> Are deepfakes so realistic that it messes with our fabric of reality?
I disagree with you here because images are very different than text. When you learn some text or even verbal statement is false, your brain is somehow able to label it as false, and when you remember it, you remember that it's false. In contrast, with images, when you "see it with your own eyes", it's very hard to keep in mind that it's false. The image sticks with you and continues to influence your thinking. I realize this doesn't account for a whole category of images that come from a fictional world, like most things we see on TV. I think we see those as real things, but things that happened in a different world. Compare that for example to seeing a picture of a real person (such as a presidential candidate) sexually abusing a child. That image of the real person is going to stay with you and influence how you think about the person even if the image is fake. So I think the ability to generate fake images does have the capacity to distort our reality in a way that text cannot.
Maybe our brains can eventually adapt to this, but I think it would take at least a few generations. But it would be cool to have the opinion of an evolutionary biologist on this.
In fact, it's that that led me to the TikTok that yesterday's AGI Friday is arguing with.
(Also I was confused in attributing it. I thought I remembered it being a TikTok and then thought I was wrong and that it was actually Bluesky. Turns out it's a TikTok that got reposted to Bluesky. I'm just gonna stick with "TwitterTok or whatever".)
PS: For the record, GPT-5 can answer the Titanic swimming pool question fine.
- The paragraph completely disregards not-commercially-published texts. What about blog posts? What about YouTube videos? You cannot usually buy access to them like you can to a book, but that doesn't mean that their creators automatically consented to AI training from their work.
- And, perhaps most importantly, the paragraph completely disregards the actual *creation* of AI art. (Yes, that wasn't the point of the question, but it's the much bigger issue when it comes to AI and copyright, imo!) I could ask AI today to create hundreds of images in the style of Ghibli films, or Banksy murals, or to write me a couple dozen AC/DC songs, or three new Harry Potter books. This is especially a bitter point for me, as a hobby artist: I hate the fact that people can capitalise on something which I do for fun and which takes me *a lot* of time in a matter of a few minutes.
Fair perspective. It's hard for me to hate it when I'm still completely flabbergasted that it's possible. I think if I were an artist (professional or otherwise) I'd feel differently.
Serious question: how much do you hate the AI-generated images that I use for AGI Friday posts?
I'm not a visual artist, I'm a writer, so I don't hate the AI-generated images as much as I probably would otherwise? I am also fascinated by what's possible; and those test images you shared once (astronaut fox etc.) were pretty interesting.
I guess I'm also a bit torn on AI art. On the one hand I regonise the possibilities. Like, maybe in ten years I can upload everything I've written and tell an AI "Write an 8k short story in my style about [whatever]." And then I'd only have to do one edit pass, and I'd be perfectly happy with the result, and I'd be done in two hours instead of eight or more.
But convenience can also be a fallacy: Sometimes, the half-hour agonising over a few sentences is exactly what you need to push the story into an even better direction. Instead, with AI-generated art, the way of less resistance is right in front of you: You just accept what it creates for you, because it's already finished, and it only took a few seconds.
But I do think that AI is going to have a hard time to perfectly emulate all forms of art. Prose is easy enough, poetry following a certain format (sonnets, haiku) is definitely doable, but what about free-form poetry? What about hybrid forms?
I think my fear is that either it will be able to emulate *everything* convincingly, and then people won't bother creating anything themselves anymore, OR it will significantly fail at some things (like very personal/emotional/associative free-form poetry), and then less of that will exist, because people won't be able to create it with AI, and those who could write it themselves would do so less because they could produce more things, and quicker, with the help of AI. And both sound like horrible options to me.
But generally, I'm a fan of lowering entry barriers, especially with art. There's so much art that doesn't get created just because people think "I'm not good enough" / "I wouldn't even know where to begin" / "I don't have the time/money/resources". Inputting prompts into a computer does at least mitigate some of these points, if not all. (You'd still need internet access, and you'd still have to pay something, at the very least.)
I just don't want the norm to be uninspired AI generated art en masse (to not say "slop"), because I think that'll only stifle human creativity instead of inspire it!
I keep thinking of two analogies: books vs. ebooks and traditional craft vs. commodities.
