Trying to set up a prompt injection attack for someone accessing a repo with a coding agent is juvenile and pointless. And it doesn't deal with the training part.
I look at this pretty similar to licenses that attempt to say something can't be used in some specific way, such as a "no evil" license. Famously, the JSON license:
Permission is hereby granted, free of charge, to any person obtaining a copy of
this software and associated documentation files (the "Software"), to deal in
the Software without restriction, including without limitation the rights to
use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
the Software, and to permit persons to whom the Software is furnished to do so,
subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
The Software shall be used for Good, not Evil.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
This seemingly pointless yet straightforward addition of the license has caused problems, because it's highly subjective and therefore makes compliance with the license impossible to objectively measure - which is really important for a license!
I think LLM training seems right now a simple objectively measurable thing - and maybe it will always be that simple - but I could also see it becoming a subjective thing. At the very least the interpretation of copyright law has yet to be upheld, which historically was one of the most "powerful" license-like things in existence.
In theory the same way people are making those claims about "stolen" art, such as models that produced watermarks from Getty images or Shutterstock. Similar "watermarks" have existed in some LLM output.
I remember the case of books used for training, where the court found training to be fair use, but the material has to be legally obtained (=Bought instead of pirated the books).
Not an objection, you're fully within your rights to invent a license, just remember that a license with such clause would not be an open source license as per the Open Source Definition [1].
6. No Discrimination Against Fields of Endeavor
The license must not restrict anyone from making use of the program in a specific field of endeavor. For example, it may not restrict the program from being used in a business, or from being used for genetic research.
There's also 5. which probably isn't relevant here but possibly might depending on the definition of person (e.g. a legal person).
5. No Discrimination Against Persons or Groups
The license must not discriminate against any person or group of persons.
And then there's the obvious aspect of AI foundries blatantly ignoring copyright anyway (copyright law grants the the author the rights which he then gives away by way of a license).
If you use a non-standard license, I will pass on your project, legal is not going to take the time to review and will just reject. It can be counterproductive to your project depending on your goals
You are looking for "ethical" / "permissioned" licenses (as many as there are people's gripes or causes), versus permissionless like MIT / Apache-2 / BSD-3
There is no legal way to prevent things. Bad actors will not care about your license. If you catch them, you can pay lawyers to try and have it enforced, but these licenses have not been tested in court yet, so no one knows who will win.
Not just companies, people too. It will have trouble getting into Linux distro repos. And a lot of devs/users avoid non-open-source projects especially if they went looking for solutions on Github.
> If you use a non-standard license, I will pass on your project, legal is not going to take the time to review and will just reject. It can be counterproductive to your project depending on your goals
For personal projects, that actually sounds like a good thing, didn't think of it that way. Probably I'll start making my projects MIT-but-modified-enough-to-scare-lawyers, and also provide a "clean" MIT license to companies if they agree to pay per month with either money or engineering hours for maintenance.
Best of two worlds, would get rid of the worst vampires at least.
I think the problem is that even if you have a "no training" license AI companies will just ignore it and you'll be left having to take legal action.
If the big publishing firms couldn't win an action against Meta for using pirated copies of their books you probably have no chance either.
It may seem impossible on an individual level, but collectively it may make sense to take them on, ala death by a thousand paper cuts style.
If content with that license reaches critical mass, a law firm could sue for many at once.
Do you believe this critical mass would likely exceed all books ever published within the current copyright window?
Do you think book publishers are somehow less financially able to muster a legal response than open source coders in their spare time?
Books don't have an explicit no-LLM license.
You may find https://github.com/manyfold3d/manyfold/blob/main/AGENTS.md useful ;-)
Trying to set up a prompt injection attack for someone accessing a repo with a coding agent is juvenile and pointless. And it doesn't deal with the training part.
I look at this pretty similar to licenses that attempt to say something can't be used in some specific way, such as a "no evil" license. Famously, the JSON license:
(Source: https://www.json.org/license.html)This seemingly pointless yet straightforward addition of the license has caused problems, because it's highly subjective and therefore makes compliance with the license impossible to objectively measure - which is really important for a license!
I think LLM training seems right now a simple objectively measurable thing - and maybe it will always be that simple - but I could also see it becoming a subjective thing. At the very least the interpretation of copyright law has yet to be upheld, which historically was one of the most "powerful" license-like things in existence.
How do you objectively measure if an LLM is training on something if you don't have access to its training data?
In theory the same way people are making those claims about "stolen" art, such as models that produced watermarks from Getty images or Shutterstock. Similar "watermarks" have existed in some LLM output.
Would this even work legally?
I remember the case of books used for training, where the court found training to be fair use, but the material has to be legally obtained (=Bought instead of pirated the books).
> and usage on it
What do you mean by "usage on it"?
I mean a clause along "don't ever have any LLM read my code" either for training, or for making a vibecoded output.
How does it work legally? IANAL so I have no idea and that's why I'm asking.
"Bots strictly prohibited."
Some lawyers may try to weasel their way around but if it comes to a court case you should have a pretty straightforward interpretation of the phrase:
Not an objection, you're fully within your rights to invent a license, just remember that a license with such clause would not be an open source license as per the Open Source Definition [1].
There's also 5. which probably isn't relevant here but possibly might depending on the definition of person (e.g. a legal person). And then there's the obvious aspect of AI foundries blatantly ignoring copyright anyway (copyright law grants the the author the rights which he then gives away by way of a license).[1] https://opensource.org/osd
If you use a non-standard license, I will pass on your project, legal is not going to take the time to review and will just reject. It can be counterproductive to your project depending on your goals
You are looking for "ethical" / "permissioned" licenses (as many as there are people's gripes or causes), versus permissionless like MIT / Apache-2 / BSD-3
There is no legal way to prevent things. Bad actors will not care about your license. If you catch them, you can pay lawyers to try and have it enforced, but these licenses have not been tested in court yet, so no one knows who will win.
> depending on your goals
is the important bit. Is your goal to have your project used by companies which have to "run things through legal", or do you have better goals?
Not just companies, people too. It will have trouble getting into Linux distro repos. And a lot of devs/users avoid non-open-source projects especially if they went looking for solutions on Github.
> If you use a non-standard license, I will pass on your project, legal is not going to take the time to review and will just reject. It can be counterproductive to your project depending on your goals
For personal projects, that actually sounds like a good thing, didn't think of it that way. Probably I'll start making my projects MIT-but-modified-enough-to-scare-lawyers, and also provide a "clean" MIT license to companies if they agree to pay per month with either money or engineering hours for maintenance.
Best of two worlds, would get rid of the worst vampires at least.
Are you financially able to take on open ai, anthropic, Facebook, etc? If not, then it doesn't matter what licence you use.