(Credit: NurPhoto / Contributor / NurPhoto via Getty Images)
Open-source AI models from Google, Meta, and others may actually be quite closed, according to a recently updated definition of the term.
The lengthy new definition comes from the Open Source Initiative (OSI), which has considered itself the steward of the open source definition since its founding in 1998. The OSI has been working on an updated definition for two years.
Mozilla endorses the revised definition as "critical not just for redefining what 'open source' means in the context of AI [but for] shaping the future of the technology and its impact on society."
Meta's Llama 3 would not be considered "open" under the new definition, says Nik Marda, Mozilla's technical lead of AI governance and former chief of staff for the White House Office of Science and Technology Policy's Technology Division. Google's Gemma models also do not make the cut because they have limits on how people can use them, which is not permitted under the new definition, he says.
"The lack of a precise definition in the past has made it easier for some companies to act like their AI was open source even when it wasn’t," Marda tells PCMag. "Many – if not, most – of the models from the large commercial actors will not meet this definition."
A loose definition of open source could undermine consumer products and services that use those systems, giving companies a license to change how the system works and restrict access if necessary to protect their bottom line.
This could lead to "disrupted services, subpar performance, and more expensive features in the apps and tools that everyone uses on their phones, in the workplace, and across society," Marda says. We saw this in July when security researchers discovered vulnerabilities in Apple devices due to flaws in open-source code.
Meta does not acknowledge OSI's definition as the new standard. Google declined to comment.
"This is very new technology, and there is no singular, global definition for 'open source' AI," a Meta spokesperson tells PCMag. "Meta – like OSI – is committed to open-source AI. We are committed to keep working with the industry on these terms."
The definition of open-source AI has been an ongoing matter of technical debate, which started well before OSI released the new definition.
"Models purported as 'open-source' frequently employ bespoke licenses with ambiguous terms," The Linux Foundation said in an April 2024 post. "This 'open-washing' trend threatens to undermine the very premise of openness – the free sharing of knowledge to enable inspection, replication, and collective advancement."
The Linux Foundation proposes a tiered approach to openness rather than a binary "open" or "closed" designation.
(Credit: The Linux Foundation)AI writer Sriram Parthasarathy also puts Llama 3 on a spectrum of openness. "It’s not as free as some open-source software but not as restricted as other AI models," he says. "In the end, Llama 3.1 is fairly open, but with some conditions."
Meta CEO Mark Zuckerberg put open source at the center of the company's strategy, calling it "good for Meta" and "good for the world." He defines open-source models as those "whose weights are released publicly with a permissive license,” and cites Llama as an example in an opinion piece published in The Economist last month.
According to Marda, Zuckerberg presents "a very narrow definition for open source AI — not one that actually provides the access needed for others to truly test and build fully upon it."