I work in the literary industry, and when ebooks were on the rise, everyone was afraid that the demand of physical books would rapidly decline and basically crush most of the industry, especially bookstores. But that didn't happen, at least not here in central Europe.
On the other hand, most people buy their furniture from Ikea and similar stores. If you own handmade furniture (or handmade ceramics, glasses, whatever), you likely either A) inherited it, or B) are wealthy enough to afford it.
I just don't want human art to end up like traditional craft items—only for a wealthy elite. I'd much rather it turns out like paper books—at most slightly more expensive than ebooks/AI art, but still widely spread and cheap enough that most people can afford it.
Great points. I do think AI slop is a big problem right now. I genuinely don't know if it will get worse as slop becomes harder to identify or if it will just get good enough that the term "AI slop" won't be apt. I admit to being a little horrified by your hypothetical of an author of fiction letting an AI write a new story emulating the author's style, editing it a bit, and publishing it. I might be being naive but it feels like that couldn't possibly be a genuinely good story, pre-AGI.
(And yet, it also kind of feels like we're on track to that being possible in a few more years. Which, yes, might be because we're on track to AGI in a few years.)
As for visual art, my very non-artist opinion is that the people who think AI art is ugly are refusing to admit that the emperor has no clothes. I mean, AI may be commonly used to create ugly art but the critics of AI art generally mean more than that. That they can spot AI art a mile away and they despise it. I think anyone in that category should try the following AI art Turing test: https://ai-art-turing-test.com/ (note that it's already a year old so presumably AI art is that much better today).
"I genuinely don't know if it will get worse as slop becomes harder to identify or if it will just get good enough that the term 'AI slop' won't be apt.« — Hypothetical worst of both worlds: good enough (in technique) that 'slop' won't technically be apt, but also abundant enough *everywhere* that it will still drown out most of the unique, creative, inventive human art out there, simply because you're gonna be flooded by a million versions of the same idea.
"I admit to being a little horrified by your hypothetical of an author of fiction letting an AI write a new story emulating the author's style, editing it a bit, and publishing it. I might be being naive but it feels like that couldn't possibly be a genuinely good story, pre-AGI." — Well, depends on your definition of "genuinely good story," maybe? (Also depends on the editing skills and how much editing is needed ;-) There's tons of middling-to-bad stories/books out there, and that was the case even before the time of the internet or the possibility to self-publish books. Maybe AI pre-AGI couldn't really, convincingly emulate Poe, but there must be hundreds (thousands?) of Poe-inspired short stories or re-writings online, probably ranging from "horrible" and "bad" to "okay I guess", with only a few exceptions that are "good" or "great".
I never tried a writing-AI, but I don't see it having tons of problems with replicating already-not-all-that-great content. I guess for me the scary aspect isn't that AI art might suddenly create the new Picasso (or Proust) and put all human art to shame, but rather that it will create *so much* middling/kinda okay/not really good art, that it will become more difficult for the new human Picasso (or Proust) to gain recognition / they will become more easily discouraged, etc. [all very broad generalisations, of course, and it's not at all unproblematic to rank art by "objective quality," to praise "canonically great" art, etc. etc.]
"AI may be commonly used to create ugly art" — Yeah, I think that's the key problem, that most AI art that we see nowadays (I was SO baffled when I found out there's uncanny AI videos *everywhere* on social media! I guess I prefer to live under rocks sometimes!!) is ucanny and/or bad and/or uninspired, and so most people think that's all you can do with AI art! I do think that some (most?) of the critics of AI art pretend they can spot it a mile away, just because they think AI art is only 6-fingered hands and weird perspectives. But maybe I'm wrong.
I did take the AI turing test a while back and was absolutely baffled by it, I'll take it again now, because I'm curious if the results will be different. (My memory is very bad, so I won't consciously remember the solution of specific pictures, so I can't "cheat").
> Are deepfakes so realistic that it messes with our fabric of reality?
