They point to copyright/licensing concerns, which I do understand considering the way datasets for most currently popular LLMs have been sourced, i.e. large-sca

They point to copyright/licensing concerns, which I do understand considering the way datasets for most currently popular LLMs have been sourced, i.e. large-scale crawling of any accessible source. In that light, most smart code completion offerings should be fine since they historically did not rely on as broad and widely crawled a data set and were generally offered with clear licenses attached. Also, the number of lines these provided historically was lower, which appears to be part of their concern. In the same light, I feel that LLMs that have been solely and verifiably trained on M̶I̶T̶ MIT-0 or 0BSD should be fine, and this may just be a temporary decision until those become more commonplace. In general, though, this will be hard to enforce, but I do understand why they took this stance considering the current flock of LLMs and their training data. Drawing a line in the sand on licensing, even if they cannot truly ensure this with absolute certainty, is understandable and feels appropriate for them, especially considering LLMs have now been in common use for coding tasks for multiple years, so this does not seem to be a knee-jerk reaction.