I think this is exposing a fundamental conceptual flaw in LLMs as they’re designed today. They can’t seem to simultaneously respect intellectual property / licensing and be useful.

Their current best use case - that is to say, a use case where copyright isn’t an issue - is dedicated instances trained on internal organization data. For example, Copilot Enterprise, which can be configured to use only the enterprise’s data, without any public inputs. If you’re only using your own data to train it, then copyright doesn’t come into play.

That’s been implemented where I work, and the best thing about it is that you get suggestions already tailored to your company’s coding style. And its suggestions improve the more you use it.

But AI for public consumption? Nope. Too problematic. In fact, public AI has been explicitly banned in our environment.