The Open Source AI Definition is not fit for purpose

Will it be effective?

The stated aim of this definition is to identify AI systems made available under terms which allow free (libre) Use, Study, Modification and Sharing. I do not think the definition is effective in this aim.

The definition of the "Preferred form to make modifications to machine-learning systems" specifically excludes the need to make the full training data available under a free or open license. Instead, it requires the provision of "Data information". At the risk of stating the obvious, "data information" is not "data". Possessing a description of the training data and how it was obtained does not guarantee that the recipient has the legal or technical ability to exactly re-create the training data. And if the training process for an AI model is followed using a similar but not exactly identical set of training data, the resulting model weights will differ. This is not a purely academic concern - without the ability to exactly re-run the training process and arrive at the exact same AI model, we cannot say that we have the preferred form to make modifications to the system. We can't answer questions like "How would the resulting model differ if we exclude some subset of the training data" or "How would the resulting model differ if we extend the training data", etc, if re-training the model also introduces other differences due to our inability to start with the exact same set of training data.

That is to say - the effective rights to study and modify an AI system require access to the complete corresponding training data.

This view was shared by most of the folks who contributed to the discussion session at OggCamp 2024. There was one counter argument that the training data was similar to "working out" or research notes made while developing software, and that software can be Open Source without the need to share those private notes. There was disagreement with this analogy from others in the room.

Why is training data excluded?

Given the above, and that this view seems to be widely shared, why would the need to share training data under an open license be explicitly excluded from the Open Source AI Definition?

The current AI hype cycle is focused on Large Language Models (LLMs) and other large-scale generative AI models. These models require vast amounts of data to train - so much data that no combination of truly open data sets can provide even a fraction of the required data volume. Instead, training data is scraped from the internet at large without regard to the consent of authors, artists or users. Even if training AI models on this data is considered fair use under copyright legislation, distributing this data set in its full and original form could never qualify as fair use.

If the OSI stated that complete corresponding training data must be shared under an open license for an AI System to be considered Open Source, no LLM or large-scale generative AI model could ever meet this requirement. So no LLM could be an Open Source AI system.

And if the OSI argued that no LLM could be considered Open Source, they would have to contend with a fear of missing out (FOMO) or fear of becoming irrelevant if they aren't able to somehow attach themselves the current AI hype cycle.

It's worth noting that there are no freestanding community projects developing such models - the training costs in both data and compute resources are simply too high for such community projects to exist independently from a corporate effort to train an AI model. These are mega-projects on the scale of a nuclear power station or an aircraft carrier.

Smaller and more special purpose AI models do exist and do have freestanding communities, but they are not the subject of the current hype cycle.

I would argue that this definition is targeted at the AI mega-projects, not at the developers of smaller AI models. It is a direct attempt to stop companies like Meta misusing the term "open source" to describe their AI systems - a thing that they are simply not going to do. They don't respect authors, artists or user consent, so why will they respect an Open Source AI Definition?

How does this compare to the original Open Source definition?

The original Open Source definition back in 1998 built on existing community efforts, licenses and policies such as the Debian Free Software Guidelines. It packaged these community values and best practices in a business-friendly way so that could reach a new audience - it was an invitation to join an existing and vibrant community.

This Open Source AI definition attempts to define such a community into existence where I don't think one exists. The definition should be downstream of the long process of community development and consensus building, not upstream of it.

What are the likely consequences of a bad Open Source AI Definition?

I expect that the most likely outcome if the Open Source AI Definition is released in its current form will be that it is not respected by the community. This does actually matter! It will damage the reputation of the Open Source Initiative and damage the integrity of Open Source as a coherent and well understood shared vision. It will also split the efforts of the OSI between two definitions which don't mesh together as well as they seem to think that they do.

In the discussion session at OggCamp 2024 it was also pointed out that such a definition would be a gift to companies like Meta, OpenAI, Anthropic, etc and their lobbyists. It would be an independent document from a respected non-profit organisation which they can take with them when they talk to governments about how AI systems should be regulated. Meta especially could argue that their AI system meets this gold standard of being open source without having to change their behaviour around the non-consensual collection of training data. So this definition could weaken the position of individuals and organisations trying to argue against these data collection practices.