Open is the new "open," because definition of something can change when adaptation is a must.
In the era of generative AI, the arms race where where tech companies race to create the best of the best generative AI tools, following OpenAI revelation of ChatGPT, have put companies in fierce battle for supremacy.
But to be the best, requires a lot of strategies and efforts.
And tech companies, like Google and Meta, have called their generative AIs as "open," claiming them to be open-source in at least one way.
But according to the Open Source Initiative (OSI), a long-time authority on open-source standards since 1998, a newly-updated definition of "open" is changing what the tech companies are claiming to be.
Here, OSI has introduced a revised definition that has been two years in the making.
In its The Open Source AI Definition – draft v. 0.0.9, the agency suggests that being open source means that others have the "freedoms" to apply changes to a fully functional system and to discrete elements of a system.
According to Mozilla, which has been a significant player in the open-source movement, said that OSI's renewed statement is crucial for not only redefining "open source" in the context of AI but also for shaping the future impact of the technology on society.
Nik Marda, Mozilla’s AI governance technical lead, points out that Meta's LLaMA 3 and Google’s Gemma models don’t meet this new definition.
The primary issue is the restrictions these companies place on how their models can be used, which violates the OSI's new criteria for open-source AI.
Meta , for example, has positioned open-source as a core element of the company’s strategy, arguing it benefits both Meta and the wider world.
According to the company, it defines open-source models as those with publicly released weights under permissive licenses.
Marda noted that Meta's definition is quite narrow, because its definition of open source doesn't allow others have the full access necessary to truly test and build upon its models.
"In the past, the lack of a clear definition allowed companies to claim their AI was open source even when it wasn’t," Marda explains.
"Now, many of the large commercial actors’ models won’t qualify."
Experts also argue, saying as long as there is a restriction, a product can never be considered open.
According to a paper by MIT Sloan professor Pierre Azoulay and co-authors, for example, tight control over access to specialized infrastructure and capabilities will likely result in a concentrated generative AI market controlled by a few key players. And this is exactly what tech companies want.
The lesser the competition, the better.
Following the announcement, Meta has not recognized the OSI’s new definition as the industry standard, and Google declined to comment on the matter.
"This is still emerging technology, and there isn’t a single, universal definition for 'open source' AI," a Meta spokesperson said. "Like OSI, Meta is committed to open-source AI, and we’ll continue collaborating with the industry on these terms."
The debate over what qualifies as open-source AI isn’t new, and OSI’s updated definition is just the latest chapter.
Weighing in to the situation, according to The Linux Foundation, the non-profit organization that promotes the growth and development of open-source software projects, the many so-called "open-source" models have been released under custom licenses with vague terms, a trend it referred to in April 2024 as "open-washing."
This, they argue, could erode the principle of openness, which is meant to promote the free exchange of knowledge to enable innovation and collective progress.
For its part, The Linux Foundation addresses this by proposing a tiered framework for openness, rather than a simple "open" versus "closed" categorization.