Meta criticised for calling its AI models ‘open-source’

Stay informed with free updates

Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.

Meta has been criticised for calling its artificial intelligence models “open-source” by the group that has spearheaded open-source technology in the software world for the past 25 years.

The social media company is “confusing” users and “polluting” the term open-source by using it to describe its Llama family of large language models, said Stefano Maffulli, head of the Open Source Initiative. The body coined the term open-source in the late 1990s and has been seen as the guardian of the concept ever since.

Speaking in an interview with the FT, he said this was “extremely damaging” at a time when bodies such as the European Commission were seeking to support true open-source technologies that are beyond the control of any particular company.

Llama, which Meta says has been downloaded more than 400mn times, has become the most popular among a wave of supposedly “open-source” AI models that have risen to challenge the leading proprietary systems such as OpenAI’s GPT-4 and Google’s Gemini.

However, most of these, including Llama, stop short of full openness, preventing the kind of experimentation and adaptation in AI that open-source brought to the software world.

By creating confusion about which models are truly open, Meta risks hampering the long-term development of forms of AI that are steered and controlled by users rather than being tightly controlled by a handful of tech companies, according to the OSI chief.

Maffulli said Google and Microsoft had dropped their use of the term open-source for models that are not fully open, but that discussions with Meta had failed to produce a similar result.

Meta said it was “committed to open-source AI” and that Llama “has been a bedrock of AI innovation globally”.

It added: “Existing open-source definitions for software do not encompass the complexities of today’s rapidly advancing AI models. We are committed to keep working with the industry on new definitions to serve everyone safely and responsibly within the AI community.”

Many backers of fully open-source technologies still credit Meta’s more limited Llama models with having broken the stranglehold of a handful of big US AI companies and opening the generative AI market to more competition.

Meta’s models have been “a breath of fresh air” for developers, said Dario Gil, head of research at IBM, giving them an alternative to what he called the “black box” models from the leading AI companies.

However, transparency is limited. Meta lets developers download its Llama models free of charge, but the only technical details it publishes are the weights, or “biases”, that influence how the model responds to particular prompts.

In addition, the licence under which Llama has been released does not conform to the open-source definitions recognised by the OSI since it does not allow free use of the technology by Meta’s largest rivals.

Other tech groups, such as French AI company Mistral, have taken to calling models like this “open weight” rather than open-source.

“Open weight [models] are great . . . but it’s not enough to develop on,” said Ali Farhadi, head of the Allen Institute for AI, which has released a fully open-source AI model called Olmo.

Developers using models such as Llama can’t see how they were developed or build on top of them to create new products of their own, as happened with open-source software, he added.

To comply with the OSI’s definition of open-source AI, which is set to be officially published next week, model developers need to be more transparent. Along with their models’ weights, they should also disclose the training algorithms and other software used to develop them.

OSI also called on AI companies to release the data on which their models were trained, though it conceded that privacy and other legal considerations sometimes prevent this.

Maffulli said bodies such as the European Commission have sought to give special recognition to open-source in their regulations to encourage its use in widely used technology standards.

If companies such as Meta succeed in turning it into a “generic term” that they can define for their own advantage, they will be able to “insert their revenue-generating patents into standards that the EC and other bodies are pushing for being really open,” he warned.