Why do people treat AI as sentient?

[LAwLz]

21 hours ago, Crunchy Dragon said:

Correct, however in this sense, the human is doing "all the work". They found what they wanted a picture of, adjusted all the necessary settings, and pushed all the necessary buttons. Human input and action is necessary through every step of the process.

The human is doing next to no work.

They "found" what they wanted a picture of, more often than not by random chance. They didn't create what they picture either, because they are just snapping a scene that was already created for them (unless they model the scene themselves, but let's say we're talking about nature photos). I didn't create the sunset just because I took a picture of it. I just happened to see it and decided to save what I saw by pressing a button on a camera.

They might not have adjusted any settings, the camera can do that for them.

"All the necessary buttons" is a single button.

 

The human input is only necessary on the "input" step of the process. The actual creation of the image is 100% done by a machine for the human. Have you ever looked at what a JPEG image looks like, like the actual file? A human did not write it. A machine did.

 

You're trying to make "point a camera at a subject and click a button" sound like a massive undertaking that has a ton of involvement from the human. Meanwhile you are trying to make AI art generation sound like it is all done by the machine and the human has no control over it.

You are overplaying the human involvement in one scenario, and underplaying it in another one. 

 

 

 

 

21 hours ago, Crunchy Dragon said:

My argument here is that with Photoshop, photography, drawing, painting, composing, etc, every single decision is made through human action.

Let's think about this sentence for a second. Are humans really involved in "every single decision"?

Do you think that when I created a picture in Photoshop I calculated the advanced mathematics necessary to draw a line in a .psd fine? Or did a machine do that for me?

When I clicked export, do you think I decided that this is what the file should look like at these address locations, or did a computer do that for me:

^ This is what the output from Photoshop actually looks like ^

Did I decide these things, or did a computer do it for me based on the input I gave it earlier and the computer "translated" my input into an actual output?

 

When I use frame interpolation in an animation software, am I really deciding what those frame should look like?

When I use the blur filter in Photoshop, am I really choosing the values of each individual pixel, or is Photoshop deciding that for me based on a rather vague input I gave it?

When I move my camera, am I really deciding where the sensor is located in the real world? Remember, my camera has OIS which physically moves the sensor by itself if the camera decides that I am "moving it wrong".

When I decide to click the shutter button, am I really deciding what I am taking a picture of? Remember, the viewfinder is just an approximation of what the sensor sees. Just because one pixel is colored a certain way in the viewfinder does not mean the image I take will have the same pixel colored the same way. In some cases (quite a lot of cases), the viewfinder doesn't even map 1:1 to the sensor. A lot of viewfinders, especially on DSLRs, do not have 100% coverage so things won't even be positioned the way the camera told you they would before pressing the shutter button.

 

As soon as we start getting into the technical aspects of things, and examining the source code for the software running on these computers, we quickly find that humans actually have very little control over what software produces. That goes for AI art generation, Photoshop, cameras, 3D programs and so on. The computer just takes vague inputs from us humans and then processes and translates those inputs into an approximation of what it thinks we want.

When I do frame interpolation I have a vague idea of what I want it to look like, and hopefully the computer matches that well enough for me to go "ok that's good". When I apply a blur filter in Photoshop I have a vague idea of what I want it to look. I would have no idea how to do it myself if you told me "edit the RGB values of each individual pixel, don't rely on an algorithm in a program someone else wrote for you". I have a picture of what I want the blur to look like in my mind, but I would have no way of actually creating that with my hands. I tell a program to do it for me, and nobody has so far ever gone "you didn't create that" just because I used a function in Photoshop to do it. Yet now that AI art programs exist all of a sudden using algorithms to generate the result you want becomes highly questioned (but only if "AI" is included in the sentence).

