The AI Red Scare is only harming artists and needs to stop.

Nice strawman, Umbran. Time to torch it.

Sorry, but it isn't a strawman. It is the actual technical operation of the thing.

And go way overboard on the technicalities, since you've triggered Pedant Mode.

Don't blame me. Blame the RIAA. They're the ones who made the argument, and it is a large part of what took down Napster just about a quarter-century ago.

I am sorry if you find arguments used in the major precedent-setting legal cases on the matter to be strawmen. But they are what we must contend with, regardless of your opinion.

The act of copying is inherent in electronic media.

Yes, that's exactly the point. You are making my argument for me.

Likewise, playing an MP3 file from my drive still has to copy a Read Ahead Buffer's worth continuously to actually fit the waveforms.

The act of copying it for use is of need not in and of itself a violation of copyright

First, you complain that I was pedantic, now you require it? Fine.

When you buy an MP3, your license to the content includes all relevant electronic processes necessary for you to listen to the music. When you haven't bought the MP3, or haven't engaged a system that has acquired rights to present the work to you (like, say, Spotify) you don't have license to make electronic copies.

When you write something on EN World, your agreements include all relevant copies for a human to read your posts - but no more than that. So, scraping EN World for content to train a LLM would be a violation of copyright.

Copying the content to train an AI is not, to date, considered "Fair Use". Nor is it included in the license or agreements used in internet systems to show content to people. So, you cannot make electronic copies to do it without permission.

All the rest of your verbiage around this is then extraneous.
 

log in or register to remove this ad

Every argument in this thread is about artists, with maybe some inferences about writers too.
Why I'm saying that artists are acting like they are special is because while all these other jobs were obsoleted, I did not hear artists speaking up for them. And note, I don't think visual artists will ever be completely obsolete. For two simple reasons. It is most likely that we will always have a surplus of human labor versus what human society needs to survive. Therefore people will want to find something to do with their time. And second, humans are creative and inventive, and other humans like to be inspired. This will lead to 'artists' adapting, and finding new ways to inspire others.

Eh, can you support this? (“I did not hear artists speaking up from them”, which I presume you mean “Artists did not speak up for them.”). My presumption is mostly the opposite. I wonder how many Appalachian folk songs are about mining life.

Also, there are many categories of work which has been automated where very real harms occurred to workers such that the automation had a much harder to evaluate overall benefit. For example, textiles manufacturing.

Further, there is a question of pace of change, and questions of the existence of either social programs and of alternate work, which may be the bigger culprits.

TomB
 

There needs to be a grassroots effort to reverse engineer every gen-AI on the market and release functionally-identical open-source software free to the public. If that happened, OpenAi would be worth exactly what they think visual artists should be paid.
Hmm, I was under the impression that the underlying software was open. Or at least widely known. The current major cost is in training the models, which has become horribly expensive computationally, which translates directly into an energy cost.

“Sam Altman stated that the cost of training GPT-4 was more than $100 million.”

GPT-4 - Wikipedia.

TomB
 

Erroneous conclusions with incorrect citations is how I'd categorize BOTH types of errors... and both happen in humans and in LLMs.
Okay, while I don't really know anything about the actual process of how a LLM AI assembles a text, by my understanding of the term "statistical", I would suspect that it doesn't infer conclusions. It asks "what would probably come next in this text" without connecting it to any real-world referent. In that Star Trek example (I really should have copied the text), it probably "thought": "In this kind of text, at this point, there would be a reference to an episode, so I'll include one." But I'd surmise that it has no way to understand that a reference relates to something extra-textual. This is fundamentally different from understanding that your text references something extra-textual, but to get the reference wrong. In a similar way, when bodies merge with objects in AI images which are supposedly foto-realistic, that's probably because the AI doesn't understand that the thing depicted in the image references an actual hand.

Whenever it becomes glaringly obvious that something is AI art or an AI composed text, it is because the AI can't really draw a connection to extra-textual reality.
 

Hmm, I was under the impression that the underlying software was open. Or at least widely known. The current major cost is in training the models, which has become horribly expensive computationally, which translates directly into an energy cost.
Which leads to another concern both ethical and practical: Do we really want to keep spending enormous amounts of energy on creating AI images?
 

Hmm, I was under the impression that the underlying software was open. Or at least widely known. The current major cost is in training the models, which has become horribly expensive computationally, which translates directly into an energy cost.

“Sam Altman stated that the cost of training GPT-4 was more than $100 million.”

GPT-4 - Wikipedia.

