The problems with AI art

[Kuroneko]

https://paizo.com/community/blog/v5748dyo6si91

<3

[Sludgewave]

The big problems I see are:
1) It'll put artists out of work, unless they can make it in the "get rich people to pay for your human art as a luxury". (Which requires there to be rich people with lots of luxury income in the first place).
2) People who have control of the AI will become wealthy off its use, when all they do is own it.
Really, just shows the real core problem is capitalism, which itself is something to fight against for a myriad of other reasons.



As far as the terrors of AI go, I'm more worried about the implications of its use in disinformation/propaganda, or the idea that once we create Skynet it'll be all transphobic or something.

[Lilias]

This is about chatbots, rather than art, but I think it nails the problem with AI.

https://www.christiancentury.org/article/features/problem-artificial-intelligence-us

[LemonMelon]

In general I'm not against AI-generated artwork—I'm even moderately invested into it myself—but in my opinion it kills most incentives for humans to pursue a career in art. I'm certain there are many would-be artists who enjoy drawing/painting or even doing character 3d modeling, but who at the same time couldn't justify investing many hours daily if they didn't have at least the slim hope that at some point in the future they could earn a living with their skills, doing what they like. Much of this is now replaced even at a professional level with rather boring "prompt engineering". If one's goal of making art is just obtaining a pretty picture, certainly AI-generation allows to reach "close-enough" results in a fraction of the time and costs.

A second related problem I have with it is that there's not really much to be learned in terms of fundamental skills with prompt engineering. Furthermore, the current interfacing methods for generating AI artwork will certainly change in the near future, potentially making soon useless again the time invested into learning what effect words (or their combinations) on the final output.

[Vekseid]

Prompt engineering is a kind of funny concept. It exists because the AI still doesn't possess a genuine model of what it is creating. Words are loosely associated with objects presented in its training data, but it doesn't actually have an inherent conception of that object, and there is a lot of noise in the training.

And if what you want isn't in the training data it becomes rather useless.

Stuff is going to start getting crazy (and possibly scary) once these things are used to build models instead.

[Dice]

Here's a short post I wrote a bit ago about this from the creators side and my views on it, call me old fashioned, but I think this true.

I can't draw, at all. But I still partake in a form of art. I sat in the darkness, in a forest for hours to take three photos. (Many photos, stacked. End product was 3 images) If you can do that in a few seconds with a touch of your keyboard you don't get to understand what it is to build a craft.

There is no soul is something that is made by throwing darts at a board and keeping the one you like. There is nothing like working out ahead of time what you plan to show, framing, taking steps towards a goal and then learning from your mistakes. Adjusting your skills in the future.

AI art's true flaw is that it asks so little from the "creator" and thus leaves very little in the ways of development, self reflection or healing that other forms do. Asking someone to sit alone with their thoughts for hours innately requires time for self reflection, understanding and care. Prompted by one's art is more meaningful then prompt and pray.

[GloomCookie]

Framed in that way, I can understand some of the reluctance by artists to accept AI art. It's similar to my skills as a designer/engineer. We have a lot of tools at our disposal, and we're developing new tools to try and make our jobs faster, but they'll never replace the actual person, at least not for a while.

I get it though. There are a lot of designers out there who are so overworked that they'll take any shortcut they can, including having some tools that just drop their work in and call it good, without having to do much more than make a few connections. And for some situations, that's perfectly acceptable, since they're routine and even boring because they do the same thing time and time again. But, the program we're developing doesn't know what to do after that, and that's where the designer must step in.

What I think will end up happening is that a lot of people will look at AI art as "Good enough" for their average purposes and then when they need something that AI can't do, they'll start turning to the real thing. If, say, the local university wants to add some people to make their brochure look better (That happened once at the University of Wisconsin) then AI might be good enough. But if you want something super specific, then you'll need to get someone who is capable of creating actual art.

That's kinda how I see things. Where I think AI will flourish is as a personal assistant, able to start custom tailoring recommendations for specific applications. For example, a lawyer might use it to look up similar case law, or someone like me might use it to research codes and similar projects. I doubt we'll ever get rid of AI given how much money's been poured into it at this point, but we can certainly curtail some of its more negative aspects.