 

 

 

22 hours ago, Crunchy Dragon said:

If I asked you to draw me a picture of "a reindeer riding a can of Miller Lite across the Milky Way", then walked away and left you to draw a picture based on my prompt, I have no creative input or control beyond just giving you the input and waiting for the output. Every creative decision is made by you, with no influence from me other than the prompt I gave you.

I would argue that "a reindeer riding a can of Miller Lite across the Milky Way" qualifies as "creative decision", and that prompt is a lot of "influence".

I mean let's break it down. You want to create a picture of a reindeer riding a can of Miller Lite across the Milky Way. You start the program, decide which settings to use, input the commands (with your mouse and keyboard) and then click generate.

Let's break down how it would have been done in Photoshop. You star the program, decide which settings to use, input the commands (with your mouse and keyboard) and then click save.

 

And before you go "but with photoshop you would have had more input!"

1) Trying to argue that the amount of input is a deciding factor opens up a whole can of worms, especially when it comes to photography where a camera where someone literally only has to press a single button and the camera itself, using AI, will generate an image based on hundreds of parameters that are automatically tweaked and decided.

 

2) The amount of input the human has varies both in an AI program as well as Photoshop. Would you argue that I didn't create this image in Photoshop?

It was literally a single button press. It took me 5 seconds to make.

 

 

Meanwhile, I made this image in Stable Diffusion. It took me about 30 minutes because I had to tweak various settings (find a good model, apply a VAE, I did some inpainting, might have applied a LoRA, tweaked the prompts over 20 times, tested various samplers and sampling steps, and so on):

 

 

Would you claim that I "made" the first image, even though I had very little control over the output (I just clicked a button and it created roughly what I wanted, I certainly didn't decide the RGB values for the smooth edge the dot has), and would you claim that I didn't make the second image even though I had a very clear idea of what I wanted and tweaked the program until I got it?

[NapalmFirestorm]

The goal of the AI you are referring to is to be as useful and sentient (in regards to it's purpose) as possible. People have those complaints because they want the developers to make adjustments to continue working towards that goal and improving the AI tool. I'm not sure I find the correlation between blaming the photographer as compared to the AI not doing it's job right because negative feedback of the AI is in regards to its mistakes or poorer programming, obviously stating possible flaws that they want the developers of the AI to fix. People want the tool to be improved and there is nothing wrong with that. They don't actually believe the AI is it's own conscious person, nor are they really treating the AI as sentient.

[BlueChinchillaEatingDorito]

The kind AI people think of in their dystopian future:

 

[Mark Kaine]

On 1/23/2023 at 1:01 AM, Crunchy Dragon said:

Correct, but the people are not creating the output. They tell the machine what they want, and the machine tries to create it.

well yes, but the content that its feed on matters too and that is largely human input (even when automated) i think that is the current Achilles heel of "AI" systems like that... i also think this isn't anything new, whats new is that computational and algorithmic advances make this now accessible on a casual level..  which is interesting, but the ways it's used aren't really interesting,  it just means more low quality content.  ie a game made by "AI" might work well, but it'll lack the spirit that makes a game actually good and enjoyable -- of course that also depends on the kind of game, something like "FIFA" will always be boring samey and uninspired, may as well use "AI" to make it.  A "3D action hero" game on the other hand might fall flat without actual human inspiration (i mean those are rarely good, but if they are it's usually something special,  long-lasting) 

[Crunchy Dragon]

41 minutes ago, Mark Kaine said:

 ie a game made by "AI" might work well, but it'll lack the spirit that makes a game actually good and enjoyable

Yeah, ultimately I think my biggest gripe with AI is how it lacks the human element.

 

No computer can ever be truly random or spontaneous in the way the human mind is.

[poochyena]

22 minutes ago, Crunchy Dragon said:

No computer can ever be truly random or spontaneous in the way the human mind is.

Actually, no, computers can be MORE random than humans. Humans are actually kinda bad at randomness.