I disagree with you here because images are very different than text. When you learn some text or even verbal statement is false, your brain is somehow able to label it as false, and when you remember it, you remember that it's false. In contrast, with images, when you "see it with your own eyes", it's very hard to keep in mind that it's false. The image sticks with you and continues to influence your thinking. I realize this doesn't account for a whole category of images that come from a fictional world, like most things we see on TV. I think we see those as real things, but things that happened in a different world. Compare that for example to seeing a picture of a real person (such as a presidential candidate) sexually abusing a child. That image of the real person is going to stay with you and influence how you think about the person even if the image is fake. So I think the ability to generate fake images does have the capacity to distort our reality in a way that text cannot.
Excellent point. I'm not necessarily convinced -- maybe our brains can just adapt to this? -- but it's a key difference I neglected.
Maybe our brains can eventually adapt to this, but I think it would take at least a few generations. But it would be cool to have the opinion of an evolutionary biologist on this.
Ridiculing AI has not exactly ended. Take this hilarity:
https://bsky.app/profile/timkrief.com/post/3lyxkhf3zns2a
In fact, it's that that led me to the TikTok that yesterday's AGI Friday is arguing with.
(Also I was confused in attributing it. I thought I remembered it being a TikTok and then thought I was wrong and that it was actually Bluesky. Turns out it's a TikTok that got reposted to Bluesky. I'm just gonna stick with "TwitterTok or whatever".)
PS: For the record, GPT-5 can answer the Titanic swimming pool question fine.
I have a lot of issues with "3.) Does AI training violate copyright?"
- "Fair Use" is only one fraction of copyright law. Most of the world doesn't use Common Law. (See https://upload.wikimedia.org/wikipedia/commons/9/92/Map_of_the_Legal_systems_of_the_world_%28en%29.png)
- The paragraph completely disregards not-commercially-published texts. What about blog posts? What about YouTube videos? You cannot usually buy access to them like you can to a book, but that doesn't mean that their creators automatically consented to AI training from their work.
- And, perhaps most importantly, the paragraph completely disregards the actual *creation* of AI art. (Yes, that wasn't the point of the question, but it's the much bigger issue when it comes to AI and copyright, imo!) I could ask AI today to create hundreds of images in the style of Ghibli films, or Banksy murals, or to write me a couple dozen AC/DC songs, or three new Harry Potter books. This is especially a bitter point for me, as a hobby artist: I hate the fact that people can capitalise on something which I do for fun and which takes me *a lot* of time in a matter of a few minutes.
Fair perspective. It's hard for me to hate it when I'm still completely flabbergasted that it's possible. I think if I were an artist (professional or otherwise) I'd feel differently.
Serious question: how much do you hate the AI-generated images that I use for AGI Friday posts?
Also fair.
I'm not a visual artist, I'm a writer, so I don't hate the AI-generated images as much as I probably would otherwise? I am also fascinated by what's possible; and those test images you shared once (astronaut fox etc.) were pretty interesting.
I guess I'm also a bit torn on AI art. On the one hand I regonise the possibilities. Like, maybe in ten years I can upload everything I've written and tell an AI "Write an 8k short story in my style about [whatever]." And then I'd only have to do one edit pass, and I'd be perfectly happy with the result, and I'd be done in two hours instead of eight or more.
But convenience can also be a fallacy: Sometimes, the half-hour agonising over a few sentences is exactly what you need to push the story into an even better direction. Instead, with AI-generated art, the way of less resistance is right in front of you: You just accept what it creates for you, because it's already finished, and it only took a few seconds.
But I do think that AI is going to have a hard time to perfectly emulate all forms of art. Prose is easy enough, poetry following a certain format (sonnets, haiku) is definitely doable, but what about free-form poetry? What about hybrid forms?
I think my fear is that either it will be able to emulate *everything* convincingly, and then people won't bother creating anything themselves anymore, OR it will significantly fail at some things (like very personal/emotional/associative free-form poetry), and then less of that will exist, because people won't be able to create it with AI, and those who could write it themselves would do so less because they could produce more things, and quicker, with the help of AI. And both sound like horrible options to me.
But generally, I'm a fan of lowering entry barriers, especially with art. There's so much art that doesn't get created just because people think "I'm not good enough" / "I wouldn't even know where to begin" / "I don't have the time/money/resources". Inputting prompts into a computer does at least mitigate some of these points, if not all. (You'd still need internet access, and you'd still have to pay something, at the very least.)