[Al Terego]

That's kinda how I see things. Where I think AI will flourish is as a personal assistant, able to start custom tailoring recommendations for specific applications. For example, a lawyer might use it to look up similar case law.

Why not allow LLMs custom-trained on legal data to substitute lawyers?  After all, if ChatGPT4 can pass the bar exam in the top 10%, why must one pay a (statistically inferior) lawyer $400/hour, when they can get (statistically better) results for $0.12/1K tokens?

Oh right, because the legislators will ban that use to protect their caste.

[Keelan]

Why not allow LLMs custom-trained on legal data to substitute lawyers?  After all, if ChatGPT4 can pass the bar exam in the top 10%, why must one pay a (statistically inferior) lawyer $400/hour, when they can get (statistically better) results for $0.12/1K tokens?

Oh right, because the legislators will ban that use to protect their caste.

ChatGPT can't represent me in court as far as I know. It also can't help out at all during trials.

ChatGPT also doesn't necessarily know when and what to file or how the filings have to be formatted and bound (yes, that does matter for some courts), nor necessarily how many copies you're required to file.

ChatGPT can write up a contract, but if you don't know how to read contracts you can end up fucking yourself over and words in law often have very specific meanings.

ChatGPT can write up a contract or any legal document, but are you 100% sure that a document it created is in fact accurate? Do you know enough about the law to be able to double-check the ai?

ChatGPT would potentially require intimate knowledge of evidence potentially which would be a MASSIVE legal issue since in order to feed it to ChatGPT I'm pretty sure you run into evidentiary issues and can potentially ruin your case.

My point being that as much as lawyers tend to rub people the wrong way - and I've worked for lawyers before - they do actually provide a useful skillset in dealing with legal systems. This isn't to say that ChatGPT won't have utility - it absolutely will - but I doubt that we'll be able to completely price them out.

[Al Terego]

I specifically wrote "LLMs custom-trained on legal data".

And the fact that it "can't represent me in court [and] can't help out at all during trials" was the whole point of my post.

[Oniya]

And the fact that it "can't represent me in court [and] can't help out at all during trials" was the whole point of my post.

It's really hard to get that from your prior post, which stated:

Why not allow LLMs custom-trained on legal data to substitute lawyers?  After all, if ChatGPT4 can pass the bar exam in the top 10%, why must one pay a (statistically inferior) lawyer $400/hour, when they can get (statistically better) results for $0.12/1K tokens?

Oh right, because the legislators will ban that use to protect their caste.

It's not that ChatGPT isn't allowed to represent someone in court, it is that it is functionally unable to effectively do so, due to the other four points that Keelan made.  I would like to see an actual source stating the 'statistics' that you loosely reference here.

[Keelan]

I specifically wrote "LLMs custom-trained on legal data".

And the fact that it "can't represent me in court [and] can't help out at all during trials" was the whole point of my post.

And my point was that if you're not an LLM (Master of Laws) and you're using an LLM (Large-language model) for legal matters then there is potential for a lot of problems that can be caused by one's ignorance of law and legal procedure, or alternatively one's stupidity with regards to how they believe the law *should* work.

Also the stuff relating to confidentiality, evidentiary protocols, disclosure policies, privacy, security, etc.

And while your whole point seemed to center upon the "Legislators [banning] to protect their caste", my point of why it wouldn't work stemmed largely from the fact that ChatGPT is not a legal person and has no legal standing. I cannot secure ChatGPT as counsel not because they want to 'secure their caste', but because my counsel has to be a legal person registered to practice law in the state and in that area of law.

IF ChatGPT were able to be used in the courtroom, you would legally be representing yourself as you do not have counsel representing you. This means that Average Joe with ChatGPT is up against Bigshot Lawyer with ChatGPT in court, or worse yet Bigtime-Election-Year Prosecutor with ChatGPT (because why wouldn't they use that tool too?). You likewise then couldn't appeal the results due to ineffective counsel because you never secured counsel and were representing yourself.

[Al Terego]

Reply to Oniya

Hmmm...  Alright, I'll attempt a better presentation.

Fact: OpenAI claims that ChatGPT4 scored in the top 10% of the Uniform Bar Exam (source: Forbes).

