A wave of visual art generators and chatbots have surfaced over the past year, powered by artificial intelligence. Artists have expressed concern that Midjourney and Dalle-2 are not only taking away jobs from humans, but training without limit on the work of human artists, without any credit or royalties.
That's where the Human Artistry Campaign comes in. On March 16th, the HAC made an announcement at South by Southwest Music Festival, explaining their mission to promote responsible and fair use of human art by generative AI tools.
SXSW is a popular forum where conversations like this can happen in the open. They dedicate a full section of their website to covering artificial intelligence. Greg Brockman, president of OpenAI, gave a lengthy speech about ChatGPT and Dall-e.
Right now, it is “very expensive” to run large language models like the one behind ChatGPT, she said. It costs 2 cents per query, or seven times more expensive than one basic Google search. But she believes prices will fall even as computing advances, as has been the trend in tech.
Most of the publicly available AI music generators have been very carefully designed to avoid copyright violations. So it's difficult to know whether they have any active targets in their crosshairs yet. In the next section I'll share my best guess of the apps that are on their watchlist.
AI music on the Human Artistry Campaign watchlist
The Human Artistry Campaign has been joined by the IFPI, RIAA, BPI, NMPA, ASCAP, SESAC, SoundExchange, Songwriters of North America, Black Music Action Coalition, Trichordist, and others. Find the full list on their website.
A variety of AI music app categories are emerging today. Some of them, like the AI mastering tools, may be already be on the Human Artistry Campaign's watchlist:
GPT MIDI and Audio generators: OpenAI published Musenet and Jukebox in 2019 and 2020 respectively. Both apps create music out of thin air, one focusing on MIDI and the other on audio. They train on human music and then formulate their own ideas with a neural network. Access and use has been limited, due to the technical challenge of using their API and the long processing times associated with generating a single song.
Text-to-music: Google's research team recently published a paper called MusicLM. The premise of the app was to begin turning descriptive text prompts into music. You can hear audio samples but the actual app is not publicly available. People are using AudioCipher as a text-to-music MIDI generator, but it doesn't feature artificial intelligence under the hood.
Commercial AI music generators: Mubert is a great example of an AI music app in this category. They openly market their tools as being a collaboration between human-made music loops and layers of artificial intelligence on top. Commercial AI music generators let you set parameters like key signature and tempo to modify the underlying loops, but they aren't fully generative like Musnet and Jukebox.
Text-to-speech: Uberduck is probably the most popular TTS software for musicians, because it's free and allows users to imitate famous people. Uberduck and similar text-to-speech tools been used for internet memes, mostly to create parody deepfakes of famous artists. Voicemod offers a variation on this with their "musical meme machine".
AI music videos: Sites like Kaiber and Video Killed the Radio Star make it possible to generate videos with AI based on text prompts. They've been featured in a recent music video from Linkin Park called Lost.
There are also AI tools available in digital audio workstations, for mixing and mastering audio. If you’re interested in getting started with learning the basics of machine learning, Dadabots run a free Discord channel where hobbyists discuss the latest tools like Jukebox, RAVEv2, and DDSP.
Companions to the Human Artistry Campaign
Like what you've seen from the Human Artistry Campaign? I'll share a quick overview of some other projects that might interest you. Spawning, HaveIBeenTrained, an "art cloaking" project called Glaze, and the Zora NFT marketplace will each be covered briefly.
Spawning and HaveIBeenTrained
Spawning is an organization that advocates for the fair and ethical use of artwork. They want to build a future where data can be shared consensually for the benefit of AI development and the artist data that it trains on.
Spawning are behind HaveIBeenTrained, a separate site that allows users to search for art and opt it out of existing training sets.
On March 7th, 2023, HaveIBeenTrained announced that 78 million artworks had been successfully opted out of AI training. In a perfect world, some say that these training sets should have been opt-in to begin with. This would have eliminated the burden of action for the artist. But at least an opt out exists.
Glaze: Digital art cloaking
Glaze is another tool that's been created in an alliance with artists, to make it difficult for AI to mimic. The tech was developed by professor Ben Zhao at the University of Chicago. He was interviewed by Tech Crunch in an article published yesterday.
Spawning recently published an article about Glaze, celebrating the app's efforts while exposing some of its shortcomings. This peer review process will lead to improvements in the software over time.
Zora NFT Marketplace
Zora claims that they make it easier for artists to collect royalties on their work. The platform wants to give people exposure and help them build a community. Artists can earn more money here than on conventional streaming platforms like Spotify or Apple music.
As a blog that covers AI music culture, we sometimes receive messages from readers and customers who feel concerned for our future as musicians. It's a relief to know that the Human Artistry Campaign is out there, along with these other groups, taking steps toward a reasonable future for artists.
Thank you SXSW for hosting these conversations and getting the ball rolling!