Open-source AI has become one of the most contested ideas in technology. Companies want the credibility and ecosystem benefits that come with openness, while researchers, lawyers, activists, and policymakers are still debating what the term should actually guarantee.
The result is a high-stakes definition fight. If the meaning stays vague, powerful companies can use the language of openness while keeping control over the ingredients that matter most.
Why the definition matters now
Meta has pledged to create open-source artificial general intelligence, and Elon Musk is suing OpenAI over its lack of open-source AI models. At the same time, more companies are presenting themselves as open-source champions.
On paper, open-source AI suggests a system that more people can inspect, adapt, and improve. That could support faster innovation, more transparency, and greater user control over systems that may affect many parts of life.
But AI does not map neatly onto the older software model. In conventional software, the key question is usually whether developers can access and modify the source code. With AI models, the important components can include the trained model, the training data, preprocessing code, training code, model architecture, and other implementation details.
That is why the Open Source Initiative, or OSI, has assembled a 70-strong group of researchers, lawyers, policymakers, activists, and representatives from companies including Meta, Google, and Amazon. The organization, founded in 1998, maintains the Open Source Definition for software and is now trying to build a working definition for open-source AI.
The license problem
The lack of a settled standard has not stopped companies from using openness language. Meta made Llama 2 freely available last July and described it as open source. Google released Gemma last month as freely accessible and described it as “open” rather than “open source.” Stability AI and Aleph Alpha have also released models described as open source, and Hugging Face hosts a large library of freely available AI models.
Still, availability is not the same as open source. Both Llama 2 and Gemma come with licenses that limit what users may do with the models. That conflicts with a core clause of the Open Source Definition, which rejects restrictions based on use cases.
OpenAI has moved in the opposite direction from broad disclosure. It has shared fewer details about its leading models over time, citing safety concerns. A spokesperson said, “We only open-source powerful AI models once we have carefully weighed the benefits and risks, including misuse and acceleration.”
The disagreement shows why the label matters. A model can be downloadable, usable, and adaptable in some ways while still failing to meet the expectations many people associate with open source.
Training data is the biggest fault line
The most difficult dispute is data. Major AI companies have released pretrained models without releasing the datasets used to train them. For people who favor a stricter standard, that missing data prevents meaningful study and modification, which should disqualify the models from being called open source.
Others argue that full access to the original training data is not always necessary. A description of the data may be enough for some forms of inspection, and many pretrained models can be adapted through fine-tuning on smaller, application-specific datasets.
Roman Shaposhnik, CEO of open-source AI company Ainekko and vice president of legal affairs for the Apache Software Foundation, points to Llama 2 as an example. Even though Meta released only a pretrained model, developers have downloaded it, adapted it, and shared modifications around it.
Zuzanna Warso, director of research at Open Future, argues that data remains central. If openness is meant to challenge concentrated power, the openness of the data becomes especially important.
The business incentive is clear. High-quality training data is a major bottleneck in AI research and a competitive advantage for large firms. Sharing pretrained models while withholding training data lets companies gain some benefits of openness without giving up a key asset.
Business, regulation, and open washing
The word open source carries strong positive associations, which makes it useful for public relations. Warso describes superficial openness as “open washing.”
There are also practical business reasons companies may want the label. Economists at Harvard Business School recently found that open-source software has saved companies almost $9 trillion in development costs by letting them build on free, high-quality software rather than creating everything themselves.
Open source can also help major companies build ecosystems around their products. The article points to Google’s open-sourcing of Android as a classic example of how open software can support a dominant position. Meta’s Mark Zuckerberg has also said in earnings calls that open-source software often becomes an industry standard, and that when companies build with Meta’s stack, it becomes easier to integrate new innovations into Meta’s products.
Regulation adds another incentive. Warso points to the EU’s newly passed AI Act, which exempts certain open-source projects from some of its stricter requirements.
These incentives make the definition of open-source AI more than a technical matter. It can shape who gets lower compliance burdens, who builds developer ecosystems, and who benefits from the reputation of openness.
What openness is supposed to achieve
Even a strong definition may not fully level the AI field. Sarah Myers West, co–executive director of the AI Now Institute, coauthored a paper published in August 2023 that examined the lack of openness in many open-source AI projects. The paper also highlighted deeper barriers: the large amounts of data and computing power needed to train advanced AI systems.
Myers West argues that the AI community needs to be clearer about the goal. Is the purpose safety, academic research, competition, transparency, autonomy, or something else? The answer changes what kind of openness matters most.
The OSI’s draft definition mentions autonomy and transparency, but it also treats questions about “ethical, trustworthy, or responsible” AI as out of scope. Stefano Maffulli, OSI’s executive director, says the open-source community has historically focused on frictionless sharing rather than on what software should be used for.
Others are experimenting with licenses that include use restrictions. Responsible AI Licenses, introduced in 2022, allow developers to block specific uses they consider inappropriate or unethical. Danish Contractor and colleagues found in a recent analysis of Hugging Face that 28% of models use RAIL.
Google’s Gemma license follows a similar path by prohibiting various “harmful” uses. The Allen Institute for AI has also developed ImpACT Licenses, which restrict redistribution of models and data based on potential risks.
Luis Villa, cofounder and legal lead at Tidelift, sees experimentation as inevitable because AI differs so much from traditional software. But he warns that too many incompatible “open-ish” licenses could weaken the smooth collaboration that made open source valuable in the first place.
That is the core challenge. Without a shared standard, the industry may define openness for itself. With a standard that is too strict, major players may ignore it. The future of open-source AI depends on whether the community can draw a line that is clear, meaningful, and widely adopted.