Timeline for Why are humans and AI often treated differently in cases where they perform nearly identical processes?
Current License: CC BY-SA 4.0
18 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Sep 27, 2023 at 0:07 | comment | added | Scott Rowe | @user4574 a lot of research has gone in to evaluating the difference in children's school performance relative to the amount and kinds of language use they are exposed to from birth (or before) to school age. More input of complex language makes a big difference. Same as we are pondering here. | |
| Sep 26, 2023 at 1:24 | comment | added | user4574 | "The training sets for generative AI systems are orders of magnitude larger than the number of images or words that a human being sees in a lifetime" Thinking about this, I am not sure it's true. If the human eye can see like 20 equivalent frames per second, and we are awake for 16 hours a day we see like 420 million images per year. Many of those images a human sees are similar to each other, but that is probably extremely useful for building models of how objects move/behave (and people training AI should probably take that as a hint to train AI on Cat videos rather than cat pictures). | |
| Sep 24, 2023 at 21:10 | comment | added | Scott Rowe | Maybe AI could provide for people's needs, then they wouldn't have to sell their time making wants? | |
| Sep 22, 2023 at 13:34 | comment | added | NotThatGuy | @leftaroundabout A lot of research has gone into what the structure of neural networks should look like, and how to train it, which could similarly be considered "other things" that isn't copyrighted. Much of the material in the training set also isn't copyrighted. I have no idea what fraction it is, but I'd certainly take issue with an un-backed-up "pretty much exclusively copyrighted". Although it shouldn't really matter what fraction is copyrighted. To me, it seems that either it's okay to use a single piece of copyrighted material in this way, or it isn't. | |
| Sep 22, 2023 at 13:09 | comment | added | David Gudeman | @Brondahl, and besides the point leftaroundabout made, the human is able to do thousands of things with that input: discuss religion, drive cars, play volleyball, wash dishes, play with dogs, trouble-shoot technical problems, talk someone into doing something, mow grass, recognize when the party is getting out of control, grocery shop, and much more. An AI, with more data more carefully curated can only do one of those things and some it may not be able to do at all. | |
| Sep 22, 2023 at 12:59 | comment | added | leftaroundabout | @NotThatGuy "Does using more data imply plagiarism? Shouldn't it be the other way around?" I'd say neither more or less data should make it more or less plagiarism, but it seems reasonable that a higher fraction of peoples' copyrighted work in the data plays a role. The AIs discussed here are trained pretty much exclusively on copyrighted material. A human artist is trained on a mixture of copyrighted material and personal sensory input. Hard to attribute how much of each turns up in the their artworks. | |
| Sep 22, 2023 at 12:44 | comment | added | leftaroundabout | @Brondahl "45fps. That's 1.4 billion frames a year" - quite exaggerated; most of the time you'll have thousands of consecutive frames that are just slight variations of the same image, and the scenes also tend to repeat daily. Whereas the billions of images an art AI is trained on are mostly really different images with only some repetitions. A human does certainly encounter millions of image during their lifetime, but at least most of these aren't copyrighted. | |
| Sep 22, 2023 at 9:49 | comment | added | NotThatGuy | I'm struggling to follow the train of thought and implication in this answer. Does using more data imply plagiarism? Shouldn't it be the other way around? Is the brain not plagiarism because of "other things going on in the brain"? What other things and why do those not make it plagiarism? There are also other things going on in AI, that isn't in the brain, so there simply being other things doesn't say much, one way or the other. This answer doesn't really seem to make the case that they should be treated differently (i.e. what was asked), but merely that question "assumes" they shouldn't be. | |
| Sep 22, 2023 at 9:13 | comment | added | NotThatGuy | The profits of artists may be a valid concern, but that's distinct from whether AI produces something "new" in a philosophical or functional sense. To rely on that argument would be to say that, yes, AI might be transformative and doesn't functionally violate copyright, but we'll treat it as if it does in any case, because we think that leads to a better outcome (which is in itself questionable). We could do that, but it's important to separate concerns. | |
| Sep 22, 2023 at 9:13 | comment | added | NotThatGuy | "If the art student and the generative AI, trained on the same images, produced similar output..." - that seems like an odd argument. That'd be like saying "if Alice and Bob, looking at similar images, produced similar output, it would be unreasonable to treat them differently". Obviously the output is likely to be different on account of different processing steps (even between humans). It's unreasonable to treat them differently (with respect to copyright) because for both, any given the input image is just one tiny part of what ultimately produces some output. | |
| Sep 22, 2023 at 8:57 | comment | added | Brondahl | Human vision is 30-60 fps; so call it 45fps. That's 1.4 billion frames a year; Top Google search says Art AI are trained on 1-10 billion images. So no, their training set isn't an order or magnitude larger. | |
| Sep 22, 2023 at 8:15 | comment | added | Brondahl | @DavidGudeman you are right, and that's part of the point. AIs have developed in their neural net a great number of very sophisticated digital model-components of their inputs, and they interpret whatever they are given next according to those models. | |
| Sep 22, 2023 at 8:06 | comment | added | David Gudeman | @Echox, you are right, and that's part of the point. Humans have developed in their mind a great number of very sophisticated analog models of the world, and they interpret whatever they see according to those models. | |
| Sep 22, 2023 at 8:01 | comment | added | Jemox | @DavidGudeman Can a human ever only have "a few examples" ? The student asked to get inspiration from 5 artists would already have a database of twenty years of seeing the world through his eyes, and thousands of different drawings and artpiece seen randomly throughout his years. If you put someone in the dark for twenty years, then show him five pictures and ask him to draw a sixth, I'm not sure you can guarantee the work won't be extremely derivative. | |
| Sep 22, 2023 at 1:24 | comment | added | David Gudeman | @user4574, there is no algorithm that can make an ML system work with just a few examples like a human can, unless that algorithm already incorporates human-like intelligence. How is the algorithm going to recognize what is relevant in the training set without a huge number of examples to show it? | |
| Sep 22, 2023 at 0:23 | comment | added | benrg | @user4574 Well, if you're proven right, I'll concede your point. Moore's law has lasted longer than I thought it would, but I don't think it can last another 15 years. | |
| Sep 22, 2023 at 0:13 | comment | added | user4574 | "The training sets for generative AI systems are orders of magnitude larger than the number of images or words that a human being sees in a lifetime." While true today, it's not a fundamental aspect of the technology. Due to lack of processing power, most neural network models don't necessarily have the right structure or scale. Chat GPT or Google Bard might have like 1% as many neurons as a human, and like 100X less connections per neuron. Plus, training algorithms are different (back propagation vs Hebbian learning). If Moor's law holds, that will change in about 15 years or so. | |
| Sep 22, 2023 at 0:00 | history | answered | benrg | CC BY-SA 4.0 |