Conjecture: If we assume that a lawyer's score on the bar exam has some correlation with how "good" of a lawyer they are, it follows that ChatGPT4 is "statistically better" than most lawyers.

Now, I do realize that experience plays a large role, and that's why the rest of my comments below will not refer to ChatGpt4 specifically but to a hypothetical LLM extensively trained on legal data.  Given the huge amount of court transcripts, case law, and other documents that is available, this is quite doable.

Assertion: My personal opinion is that there isn't anything that a defence lawyer can do that a specialized "AI" cannot.  I believe that 10 years from now (a very conservative estimate given the breakneck speed of this field's advancement), machines will outclass human lawyers just as they outclassed chess and go players (and soon to outclass Starcraft II players).

Facts: Lawyers charge hundreds of dollars per hour, and the adversarial legal system can make the ability of the lawyer matter more than the facts of the case.  I have quoted the pricing of ChatGPT4 in my post, they are significantly lower.

My conclusion: An AI can do a better job as a defence lawyer than most human lawyers for a fraction of the cost.
My prediction: However, given that the legal system is a protection racket, people will never be allowed to take advantages for an AI without paying a lawyer regardless.

Feel free to poke holes in my arguments, I've been known to change my views in light of sufficiently compelling evidence.

[Al Terego]

And my point was that if you're not an LLM (Master of Laws) and you're using an LLM (Large-language model) for legal matters then there is potential for a lot of problems that can be caused by one's ignorance of law and legal procedure, or alternatively one's stupidity with regards to how they believe the law *should* work.

In legal matters, I am stupid.  But a properly trained AI won't be.

Also the stuff relating to confidentiality, evidentiary protocols, disclosure policies, privacy, security, etc.

All of them are rules that AI can be trained on.

And while your whole point seemed to center upon the "Legislators [banning] to protect their caste", my point of why it wouldn't work stemmed largely from the fact that ChatGPT is not a legal person and has no legal standing. I cannot secure ChatGPT as counsel not because they want to 'secure their caste', but because my counsel has to be a legal person registered to practice law in the state and in that area of law.

That is circular reasoning.

IF ChatGPT were able to be used in the courtroom, you would legally be representing yourself as you do not have counsel representing you.

Sure.

This means that Average Joe with ChatGPT is up against Bigshot Lawyer with ChatGPT in court, or worse yet Bigtime-Election-Year Prosecutor with ChatGPT (because why wouldn't they use that tool too?).

The AI would be the equalizer.

If we face each other in hand-to-hand combat, the bigger, stronger, and better trained person will have a decisive advantage.  Give both of us a machine gun, and that physical advantage goes out the window.

I do not believe that Bigshot Lawyer + AI will do much better than AI unhindered by Bigshot Lawyer anymore than a chess International Master (Elo rating below 2500) aided by the Stockfish engine (Elo rating above 3500) can play noticeably better than Stockfish alone.

This means that Average Joe with ChatGPT is up against Bigshot Lawyer with ChatGPT in court, or worse yet Bigtime-Election-Year You likewise then couldn't appeal the results due to ineffective counsel because you never secured counsel and were representing yourself.

Caveat emptor.

[Keelan]

In legal matters, I am stupid.  But a properly trained AI won't be.


All of them are rules that AI can be trained on.

And how do you know if your AI is properly trained if you know nothing of the law? Do you even know what questions you need to ask to make sure that you're getting the right answers? If it tells you you need to file a motion, but you actually need to submit a writ but had phrased your question improperly, would you realize it?

That is circular reasoning.

No, it is not. If I am arrested and I invoke my right to an attorney - to legal counsel - that means I'm asking for SOMEONE that will advise me on the law and speak on my behalf.

Plants, animals, and objects cannot speak on my behalf as they cannot speak, comprehend language, nor comprehend law.

Computers cannot speak on my behalf as they require human input, are not sentient, and cannot be held liable for any statements they may make, and thus if I request counsel and they bring me a computer, I am my own counsel and I am the one responsible.

ChatGPT or any other LLM requires user input to output answers. It cannot spontaneously tell you to shut the fuck up if you start running your mouth, nor can it answer questions for you without input. At best you get asked questions, you get ChatGPT to produce your response with a 'high degree of accuracy', and then you review and submit it.

