There are a couple of places that I still use or have used or plan to use things that are commonly rolled up in the "AI" bucket, and that I believe are fine.

Because "AI" is a bunch of different technologies built a bunch of different ways, wearing a marketing trenchcoat.

I use an automatic subtitle generation tool that is "AI" powered, but runs locally and isn't generating anything new, it's trying to identify a thing that already exists.

The argument could be made that it was trained on stolen labor in that it was trained by matching existing subtitles to the audio tracks those subtitles correspond to. I decided I was personally comfortable with this, but I understand how other people might not be.

It's wrong a lot, but it gives me a place to start from as I'm manually subtitling a thing, and saves me hours of transcription work.

I think it's fine. I think it's good even, I like it and I want it to get better.

I also occasionally use image classifiers to sort through large volumes of photos or scans or whatever because going through them by hand would take a lifetime, and some large corporations already taught computers to do this. This makes some of my research and art projects much faster, and can again be run locally and generally doesn't involve any stolen labor (at least no more than we've been giving away to solve captchas)

I'm planning to set up a particular image classifier that is designed to specifically extract foreground images from solid color backgrounds, to provide a clean chromakey for footage that was not filmed against a properly lit green screen.

Again, it's not generating anything new, it's automating manual labor away, and letting me focus on the creative part of video production, rather than rotoscoping a person out of their background over and over again.

I think this is great!

I don't think it's really fair to call either of these things "AI" in the way that we talk about "AI" right now, but they are both built on machine learning, and ...

I have a point, give me a second.

Often, when I find myself getting roped into a conversation about generative AI, and the ways that it is bad, people will either point at these technologies to call me a hypocrite or they'll use these technologies (and a couple of similar ones) to justify that "AI" can be good and ethical and so, because "AI" can be good and ethical in general, their specific use case must also be as well.

This is marketing fluff.

So yes, there are things called AI that I think are fine, because they help me do a mechanical thing faster (transcription, rotoscoping) and can be run on my local machine, and mostly aren't powered by stolen creative labor.

But these things are fundamentally different from what (in my experience, most) people mean when they say AI.

What (in my experience, most) people are railing against is Automated Slop Generation trained on the output of real people.

Image generators, music generators, text generators.

"Tools" designed to mimic human creativity and produce something that is designed to trick people.

Generative AI is a machine for producing something that seems plausible for a given input. That's all it can do, and that's what makes it dangerous.

Well, that's one of the things that make it dangerous anyway.

I bring all that up, because it's an argument I see over and over again, it's one of the places (perhaps the only place?) for nuance in this conversation and I've seen that point abused and dismissed on both sides of this conversation.

Some folks who I otherwise agree with seem to hold that anything that gets bundled under the label AI or that involves ML is inherently evil and wasteful (I disagree) and other folks who I generally disagree with seem to hold that, because there are things labeled AI or ML that aren't evil and wasteful, there thing must therefore be okay (I disagree!)

In this particular case, we were in nuance territory. The person in question was writing music and performing vocals and then generating parts and exporting stems from a generative AI music imitator to produce something they considered a prototype or sketch of the finished track.

That, on the surface, doesn't sound all that different from me firing up garage band and sticking an autodrummer behind a track I wrote, or using a vocoder for vocals, or playing to a drum machine.

and I imagine that some folks, myself included when I first encountered this argument, might struggle to articulate what makes it different.

Resource utilization, certainly. Trained on stollen labor, for sure.

Those are problems, but they aren't the biggest problem I have.

The biggest problem that I have is that Suno and tools like it are trying to replace the act of making music. They are only being positioned as music creation tools for individuals for as long as it takes them to get good enough that most people can't tell the difference, and the people using these tools are paying for the privilege to train that machine.

When the training is done, the people get cut out, and the market get flooded with slop that's good enough that most people will struggle to recognize it.

(This is already happening. It will get worse.)

Garage band can't spit out a song without human interaction. Suno can.

It's the difference between human-machine augmentation and machine automation.

It's probably possible to use a generative audio application in a human augmentation way, but the tools that exist right now are clearly intended to automate the people out of the creative process and, as far as I can tell, any use of these applications serves to further that goal.

So even if you can over look the stollen creative labor (and you shouldn't) or the runaway resource utilization (and you shouldn't) and the obvious economic fraud (and you shouldn't) in the name of agency or accessibility, I think it's important to reckon with the fact that access to these applications today is predicated on their eventual use in place of human produced music tomorrow.

@ajroach42 I recently finished using an AI image classifier to annotate 15,700 photos in my photo database. There’s no way I’d have even attempted it without the AI.

Similarly, I have an application that will let me dump in documents, ask questions, and get responses *with citations to exactly where the answer was found* so I can click through. Basically, a fancy personal search engine. I wish the LLM hadn’t been trained on stolen data, but right now that’s what’s available, and it’s not like I’m using people’s stolen data itself — just the latent semantics of the language that were gleaned by examining the data.

The person saying they can’t afford to make music without AI is full of it, though. We’re in a world where they can download a digital multitrack studio and some software instruments for free. (If they have a Mac there’s one preinstalled even.)