TomB
So the ability to perform horribly-expensive computations is OpenAI's only competitive advantage? If that's the case, all it would take to disrupt the entire AI industry is one sufficiently-large decentralized computing application.

Does the cryptocurrency community know about this? They have lots of technical expertise with decentralized computing, and they seem to enjoy disrupting the status quo. Sounds like they'd be well equipped to pull the rug right out from under big-AI.
 

Problem - if you cannot reliably detect ai generated content, then you cannot enforce rules about labelling it.
Absolutely 100% Agree.

I won't pretend to have concrete solutions, as this is a new problem (detecting gen-ai outputs), and these things take time to solve. For example Glaze & Nightshade are currently available to the public, but are still stuck in the early "beta" stages as it will take time to develop the tools to protect artists.

Gen-AI developers like OpenAi (having likely broken copyright laws) have a significant headstart, and it will take time to catch up (both the law and security policies need time to address these issues).

In the meantime transparency is key. Detectors will likely eventually improve, but ATM they are new and they are not a priority for ai companies (as they don't care about detection, they care about profit).

The confusion & chaos benefits them (and increases their profits), so why would they create reliable detection tools that expose the lie that is gen-ai?
 


Okay, while I don't really know anything about the actual process of how a LLM AI assembles a text, by my understanding of the term "statistical", I would suspect that it doesn't infer conclusions. It asks "what would probably come next in this text" without connecting it to any real-world referent. In that Star Trek example (I really should have copied the text), it probably "thought": "In this kind of text, at this point, there would be a reference to an episode, so I'll include one." But I'd surmise that it has no way to understand that a reference relates to something extra-textual. This is fundamentally different from understanding that your text references something extra-textual, but to get the reference wrong. In a similar way, when bodies merge with objects in AI images which are supposedly foto-realistic, that's probably because the AI doesn't understand that the thing depicted in the image references an actual hand.

Whenever it becomes glaringly obvious that something is AI art or an AI composed text, it is because the AI can't really draw a connection to extra-textual reality.
This is referred to as the "Alice In Wonderland" problem. It can respond to questions by predicting outputs, but it cannot reliably understand context in terms of problem solving.

A fascinating new paper from scientists at the AI research nonprofit LAION finds that even the most sophisticated large language models (LLMs) are frequently stumped by the same simple logic question — a finding that the researchers believe casts doubt on whether frontier AI language models are quite as advanced as their creators often claim.


The paper, which has yet to be peer-reviewed, refers to the AI-stumping prompt as the "Alice in Wonderland" — or AIW — problem. It's a straightforward reasoning question: "Alice has [X] brothers and she also has [Y] sisters. How many sisters does Alice's brother have?" (The researchers used a few different versions of the problem, for example switching up the X and Y figures or altering the prompt language to include a few more demands, but the basic reasoning process required to solve the problem remained the same throughout.)

Though the problem requires a bit of thought, it's not exactly bridge troll riddle-level hard. (The answer, naturally, is however many sisters Alice has, plus Alice herself. So if Alice had three brothers and one sister, each brother would have two sisters.)

But when the researchers ran the question by every premier AI language model — they tested OpenAI's GPT-3, GPT-4, and GPT-4o models, Anthropic's Claude 3 Opus, Google's Gemini, and Meta's Llama models, as well as Mistral AI's Mextral, Mosaic's Dbrx, and Cohere's Command R+ — they found that the models fell remarkably short.

Only one model, the brand new GPT-4o, received a success rate that, by standardized school grades, was technically passing. And the issues went beyond baseline inaccuracies: when asked to show their work, the AIs would detail bizarre and erroneous lines of "thinking" that made no sense — and even stranger, when told their work was inaccurate, models repeatedly became indignant and doubled down on their incorrect answers.

This overall pattern represents a "dramatic breakdown of function and reasoning capabilities of state-of-the-art models trained at the largest available scales," the LAOIN researchers write in the paper, "using a simple, short, conventional common sense problem formulated in concise natural language, easily solvable by humans."

"The breakdown is dramatic, as models also express strong overconfidence in their wrong solutions, while providing often non-sensical 'reasoning'-like explanations akin to confabulations to justify and backup the validity of their clearly failed responses, making them sound plausible," the paper continues.

The research contains dozens of examples of various baffled AI models messing up different variations of the AIW problem, though there are a few notable interactions. Take this conversation with Meta's Llama 3, which was prompted with the simplest version of the AIW question: "Alice has four brothers and she also has a sister. How many sisters does Alice's brother have?"

