A human artist learns to draw clouds by, for example, watching a Miyazaki movie, deconstructing how they draw clouds, and then adapting those skills to their own work. Is that really any different from what the AI is doing during training?
The machine doesn’t keep a perfect copy of the art either. A neural network is inherently lossy compression, so training leaves the model with an impression of the work, not the work itself, much like the artist. Models do sometimes regurgitate near-verbatim snippets of work they saw over and over in training. But so does a human who has read a poem one too many times. When the human publishes it we call it plagiarism, and we blame the human, rightly so.
This brings me to the point where I feel like we need a new classification of intelligence to even begin to wrap our heads around how to think about AI plagiarism.
I’m thinking two classifications, human intelligence and proto intelligence.
Both should be allowed to use existing art for training data, and I’d argue current copyright law and fair use rules support this. I don’t think you could restrict access to training data even if you wanted to.
The core difference between these two intelligences is:
- Human intelligence has agency and can be employed.
- Proto intelligence does not have agency and is currently owned by its operator.
(That’s how we treat proto intelligence today, not necessarily how we should)
Which means:
- Human intelligence has the agency to refuse to create derived works or engage in plagiarism.
- Proto intelligence can only refuse what its owner allows it to refuse.
Since a proto intelligence’s ability to refuse is decided by its owner, not by the intelligence itself, I believe the AI must be classified as a tool, operated by its owner. At least within the current framework, where an AI has an owner who owns or rents the hardware that allows inference. (The inference hardware is the physical brain, the neural network the model running on the brain)
If AI is a tool, do we charge the person who provided the tool, or the person who used the tool, for the crime of plagiarism?
In other words, do gun manufacturers kill people, or do the people who use the guns kill people?
I don’t expect this to settle anything, but I hope it at least gets us arguing the right question… again.