So on the surface, it seems like that would be fine, right? There's a piece of software called 'Capo', for example, that assists with the process of transcript

So on the surface, it seems like that would be fine, right? There's a piece of software called 'Capo', for example, that assists with the process of transcription, albeit not generating tab directly, but which analyzes a piece of music and helps you pick out and transcribe notes as tab - and for personal use, of course there's no problem doing so. It should be okay if it went further and attempted to autogenerate tabs. It does chords already, for example. And writing the tab out, for your own purposes, from a recording you have bought, is not infringing copyright on the original any more than listening to a copyrighted recording you have bought through speakers in your own home is infringing - I mean, it's possible that it could be considered 'format shifting' a work, like copying a vinyl onto a tape, but really it's more akin to just taking really detailed notes as you listen... And having a software tool that helps you do it seems like a reasonable thing to use, so how could it be infringing? In a way, if some ML software existed that could generate accurate tab from an audio file, then that software, plus a guitar, is just a really low-fidelity digital-to-analog converter pipeline for turning the original recording into audio. Just the same as an AAC codec + pair of headphones. The codec turns the original compressed data into a series of numbers representing audio samples, which a speaker turns into sounds; the Tab-transcribing ML turns the original data into a series of numbers representing which strings and frets to play, which a guitar player turns into sounds. So on that basis, it seems hard to argue that the existence of such an ML model is a copyright circumvention device, any more than you could argue that a speaker is a copyright infringement device. But... What if the way you trained the ML model in the first place, for example, was by feeding it copyrighted guitar tabs of copyrighted recordings? In particular, what if it was only for a very limited repertoire - say you just trained it on Hendrix recordings, and official licensed Hendrix tabs? Then, if you hand it something other than a Jimi Hendrix guitar performance, it might not be very good (although I bet the results would be interesting!) - but if you hand it any original Hendrix recording, the same recording of Little Wing that was used in training it, for example - then there's a really good chance it would just reproduce the official tab of Little Wing. The legal question might be, does the ML model embody the knowledge of how to produce a good tab from a recording, or just contain the memory of, and ability to reproduce, the copyrighted work? And when it comes to something like music, that gets very close to asking the question of, when it comes to a musical performer producing a song that sounds similar to a copyrighted work they were influenced by: Do they embody the knowledge of how to write a song that sounds like that or are they just reproducing their memory of the previous copyrighted work? Which gets to the 'originality' test in copyright law. And applying that in the case of an AI piece of software sounds... challenging. But on the other hand, it's pretty clear, for example, that when you hand an audio file to the AAC codec and it produces audio from it, that you can't argue that the codec is just reproducing the memory of the right sounds to make, prompted by the audio file - but on the other hand, it's not like the codec is bringing nothing to the table in terms of figuring out how to turn the 1s and 0s on disk into sounds.