 

[Mark Kaine]

15 minutes ago, poochyena said:

Actually, no, computers can be MORE random than humans. Humans are actually kinda bad at randomness.

you missed this important part in the quote "in the way the human mind is"... humans aren't truly "random" there'll always be a thought behind human "randomness", thats how creativity works and how *new* things appear often seemingly out of nothing,  but there will always be some inspiration behind it, a computer can't do that, it can pick random numbers,  much better than a human, but its not the same thing.

 

42 minutes ago, Crunchy Dragon said:

No computer can ever be truly random or spontaneous in the way the human mind is.

actually i believe computers can be "intelligent" like in science fiction movies... this is usually something humans are very good at predicting... it might seem outrageous or "impossible" at first but in the end these things often become reality... but that will take time,  probably like 100 years and even then maybe computers can be truly intelligent but on the level of a mouse or other animals, doesn't mean they're automatically equal or smarter than humans... actually not just simulating intelligence is very delicate,  but i dont think its impossible and current "AI" is only a tiny step in that direction,  if at all.

 

 

 

 

 

[LAwLz]

20 hours ago, Crunchy Dragon said:

Yeah, ultimately I think my biggest gripe with AI is how it lacks the human element.

 

No computer can ever be truly random or spontaneous in the way the human mind is.

Sorry, but that sounds like pseudoscientific humbug.

"It lacks the human element, which is conveniently something that I can't and won't define but I will say anything I like has it, and anything I dislike doesn't have it".

 

AI generated art lacks no more "human element" than a photograph, 3D rendering, animation, etc, because those things are primarily done without human input. 99,9% of the word is done by a computer, just like with AI art.

 

People need to stop thinking that humans have some magical and unique property. We are just flesh computers executing mathematical formulas. We're a very complicated flesh computer, doing very complicated mathematics, but that's still all there is to it.

[xFluing]

Oh man this whole AI sentience topic is a huge theme in the talos principle. Amazing game I recommend it to whoever can handle it (its kind of a heavy game)

[graograman]

This whole argument is going in circles. The biggest issue seems to be that everyone tries to use analogies or previous cases to argue one way or another but that wont work because those pre-AI are different to anything humanity has encountered in the past. There is nothing to compare them to.

 

Also, these tools are way too useful and powerful to lock them behind law issues even if it hurts current artists and writers. A more pragmatic approach is needed.
Drawing a picture by itself is not a crime, if they sell it or distribute it, it can be. If the programmers did something illegal to create it, they commited a crime

[Kisai]

14 hours ago, graograman said:

This whole argument is going in circles. The biggest issue seems to be that everyone tries to use analogies or previous cases to argue one way or another but that wont work because those pre-AI are different to anything humanity has encountered in the past. There is nothing to compare them to.

It's gone around in circles several times, across several threads over the last 6 months. 

 

Pretty much, the reason people piss their pants when they saw AI generated artwork is because they thought the AI "created" it. Then it was discovered that, no it was trained on actual artwork, and then want to wave the "copyright infringement" hammer at it, instead of looking inward and going "oh, maybe I shouldn't be inspired by anything, wow, I must actually create things wholecloth and not use any references."

 

Trying to handicap or kneecap the Machine learning process because it's not a human. The AI will never be "Creative" in any sense, but it still needs to learn from what exists "now", not what existed and is more than 100 years old. Do you really want an AI that trained on material where women were not treated as equal to a man?

 

The problem really with letting the AI use uncurated datasets is that it has no concept if that data is bad and should reject it before training on it. 

 

Here I'll give you an example, since this image was floating around twitter for the last 3 days and is a perfect example of a HUMAN artist being tone deaf on purpose.

https://twitter.com/khyleri/status/1619002424143798272

 

Do I even need to being to explain what's wrong with it? Would an AI know what's wrong with it? Would an AI generate something like this image from a prompt? Would an AI get trained off this image in a dataset without knowing what's wrong with it?

