The Open Source Initiative has released the official definition of open source AI. A lot of companies have been promoting their models as open source, but even before this official definition it was clear they were not in fact open.
Most of the released "open" model have been just the model. It's a good start and the most useful thing for most users. They don't really care about how to train their own model as long as they can use these "open" models freely for their needs.
But there are many benefits of having a truly open source AI. Even if you are not going to train one by yourself. The usage terms are just one thing. The most important thing, and what I'm glad OSI included in their definition is the requirement of the source data used to train the model must also be openly available.
There has been a lot of debate about the usage of source data that the big companies have been using to train their models. Including that to the official definition could be a big step forward considering that there are already many legislations that treat open source models differently and which the big corporations have been able to exploit due to the vague definition of what open means in the context of AI.