You are essentially constructing written self-statements every time you use ChatGPT - thus denying yourself the right to remain silent in the process - and you better pray to GOD that each instance of a statement by you matches up with each other because if it doesn't, well ChatGPT can't be liable because it's just a tool and you had every chance to review the statements before you submit them.

The AI would be the equalizer.

If we face each other in hand-to-hand combat, the bigger, stronger, and better trained person will have a decisive advantage.  Give both of us a machine gun, and that physical advantage goes out the window.

I do not believe that Bigshot Lawyer + AI will do much better than AI unhindered by Bigshot Lawyer anymore than a chess International Master (Elo rating below 2500) aided by the Stockfish engine (Elo rating above 3500) can play noticeably better than Stockfish alone.

No, that very specific physical advantage goes out the window; you take some random dude with no firearm experience and give him a M249 and put him up against a highly-competent 0331 Veteran and something tells me that the guy who has used that machinegun for several tours is probably still going to have the advantage.

(This is why everyone who is a halfway decent firearms advocate will tell you if you have $300, it's better to buy a cheap gun, ammo, and get more practice putting rounds down range than it is buying a nicer gun and being able to do less - maybe even no - training.)

Same with the law; random dude with what is essentially the best legal template generator and legal dictionary up against an experienced lawyer with that same legal template generator and legal dictionary?

Not to mention that the same tool that can be used to fight FOR you can be just as effective at CRUSHING you; a substantial component of law EVERYWHERE is the ability to argue your case. LLMs would be completely mercenary in that regard.

And that's not even getting into the fact that unlike chess - which is a highly controlled, highly systematic, highly regulated and highly controlled environment that ultimately has no room for human error intrinsic to it - even with ChatGPT as a replacement for lawyers you STILL have Judges and juries (and other human components of the legal process) where applicable, whom are still human.

The only way this theoretical 'let's get rid of lawyers' notion works is if you replace the ENTIRE legal system with what would essentially be a fully-automated techno-tyrannical judicator which could be easily programmed with biases (or simply have them in there accidentally), and otherwise would be prone to exploitation by state powers that be.

As much as I find lawyers expensive and some to be quite scummy assholes, I would much rather them than some faceless series of if-then-else equations state-obediently ruling with the compassion of a functional psychopath against anyone the state ultimately decides to legislate and weaponize this 'totally impartial' legal network against.

[Oniya]

I do appreciate that Mr. Koetsier (who does not claim any legal experience) provided the data on which exams the ChatGPT was tested on.  I am also not a lawyer, nor have I stayed in a Holiday Inn Express recently.  I am, however, a keen observer of people, and spent a lot of time watching real lawyers in action.

ChatGPT or any other LLM requires user input to output answers. It cannot spontaneously tell you to shut the fuck up if you start running your mouth, nor can it answer questions for you without input. At best you get asked questions, you get ChatGPT to produce your response with a 'high degree of accuracy', and then you review and submit it.

This is something that I have observed lawyers doing for their client's benefit, particularly when the defendant chooses to take the stand in their own defense.  Our hypothetical AI would have to be able to listen to its client's off-the-cuff testimony and anticipate that the client is entering dangerous territory before it happens.  While an AI trained on legal matters might have the ability to predict legal arguments, would it have the training necessary to - for example - anticipate that its client would invoke some half-baked conspiracy theory on the stand?  A human lawyer would have the experience to discuss testimony with their client before putting them on the stand, and the caution to ask questions that don't lead to 'surprises'.  One of the common bits of advice I've heard is 'don't ask a question that you don't already know the answer to'.

[Keelan]

So, this is relevant I think, though I haven't read it in full:

https://www.reuters.com/technology/plagiarism-machines-hollywood-writers-studios-battle-over-future-ai-2023-05-03/

[Oniya]

When I was younger (so much younger than today), I read Robert Heinlein's 'The Moon is a Harsh Mistress', which includes a supercomputer that becomes sentient.  There's an entire section where the protag is trying to explain to this AI 'what is funny'.  That particular bit comes to mind because it was so difficult to distill comedy into a set of rules (which is something that the writers could possibly use to their advantage.)  If I haven't missed anything, the last time that the WGA went on strike, we ended up with 'reality' TV, since that was pretty low-effort as far as actual plot arcs are concerned.  AI might be able to handle the plotting of your average daytime drama, but can it produce moderately reliable humor? 

