My stance on AI produced content for wellbeing of humanity: 2023 (1/3)
I am always open to discussion and correction.
If a model is trained on content of given license, say GPL or Creative Commons, I believe it's fair that content produced from this model must also comply with the original license.
(2/4) I think, at least for now, while there is not a social consensus of a good outcome, models should only be allowed to be trained on content with compatible licensing, to follow the previous rule. So for code, it should be either all-GPL-compatible, all-proprietary (that the model producers have a license to), and so on. For images, all-public domain, all-creative commons, and so on. Mixing incompatible license should not be allowed.
(3/4) in case a model uses public resources (such as GPL'd or CC'd content), it should disclose the content (such as images) used, crediting each author. An author may choose to make his work not available for training of (non-conscious) machines in his license.
(4/4) Long term solutions that I see include have "inspiration tracing" technologies that find the images a certain content are most inspired by, and rewarding them according to their impact. Also, we should start thinking about the consequences of widespread automation and rewarding humanity more generally with solutions like Universal Basic Income, in addition to more direct rewards that seem both fair and useful to provide feedback to real human artists.
Request for comments @smallcircles @notes @tripu @pluralistic @cstross @codinghorror @torvalds
I'm no expert on this, but I like the idea of not mixing licences when training models. It's only fair and logical. We pick licences for a reason.
But I fear it's not feasible.
@tripu @gnramires I agree. It should be enforced by regulation then, which is quite hard esp. with a global AI "arms race" to conquer the market & monetize.
LLM's are getting more opaque, and depending on how they are integrated it may be impossible for 3rd-parties to know which data they were trained on. Requirement to disclose the entire training set may be another regulation.
Asking consent and even "inspiration tracing" will be fiercely battled against by the big tech lobbyists.