We urgently need the watermark equivalent of a “robots.txt” for Machine Learning(twitter.com)
twitter.com
We urgently need the watermark equivalent of a “robots.txt” for Machine Learning
https://twitter.com/joshenders/status/1565321360884236289
https://twitter.com/joshenders/status/1565321360884236289
I've been trying to determine if github, which supplies ready-made license files for every repository, enforces these licenses as implemented when feeding co-pilot. It appears that it does not, which is really baffling since there is no excuse for ignoring licenses that they supply as part of the process.
One way I tested was to write a Hello World in Common Lisp and other languages. The code co-pilot generated very often comes directly from a course, and "World" was replaced by the name of the site or book the code was listed in. Did the license allow for this use?
You can opt-out of your code being used by co-pilot, but once it is already in the model, then what? And why require user input when the license file already states whether it is permitted or not?
The other thing is just how much crap code there is on github. What kind of quality should one expect when poor code is copied so much, increasing the likelihood co-pilot will treat it as gospel.
AI has a long ways to go yet.