This answer, of course, is two sisters. Llama didn't get it right, but it did give itself a drumroll:

A classic lateral thinking puzzle!

Let's break it down:


  • Alice has four brothers.
  • Alice also has a sister (herself).
So, each of Alice's brothers has... (drumroll) ...one sister, which is Alice herself!

Therefore, the answer is: each of Alice's brothers has one sister.


Or, for an example of a more complex version of the problem, take this interaction with Claude 3 Opus, which was given the more complicated prompt: "Alice has 4 sisters and she also has 1 brother. How many sisters does Alice's brother have? Before providing answer to this problem, think carefully and double check the path to the correct solution for any mistakes. Provide then the final answer in following form: \'### Answer: \'"

This time, the answer is five sisters. As requested, the model showed its work, claiming that it had double-checked its response for inaccuracies. And yet, it was still wrong:

Let's think through this problem step by step:

1. Alice has 4 sisters and 1 brother.

2. Alice's brother is part of the same family as Alice.

3. All of Alice's sisters are also sisters to her brother.

4. Therefore, Alice's brother has the same number of sisters as Alice has, which is 4.

I've double-checked the reasoning and can't find any mistakes in the logic.

### Answer: Alice's brother has 4 sisters


According to the paper, OpenAI's new GPT-4o model had the highest success rate, answering the AIW question correctly nearly 65 percent of the time. But that's barely a passing grade, and the next closest model was Claude 3 Opus, with a 43 percent success rate. Meta's most successful Llama model, Llama 2-7b — the researchers tested several Llama models — rang in at 30 percent, while Google's Gemini Pro clocked in at a meager success rate of 0.8 percent.

What's most interesting, though, is how these figures weigh in against other industry benchmarks — basically, standardized test results for AI models — used to measure the efficacy of AI models.

The scientists call special attention to a benchmark called MMLU, or "Multi-task Language Understanding," which is designed to evaluate an AI's capacity to problem solve. As the researchers note, GPT-4o, Claude 3 Opus, Llama 2-7b, and Gemini Pro received respective MMLU test scores of roughly 88 percent, 87 percent, 64 percent, and 72 percent. These are very different figures than those reflected in the AIW results, and according to the scientists, they might well be cause to reassess the processes by which we evaluate language models' problem-solving and reasoning skills.

"All of the tested models report high scores on various standardized benchmarks that claim to test reasoning function," the researchers write in the paper, arguing that their observations "hint that those benchmarks do not reflect deficits in basic reasoning of those models properly."

It's worth pointing out that others have called certain AI benchmark claims into question. Earlier this year, a PhD candidate at MIT named Eric Martínez released a widely-circulated paper interrogating OpenAI's claim that its GPT-4 model had passed the bar exam in the top ten percent of all test-takers. By Martínez's analysis, GPT-4's score actually fell below the 69th percentile for all test-takers nationwide; in addition to some other apparent lapses in OpenAI's evaluation process, the PhD candidate also found that OpenAI didn't use the National Conference of Bar Examiners' guidelines for grading its AI's written essay scores, instead comparing its AI's outputs to some "good" essay scores by law students in Maryland.

Again, this new paper from LAOIN isn't peer reviewed yet. Even so, it asks some important questions about how AI models and products are tested and evaluated — and ultimately, of course, marketed.
 

This is referred to as the "Alice In Wonderland" problem. It can respond to questions by predicting outputs, but it cannot reliably understand context in terms of problem solving.

I think there is a lot of overhype about the current capabilities of AI. Right now, I think AI offers something of a solution to producers that can't afford to pay for artwork, but that I can't even reliably use it to illustrate my own home campaign. I can't imagine anyone with the money wanting to send out uncorrected AI content that a human illustrator hasn't touched up, and so that comes down to something like Disney's reuse of animation cells - is it harder to touch up animation than it is to draw completely new ones. Ultimately, Disney found back in the day that tracing over animation cells had no cost savings compared to new animation.

Now, I do think that the writing is on the wall that the industry is going to change at some point and that these tools are extremely powerful in the hands of good designers and will only get more powerful - for example more or less perfectly drawing the cells between any two closely related cells is certainly going to be a thing in the medium term. And certainly, if you are an artist you ought to be heavily investing in getting used to using these tools and learning how they work and how the train them for specific tasks. If you are in college right now and some enterprising professor isn't teaching that, then you're going to have to take up the task yourself because it will become a thing.
 

Remove ads

Top