For anyone who actually used those models for more than few days and learns their strengths and weaknesses it is completely obvious all this talk of "societal impact", or as you called it self imposed limits are 100% bulls**. Everyone in the field knows it.

99% of those using this tactics use it to justify not releasing their models to avoid giving competition a leg up(Google, openAI) and to pretend they are for "open research". As I said this is 100% bull.

The remaining 1% are either doing this to inflate their egos ("hey look how considerate and enlightened we are in everything we do!"), or they pander to media/silly politicians/various clueless commentators whose level of knowledge about this technology is null. They regurgitate the same set of "what ifs and horror stories" to scare the public into standing by when they attempt to over regulate another field so they can be kingmakers within it(if you want an example how it works look at the energy sector).

All this sillyness accomplishes is to raise a barrier to entry for potential commercial competition. Bad actors will have enough money/lack scruples to train their own models or to steal your best ones regardless how "impact conscious" your company it.

Now, I don't claim everyone should be forced to publish their AI models. No, if you spent lots of money on training your model it is yours. But you can't lock all your work behind closed doors and call yourself open. It doesn't work like this. One important point is that there is value even in just publishing a paper demonstrating some achievements of a proprietary model, but if the experiment can't be reproduces based on description given that is not science and for sure it is not open.