"Are two types of jokes. One sort goes on being funny forever. Other sort is funny once.
Second time it's dull. This joke is second sort. Use it once, you're a wit. Use twice, you're a halfwit."
"Geometrical progression?"
"Or worse."

[Al Terego]

@Keelan

It seems that we are arguing different things.  I am not advocating for Dick The Butcher's suggestion, just for the option for a person that chooses to represent themselves to be able to use AI as an aid.

And how do you know if your AI is properly trained if you know nothing of the law?

And how to you know that your lawyer has been properly trained?

Or what if you don't have the means to afford an attorney (like sample source)

No, it is not.

It is.  I am arguing against the established rules, and you are saying that my position is invalid because it is at odds with the established rules.  This is the textbook definition of circular reasoning.

If I am arrested and I invoke my right to an attorney - to legal counsel - that means I'm asking for SOMEONE that will advise me on the law and speak on my behalf.

See above for legal counsel affordability.

Plants, animals, and objects cannot speak on my behalf as they cannot speak, comprehend language, nor comprehend law.

A person can choose to represent themselves.  I argue that the use of computerized aid should be allowed.

ChatGPT or any other LLM requires user input to output answers.

Not anymore, and definitely not 10 years from now.

It cannot spontaneously tell you to shut the fuck up if you start running your mouth

Just add voice recognition.

nor can it answer questions for you without input. At best you get asked questions, you get ChatGPT to produce your response with a 'high degree of accuracy', and then you review and submit it.

Are you basing your argument on the limitations of a single implementation (ChatGPT) at his point in time?  Just take a look at how fast the technology evolved since its inception, and it will evolve even faster now that it got the interest of the open source community.

[...] ChatGPT can't be liable because it's just a tool

On this we are in agreement.  My point is that one should be allowed to use this tool if they wish to.

But why limit the discussion to criminal defence in court?  Why not use software to formulate contracts, write wills, find and cite applicable case law, etc.?
There are many fields that require a lawyer who basically acts like a glorified expert system, with the additional fallibilities stemming from being human.

Several years ago I was involved in collision and was charged under the traffic act.  I represented myself, did my homework, spent time in the law library of the local university researching case law, and ultimately prevailed.  Having a computerized agent to assist me could have saved me quite a bit of time and a some anxiety.

Most lawyers to not argue at murder trials.  Hell, only half a percent of US lawyers are trial lawyers.

No, that very specific physical advantage goes out the window; you take some random dude with no firearm experience and give him a M249 and put him up against a highly-competent 0331 Veteran and something tells me that the guy who has used that machinegun for several tours is probably still going to have the advantage.

Not if the machine gun is fully autonomous.

Same with the law; random dude with what is essentially the best legal template generator and legal dictionary up against an experienced lawyer with that same legal template generator and legal dictionary?

I claim that in 10 years time, the "legal template generator" will be better than a human lawyer.

As much as I find lawyers expensive and some to be quite scummy assholes, I would much rather them than some faceless series of if-then-else equations state-obediently ruling with the compassion of a functional psychopath against anyone the state ultimately decides to legislate and weaponize this 'totally impartial' legal network against.

I did not suggest this.  Let's keep strawmen out of the discussion.

[Al Terego]

So, this is relevant I think, though I haven't read it in full:

https://www.reuters.com/technology/plagiarism-machines-hollywood-writers-studios-battle-over-future-ai-2023-05-03/

They are scared for their jobs, and rightly so.

But let's look at some parts of the article:

The union is also arguing that existing scripts should not be used to train artificial intelligence, which would open the door to intellectual property theft.

“We call it the Nora Ephron problem,” August said, referring to the writer of romantic comedy hits including When Harry Met Sally and You’ve Got Mail. “One can imagine a studio training an AI on all of Nora Ephron’s scripts, and having it write a comedy in her voice. Our proposals would prevent that.”
WGA chief negotiator Ellen Stutzman said some members have another term for AI: “plagiarism machines.”

I would say that the same criteria should be applied to determine if a work is infringing regardless of whether it was produced by a human or otherwise.
What prevents a human writer from reading all of Nora Ephron's scripts (or watching the movies) and then producing a work in a similar voice?

