Openwashing: most AI systems are actually closed, experts say

Sharing source code with the public should, in theory, spur innovation and democratize the use of tech tools. But in the field of AI, “open” models are hardly that, especially ones developed by major tech companies.

Meta sounded really proud when it introduced its next generation of open-source large language models (LLMs) under the name Llama 3 back in April and said they were “the best of their class, period.”

The company has always stressed that it is releasing the Llama models largely free of charge for developers to use. The hope is that the strategy will pay off in the form of more innovative products.

Even the US government said it loved open-source. In July, the Department of Commerce issued a report in support of “open-weight” generative AI models precisely like Meta’s Llama. These models have now been opened up for US national security applications.

However, in a new paper, a group of AI researchers and experts say that, actually, claims about “open” AI often lack precision, frequently eliding scrutiny of substantial industry concentration in large-scale AI development and deployment.

Moreover, the analysis published in Nature says powerful actors incorrectly – and possibly deliberately – apply a concept of “open” imported from free and open-source software to AI systems.

The problem is that most AI models today – whether “open” or “closed” – are, of course, controlled by Google, Meta, and other tech giants, which have the resources to guide the evolution of AI to meet their financial goals.

The “rhetoric of openness is frequently wielded in ways that exacerbate the concentration of power” in large tech companies, write David Widder at Cornell University, Meredith Whittaker at Signal Foundation, and Sarah West at AI Now Institute.

According to the experts, old-school open-source software – such as free word processors similar to Microsoft’s – could indeed level the playing field for the little guys. Software engineers could also play with code, discover its flaws, and make it more usable and secure.

But with AI, the story’s different, the paper says – simply because the LLMs are built in a different way, with numerous layers of interconnected artificial “neurons.” The underlying training data isn’t usually shared, and that’s a problem.

“Understanding these systems’ internal processes isn’t straightforward. Unlike traditional software, sharing only the weights and code of an AI model, without the underlying training data, makes it difficult for other people to detect potential bugs or security threats,” the team writes.

“This means previous concepts from open-source software are being applied in ill-fitting ways to AI systems.”

For example, Llama 3, although described as open, only allows people to build on their AI through an API without sharing the underlying code, or to download just the model’s weights – the strengths of those artificial neurons – to tinker with, with certain restrictions on their usage.

“This is ‘openwashing’ systems that are better understood as closed,” write the authors of the paper, adding that such AI models may actually be a barrier to the democratization of AI.

“The pursuit of openness on its own will be unlikely to yield much benefit.”

According to the experts, the entire cycle of AI development – from setting up, training, and running AI systems to their practical uses and financial incentives – should be considered when building open AI policies.