 

A human knows the context. An AI does not. An AI will never understand context. So if you tell it something that is obviously hateful, edgy, illegal, or whatever, it's going to try and take whatever shortcuts that spit out something that matches the prompt. The AI doesn't actually understand what you asking it to do.

 

There in lies the problem with having AI's use uncurated datasets. It's very likely the AI will produce hateful piece there is even a single source for it in the data set that it can build off.

 

The real risk to this technology comes from the AI lacking any kind of "sense of right and wrong", like the training software could strip known racist words from the input descriptions, so that there is no possibility of it being able to purposely generate anything with that keyword when it no longer exists in the model. But a better direction would be to curate training data that was made ON PURPOSE to train the AI to undermine the use of "evil" language. So asking it to produce WWII propaganda would instead produce stick figure drawings in crayon.

 

But still, there is no sentience at all in existing AI.

14 hours ago, graograman said:

Also, these tools are way too useful and powerful to lock them behind law issues even if it hurts current artists and writers. A more pragmatic approach is needed.
Drawing a picture by itself is not a crime, if they sell it or distribute it, it can be. If the programmers did something illegal to create it, they commited a crime

The thing is, if left-wing artists had their way, the tools would not exist (nobody would ever release a model or source code if it could be used for "evil" purposes,) and if right-wing artists had their way, they would be locked behind very expensive cloud computing systems that only people with a lot of money could use it. Thus it remains a black box that idiots with deep pockets would use.

 

As things exist right now, a paper is produced nearly every time some "cool machine learning thing" comes out, and a description of what dataset was used. Rarely is the model itself released, and usually when a model is released, it's not the production model that the commercial version uses.

 

Like forget Image and Text AI for a minute and go back to ASR, and TTS AI.

ASR's are ONLY trained on existing works. Mainly audiobooks in the public domain. This results in over-fitting of obsolete language, such as words used to describe race and mental illness. The ASR's are also trained on TedX and Call Center Quality recordings.  TTS AI are trained on the same audiobooks. But unlike the ASR, the TTS ONLY generates words based on the input. It does not make any guesses, but if a word is not in the pronunciation dictionary (which guess what, is largely biased towards American-English) it will make it's best effort at assembling that word based on other diphones seen before in the training. The TTS can say all the profanity that currently exists, or will ever exist, because it's not trained on words, but phones.

 

Image/Art AI, can not do the equivalent of that. There are no basic symbols that make up "art", all the symbols we use in the visual realm are based on how humans experience it. What we have in art as basics are color theory and geometric shapes (eg, roundness, lines, circles, etc.) Asking an AI to produce a human from only a palette, when it doesn't know what "skin color" is, will not help it. Yet even an actual human would not describe a living creature in terms of just colors or shapes. They use words like "sexy" or "cute" and there is no way for an AI to know what that is. Something "sexy" or "cute" is a subjective human emotional reaction.

 

As I've mentioned before, on this site,  You and I can agree what "a dog" symbol is, but an AI does not know what a dog is. You can show it 10,000 images of dogs, but when you ask it to just create a "dog" it will be whatever is the most overfit "dog" it trained from, which may end up being a kitbash of several breeds. Which "dog" is not exclusively used to describe the animal either. The AI doesn't know the difference between a "a hot dog" sausage and "a hot dog" of a golden retriever laying in the sun.

 

But it should also be stated that "curated" datasets for AI's to train on will produce much better results. So A fictional "DogML" might only be trained on dogs, but because it theoretically might be trained on high quality data, it might be capable of generating 10,000 different breeds of dogs in various poses, and not just what LAION scraped from the internet.

 

Personally I feel the "most creative" thing ML can do, is by accident. Like by fine tuning the data set, you can leverage train models that do things that are impossible. Like taking a voice of a popular actor/actress, and have them able to speak 200 languages fluently, and accurately. AI Polyglot's would be immensely useful, and bring us very close to "universal translator" technology. AI art generators are just precursors to "holodeck" experiences. Not replacements for artists.