4 ms·

To the people arguing it's "fair use" to use this for training an ML network. Where do you draw the line? What if you train an "ML network" with one or two inpu

by 0x0 5y ago

To the people arguing it's "fair use" to use this for training an ML network. Where do you draw the line? What if you train an "ML network" with one or two inputs... so that they almost always "generate" exact copies of the inputs? Five inputs..? Ten? A thousand? A million?

Tenoke 5y ago
I can imagine a requirement of the sort 'generated code needs to match at most X% to snippets of the training data as shown over Y amount of sampling' but I am not sure if you can get a much better requirement than that. Forbidding the training of AI on public code would definitely be a step too far though. Edit: I'd also like if they provide a tool for checking if your code matches copyrighted code too close so you can confirm if you are violating anything or not when you use copilot.
saint_abroad 5y ago
> Where do you draw the line? My simplistic view is that the following is legally equivalent: input -> ai network -> output input -> huffman coding -> output So, whilst: * compressing and decompressing a copyright work is permissible; * output and weights are deterministic transformations of the inputs; thus: * not eligible for copyright (lacking creativity); and * are derivative works of the inputs;
- ghoward 5y ago
  But at the same time, a compiler does a deterministic transformation of its inputs, and we still count its output as under copyright and license. copyrighted input -> compiler -> copyrighted output
  - saint_abroad 5y ago
    Perhaps I wasn't clear enough on this point: copyright of a derivative work is distinct (but not inseparable) to the copyright of the original work. So portions of a derivative work are covered by the original copyright, and other portions may be under a distinct copyright as a derivative work, and several copyrights may apply to a work as a whole. In the case of a Huffman transform, the transformed work does not meet the "creativity" requirements to be eligible for copyright, over that of the original works.
- kevincox 5y ago
  > output and weights are deterministic transformations of the inputs; That may be true but I fail to see how any process that produces the same content that was input into it somehow strips the license. If the generated code is novel, then there is no copyright and it is just the output of the tool. If the code is a copy, but non-creative (example a trivial function) then it isn't covered by copyright in the source anyways, so the output is not protected by copyright either. However if the output is a copy and creative I don't think it matters how complicated your copying process was. What matters is that the code was copied and you need to obey copyright. Again, I don't think that novel code generated from being trained on copyrighted code is the problem. I think it is just the verbatim (or minimally transformed) copying that is the issue.
diffeomorphism 5y ago
There obviously is no sharp line (e.g. it is 37. Immediate question: why not 36?), but that does not matter at all. We already have the same fuzzy line for writing. Am I forbidden from ever reading other author's books because I might accidentally "generate exact copies" of some of the sentences? Clearly not, that is how people learn a language. Does that mean I am allowed to copy the whole book? Also clearly not. Where do you draw the line? Somewhere.
- criddell 5y ago
  And somewhere is determined for your particular case in court. And tomorrow, a similar case may be determined differently.
  - diffeomorphism 5y ago
    Not really, no.
crazygringo 5y ago
The line is exactly the same line that's always been drawn in fair use cases. There's absolutely nothing different whether the creator is ML or a human. Generally, if you train an ML network to generate an almost exact copy of a thousand lines, it's obviously not fair use. If it's five simple lines, it obviously is fair use. If it's somewhere in between, there are a lot of different factors that need to be weighed in a fair use decision, which you can easily look up.

by 0x0 5y ago

Tenoke 5y ago

I can imagine a requirement of the sort 'generated code needs to match at most X% to snippets of the training data as shown over Y amount of sampling' but I am not sure if you can get a much better requirement than that. Forbidding the training of AI on public code would definitely be a step too far though. Edit: I'd also like if they provide a tool for checking if your code matches copyrighted code too close so you can confirm if you are violating anything or not when you use copilot.

saint_abroad 5y ago

> Where do you draw the line? My simplistic view is that the following is legally equivalent: input -> ai network -> output input -> huffman coding -> output So, whilst: * compressing and decompressing a copyright work is permissible; * output and weights are deterministic transformations of the inputs; thus: * not eligible for copyright (lacking creativity); and * are derivative works of the inputs;

ghoward 5y ago
But at the same time, a compiler does a deterministic transformation of its inputs, and we still count its output as under copyright and license. copyrighted input -> compiler -> copyrighted output
- saint_abroad 5y ago
  Perhaps I wasn't clear enough on this point: copyright of a derivative work is distinct (but not inseparable) to the copyright of the original work. So portions of a derivative work are covered by the original copyright, and other portions may be under a distinct copyright as a derivative work, and several copyrights may apply to a work as a whole. In the case of a Huffman transform, the transformed work does not meet the "creativity" requirements to be eligible for copyright, over that of the original works.
kevincox 5y ago
> output and weights are deterministic transformations of the inputs; That may be true but I fail to see how any process that produces the same content that was input into it somehow strips the license. If the generated code is novel, then there is no copyright and it is just the output of the tool. If the code is a copy, but non-creative (example a trivial function) then it isn't covered by copyright in the source anyways, so the output is not protected by copyright either. However if the output is a copy and creative I don't think it matters how complicated your copying process was. What matters is that the code was copied and you need to obey copyright. Again, I don't think that novel code generated from being trained on copyrighted code is the problem. I think it is just the verbatim (or minimally transformed) copying that is the issue.

diffeomorphism 5y ago

There obviously is no sharp line (e.g. it is 37. Immediate question: why not 36?), but that does not matter at all. We already have the same fuzzy line for writing. Am I forbidden from ever reading other author's books because I might accidentally "generate exact copies" of some of the sentences? Clearly not, that is how people learn a language. Does that mean I am allowed to copy the whole book? Also clearly not. Where do you draw the line? Somewhere.

criddell 5y ago
And somewhere is determined for your particular case in court. And tomorrow, a similar case may be determined differently.
- diffeomorphism 5y ago
  Not really, no.

crazygringo 5y ago

The line is exactly the same line that's always been drawn in fair use cases. There's absolutely nothing different whether the creator is ML or a human. Generally, if you train an ML network to generate an almost exact copy of a thousand lines, it's obviously not fair use. If it's five simple lines, it obviously is fair use. If it's somewhere in between, there are a lot of different factors that need to be weighed in a fair use decision, which you can easily look up.