting but leaves users unable to fully audit how the system was built or verify claims about its training process. True open-source AI, by contrast, is designed to let anyone “read their inner workings like a book,” as The Conversation‘s software engineering explainer put it, enabling independent verification, reproducibility, and modification at every layer of the system.
As AI companies continue to compete for developer mindshare and public trust through openness claims, the underlying question, whether a given release actually meets the Open Source Initiative’s four freedoms across code, weights and data, or merely echoes open-source language while withholding key components, is likely to remain a central point of scrutiny across the industry for the foreseeable future.
