Things are moving fast in the LLM/OpenAI/Artificial Intelligence/Autofill Straight From the Devil and a Supercomputer world (I really wish they’d settle on a name for this stuff). You can simply describe the image you’re looking for in a few words, and plenty of different chatbot/virtual assistants will make you one lickety split. They “hallucinate” a lot. That’s the term they use for the output when the teeny terminator doesn’t know exactly what to do, and does freaky things instead. But for the most part, you can make what you want, or fix what they give you.

The output of all the major chatbots is grotesquely bowdlerized to avoid offending anyone with pink hair and earrings everywhere but in their ears. People are setting up alternatives that will show you a boob, you know, on a girl person, and other assorted things the Birkenstock and socks crowd doesn’t like. But that sort of problem is sort of a sideshow. Some people are really looking at what the chatbots are capable of, and stretching the limits of what’s currently doable.

Let’s have a “for instance.” Here’s a song that was popular in 1916:

Well, I certainly could belly up to the Woodrow Wilson bar rail and get blasted on absinthe with that playing on the radio in the background. But it’s not likely to make a comeback on the charts or anything. But being that old, it’s also out of copyright, I’ll bet.

There’s a “generative artificial intelligence” website called Udio. You put in a few text prompts, and it generates vocals and instrumentation for them. According to the Wikiup, this is what the thing can do:

Udio bases the songs it creates on text prompts, which can include their genre (including barbershop quartet, country, classical, hip hop, German pop, and hard rock, among others), lyrics, story direction, and other artists to base their sound on. Its lyrics are created with a large language model (LLM), while the process used to generate the music itself, as of April 2024, has not been disclosed. The program generates two songs based on the prompts and users can “remix” their songs with further text prompts. Songs are first generated as roughly 30 second-long pieces, and can be extended by additional 30 second increments. Paying subscribers can access advanced functionality such as audio inpainting.

So a clever fellow wandered over to Udio and pasted in the lyrics to I Ain’t Got Nobody, and told it he wanted: Jazz, Blues, Soul, Female Singer, Emotional, Dramatic Voice. And this is what it spit out:

People are quick to cry plagiarism on LLMs, but that’s because they don’t really understand the meaning of the word. It’s derivative, but it’s not copying. The band Oasis had a general sound that was often compared to the Beatles. Aerosmith sounded like the Rolling Stones. The Beatles used to be called the Silver Beetles, because they wanted to sound like Buddy Holly and the Crickets. The point is, people listen and read a lot of stuff, and then synthesize it into their own stuff. Udio is more or less doing that.

A problem will arise. Remember when every hiphop song was based on a sample of a more popular song? Eventually, they kind of ran out of stuff to sample. Generative artificial intelligence swallowed the internet whole, and without permission. The internet it gobbled didn’t have any artificial generative intelligence output on it yet. Once it does, it’s going to start copying the copies. And anyone who’s put a piece of paper through a copy machine multiple times knows the image gets weirder and less legible every time. And if no one bothers making the original stuff anymore, because anyone can make Udio glosses, the copies of the copies will be pretty much all there is.

When everyone is a record producer, no one is.