Why Generative AI Violates Copyright – For Skeptics

Another retread of the misconceived "AI as collage machine" idea behind some the early lawsuits, i.e. training data goes in and something intelligible comes out, therefore the output is "made of" the input. Non sequitur.

The author glosses over the actual point of contention, saying "through some mechanical trickery it will blend then together with some randomness to create something that is a new song".

'Blending together' is not what's happening. The "mechanical trickery" is the transformative part.

I'm really not sure this sells me on it. Sam Altman is a clown of a character to be sure; his success will be the embarrassment of venture capitalists for decades to come. Part of me suspects the startup industry may never recover from it, but there's always more fools with money.

The biggest thing I disagree with is the representation of AI and the definition of 'transformative' work. I think it's best argued in reverse-order.

Imagine for a moment, that we take AI out of the situation. No training data, no neural network, just a computer generating images or text along some PRNG function like a fractal or a Markov chain. These are, in essence, original works. They might be nonsense, offensive, or just plain bad, but they are quite likely unique and could feasibly be copywritten by a human that purports to be it's discoverer. So in that sense, I think there is both a world in which computers not only create original art, but also do meaningfully transform inputs. The lack of conscious artistry doesn't stop 3D animated movies from being copywritten for using randomized particle effects, nor should it stop computer-generated images and text.

My second issue is really how you define AI in your player-piano analogy. Analogous to AI, I don't know if we could definitively say the piano isn't playing unique or original music. If it produces an tear-jerking rendition of Farmer in the Dell, I don't think we should crucify it for stealing or abusing the original work. You could argue that it's not derivative enough to warrant it's own copyright protections (and may be right), but I don't think it inherently violates copyright any more than human derivation does. The means are separate, but the ends are identical.