“We have made a reasonable proposal that the company should keep AI out of the business of writing television and movies and not try and replace writers,” she said.

Where were you when manual workers started being replaced with automation?

[Skyguy]

I like the idea of AI art. I think it’s got huge potential to allow a lot of people access to get pretty decent quality art for their ideas that they might not otherwise have been able to get. It’s not on the level of a properly drawn and commissioned piece but if I just need a quick picture for an idea it’s great.

I don’t like the tech bros making the AI. I think it was the CEO of Midjourney was interviewed about the ethical issues with having scrapped millions of images without consent to train his AI and now it was cranking out images that were pretty damn plagiaristic. His response was a decidedly uninspiring blend of “No one told us we couldn’t,” and, “LOL already did it.” The main impression I got was that he’d barely even considered it, much less the implications of having ripped people’s art off to make his AI and now he was selling access to it. Like most of their breed, tech bros, there is zero consideration for the ethical implications of anything and barely any for the legal ramifications. Most of the explanations I hear for the ethics are “wells someone’s gonna do it so it may as well be me getting rich off it,” and the legal, “No’s one’s written a law that specifically makes the legally dubious thing I’m doing illegal so that means it’s totally cool.”

I cannot believe I’m saying this but I honestly hope that Getty eats Stable Diffusion’s face.

[ninjawriter]

Where were you when manual workers started being replaced with automation?

You are literally the only other person I have heard make this argument.  One of the frustrating things, for me, about this discussion is that it has already happened again and again to people in other professions.

[ninjawriter]

I believe the biggest issue regarding AI art is that the arguments that are being used to "defend" the artists' rights don't protect any artists.  There are two arguments I see advanced:

1.  Artists whose works are used for training should be paid.

2.  Styles should be protected as intellectual property.

Both concepts are novel.  Until a year ago, the idea that either of these things should be true for any class of artist wasn't on my radar because all artists train from existing art and styles have never been a protected class.  I'm allowed to read a bunch of Stephen King book and write a book that is stylistically identical to Stephen King!  The general artistic term is pastiche, and while it is usually pejorative, it is unarguably legal.  Or it has been legal. 

Regardless, imagine if those two concepts become legal.  A person might say, "Well, we only want to apply these laws to AIs."  This doesn't even stop AI art, of course, it just means that the people who own the most images have the best AI for a while.  Disney trains its AIs on its database of stories, scripts, comics, books, etc., and its vast art collect - many millions of high-quality pieces of art.

And then they come after everyone else.  Everyone who makes a character LIKE a Disney princess might find that their work is now Disney's work because they trained themselves on Disney art and copied the Disney "style," even if the character isn't particularly like a Disney character.  They will be the people with the resources to "prove" that you "stole" from them.

These arguments do not help artists.  This is why, I think, the big media companies are largely silent.  They're okay with what happens either way.

[Dice]

I believe the biggest issue regarding AI art is that the arguments that are being used to "defend" the artists' rights don't protect any artists.  There are two arguments I see advanced:

1.  Artists whose works are used for training should be paid...

Both concepts are novel.  Until a year ago, the idea that either of these things should be true for any class of artist wasn't on my radar because all artists train from existing art and styles have never been a protected class...

These arguments do not help artists.  This is why, I think, the big media companies are largely silent.  They're okay with what happens either way.

This is, bluntly, wrong. There is a reason why when the company made a version of the program that made music they did not fuck with music companies. The issue is not that our work is not protected and we are not owed for the rights of that which we create, it is that we as a whole are not powerful.

But clearly big media companies have in part had their say because when it came down to it, the folks running this clown show stayed the fuck away from the music giants for a reason.

[ninjawriter]

This is, bluntly, wrong. There is a reason why when the company made a version of the program that made music they did not fuck with music companies. The issue is not that our work is not protected and we are not owed for the rights of that which we create, it is that we as a whole are not powerful.

But clearly big media companies have in part had their say because when it came down to it, the folks running this clown show stayed the fuck away from the music giants for a reason.

What company is this?  I know several companies making generative music AIs, including Google.  I can think of no company that stopped working on music AI because they were worried about music companies.