A better solution would be to just not have AI at all, outside of the few research roles where LLMs actually make sense.
More to the point it's trained on copyrighted material, so why entertain any use at all on that front if anything.
If it's trained on the world's information, give the world the model.
It doesn't need a tech company to pilfer everything and charge X if we're going to ignore the IP.
The models free to download aren't the models used by OpenAI, Anthropic and Google. You aren't going to get all of OpenAI downloaded to your desktop and running fully on just your hardware.
And in each case (search and AI) the potential to decentralize and maintain "usable quality" is limited by these technologies requiring physical infrastructure at a scale that isn't available to the home consumer.