I just don't want the norm to be uninspired AI generated art en masse (to not say "slop"), because I think that'll only stifle human creativity instead of inspire it!
I keep thinking of two analogies: books vs. ebooks and traditional craft vs. commodities.
I work in the literary industry, and when ebooks were on the rise, everyone was afraid that the demand of physical books would rapidly decline and basically crush most of the industry, especially bookstores. But that didn't happen, at least not here in central Europe.
On the other hand, most people buy their furniture from Ikea and similar stores. If you own handmade furniture (or handmade ceramics, glasses, whatever), you likely either A) inherited it, or B) are wealthy enough to afford it.
I just don't want human art to end up like traditional craft items—only for a wealthy elite. I'd much rather it turns out like paper books—at most slightly more expensive than ebooks/AI art, but still widely spread and cheap enough that most people can afford it.
Great points. I do think AI slop is a big problem right now. I genuinely don't know if it will get worse as slop becomes harder to identify or if it will just get good enough that the term "AI slop" won't be apt. I admit to being a little horrified by your hypothetical of an author of fiction letting an AI write a new story emulating the author's style, editing it a bit, and publishing it. I might be being naive but it feels like that couldn't possibly be a genuinely good story, pre-AGI.
(And yet, it also kind of feels like we're on track to that being possible in a few more years. Which, yes, might be because we're on track to AGI in a few years.)
As for visual art, my very non-artist opinion is that the people who think AI art is ugly are refusing to admit that the emperor has no clothes. I mean, AI may be commonly used to create ugly art but the critics of AI art generally mean more than that. That they can spot AI art a mile away and they despise it. I think anyone in that category should try the following AI art Turing test: https://ai-art-turing-test.com/ (note that it's already a year old so presumably AI art is that much better today).
"I genuinely don't know if it will get worse as slop becomes harder to identify or if it will just get good enough that the term 'AI slop' won't be apt.« — Hypothetical worst of both worlds: good enough (in technique) that 'slop' won't technically be apt, but also abundant enough *everywhere* that it will still drown out most of the unique, creative, inventive human art out there, simply because you're gonna be flooded by a million versions of the same idea.
"I admit to being a little horrified by your hypothetical of an author of fiction letting an AI write a new story emulating the author's style, editing it a bit, and publishing it. I might be being naive but it feels like that couldn't possibly be a genuinely good story, pre-AGI." — Well, depends on your definition of "genuinely good story," maybe? (Also depends on the editing skills and how much editing is needed ;-) There's tons of middling-to-bad stories/books out there, and that was the case even before the time of the internet or the possibility to self-publish books. Maybe AI pre-AGI couldn't really, convincingly emulate Poe, but there must be hundreds (thousands?) of Poe-inspired short stories or re-writings online, probably ranging from "horrible" and "bad" to "okay I guess", with only a few exceptions that are "good" or "great".
I never tried a writing-AI, but I don't see it having tons of problems with replicating already-not-all-that-great content. I guess for me the scary aspect isn't that AI art might suddenly create the new Picasso (or Proust) and put all human art to shame, but rather that it will create *so much* middling/kinda okay/not really good art, that it will become more difficult for the new human Picasso (or Proust) to gain recognition / they will become more easily discouraged, etc. [all very broad generalisations, of course, and it's not at all unproblematic to rank art by "objective quality," to praise "canonically great" art, etc. etc.]
"AI may be commonly used to create ugly art" — Yeah, I think that's the key problem, that most AI art that we see nowadays (I was SO baffled when I found out there's uncanny AI videos *everywhere* on social media! I guess I prefer to live under rocks sometimes!!) is ucanny and/or bad and/or uninspired, and so most people think that's all you can do with AI art! I do think that some (most?) of the critics of AI art pretend they can spot it a mile away, just because they think AI art is only 6-fingered hands and weird perspectives. But maybe I'm wrong.
I did take the AI turing test a while back and was absolutely baffled by it, I'll take it again now, because I'm curious if the results will be different. (My memory is very bad, so I won't consciously remember the solution of specific pictures, so I can't "cheat").