Microsoft is saying “if our tools produce copyrighted violations we will assume responsibility”, which looks great on the surface, because almost nobody would be able to find the infringements, so it is a bet worth taking.
Exceptions include places where the code is throughly audited, like acquisitions: so this might just bust acquisitions, and I don’t think this would be covered:
https://blogs.microsoft.com/on-the-issues/2023/09/07/copilot-copyright-commitment-ai-legal-concerns/
@80f08b97 the part that gets me scratching my head is that it's a low-key admission that copyrighted material was used to train these models without consent. They'll just filter it on the way out.
@80f08b97 "We are sensitive to the concerns of authors, and we believe that Microsoft rather than our customers should assume the responsibility to address them." We are so sensitive to author's rights that we are going to use our legal muscle to crush any authors who may find that their works have been plagiarized via Copilot laundering.
@80f08b97 where are the exceptions documented, I couldn't find it on that page
@80f08b97 It would have been a shorter blog post if Microsoft had been able to say Copilot is trained only on content that they have the rights to and said that they have the chain of evidence to prove it.
How much effort & expense are customers going to put into policing this clause:
“Our new Copilot Copyright Commitment requires that customers use [guardrails built into Copilot]”? Worthless guarantee unless customers are confident in getting that right.
@80f08b97 But they're still not confident enough to feed their own private source code into the public model, right?
Oddbean new post about login | logout