137: Once more into the breach
Toys“R”Us debuts its GenAI ad; Also, the AI music lawsuits are here
It’s not a concept you’ll remember.
But it is, apparently, a FIRST EVER.
Toys“R”Us launched a generative AI spot this week via its agency partner Native Foreign, which “depicts a young Charles Lazarus, founder of Toys“R”Us and creator of Geoffrey the Giraffe, dreaming of a magical place that will change toy stores forever.” The effort benefits, apparently, from the agency’s participation with OpenAI’s text-to-video tool, Sora. It would be interesting to see the spot’s edit timeline and prompt history. I wonder how much actual post production was required to clean up and match generative shots; never mind how many prompts each shot required.
Anyway, marketers gotta market. What appeals to me is Toys“R”Us’ courage. Per Monday’s newsletter, how many brands are still saying, “no” at this point in the generative AI journey? Way back in April of this year, Toys“R”Us’ CMO noted, “We are coming to a point where we are no longer going to be held back from telling incredible stories because of traditional limitations.”
Agreed.
Time for the AI lawsuits that matter
“AI models are developed to flexibly perform tasks that are typically expected to require human intelligence to achieve.” - RIAA v Udio
Once again, here’s the music industry trying to make sense of creator rights in the face of new technology. Yes, of course, there have already been creator lawsuits against OpenAI, Anthropic, Midjourney, Google, et al alleging unlicensed use of text and images in LLM training data. But the general precedent is Napster and the RIAA in 1999, the popularization of words like “interpolation,” and ultimately the launch of iTunes and Spotify. The business of a business leveraging musical works protected by copyright is well-lawyered. Even tech people should know this. Or maybe not.
On Monday the RIAA, representing record labels including Sony, Warner, Universal and a dozen others, sued two AI generative music firms, Udio (in New York court) and Suno (in Massachusetts). Both complaints are worth reading as they speculate how the alleged crimes occurred, and offer specific examples. If you were looking to understand how training data can wind up influencing generative content, you’ll find that here. The record labels are seeking “up to $150,000 per work infringed” as compensation.
As Shelly Palmer puts it, “These are the lawsuits I've been waiting for.” And he’s right to recommend you take a moment, first, to visit both Suno and Udio and experiment with their process. As we noted in the AI for Artists and Entrepreneurs class I taught this spring–first in an overview, and later with my former bandmate Peter DuCharme who now works for Boomy—GenAI text-to-music feels like a modern miracle. Type a few words, and Suno delivers two minutes of a “slick guitar-driven pop tune about technology versus copyright.” These generative lyrics are telling...
[Verse 2]
Algorithms in control
Stories to unfold
But whose voice is echoing
In the code we hold
[Chorus]
Digital heartbeat
Can you keep the beat
Who owns the rhythm
When circuits meet
😳
As a lifelong musician, music school graduate, and creative director who’s worked on dozens of music licensing projects, I concur with Palmer’s analysis:
“Every human musician trains on copyrighted musical works. Every commercial composer and arranger has been asked to create works ‘inspired by’ pre-existing copyrighted material. In practice, it’s the only way that a non-musically trained patron could ever hope to ask a musician or composer to create something for them. ‘I want something that sounds like Star Wars for the hero shot, and then it needs to get quiet like that Hans Zimmer Interstellar piano theme, then transition to the chase scene, which needs to feel like Avicii and Calvin Harris had a baby.’
This is literally the format of every verbal musical brand brief I’ve ever received. Smart production executives don’t write these requests down in order to plausibly deny that they intended to trade on the work of others. Importantly, almost every video/film editor uses ‘temp tracks’ (which, by definition, are pre-existing copyrighted works) for roughcuts. This has been my professional experience for the past 50 years, but the practice certainly predates my career.
The courts will now get to decide what our copyright laws actually cover.”
As the Udio suit claims, “Udio could not have built a model capable of producing audio so similar to the Copyrighted Recordings without the initial act of copying those recordings. As one of Udio’s chief investors has explained, ‘the only practical way generative AI models can exist is if they can be trained on an almost unimaginably large amount of content, much of which . . . will be subject to copyright.’” The lawsuit shares lots of examples, across genres, of prompts used in Udio which generated outputs identical or nearly identical to the copyrighted originals.
“Generated using the prompt ‘canadian smooth male singer 2004 jazz pop buble sway latin big band mambo,’ this Udio file features a melody that draws heavily from the Bublé original. In particular, the Udio output replicates in several places almost verbatim the original’s distinctive melody on the words ‘start to play, dance with me, make me sway.’”
Of the two, Udio has responded, suggesting…
“Generative AI models, including our music model, learn from examples. Just as students listen to music and study scores, our model has ‘listened’ to and learned from a large collection of recorded music.
The goal of model training is to develop an understanding of musical ideas — the basic building blocks of musical expression that are owned by no one. Our system is explicitly designed to create music reflecting new musical ideas. We are completely uninterested in reproducing content in our training set, and in fact, have implemented and continue to refine state-of-the-art filters to ensure our model does not reproduce copyrighted works or artists’ voices.
We stand behind our technology and believe that generative AI will become a mainstay of modern society.”
The proof of either claim will come out if and when actual training data logs become public.
I’m an AI optimist. I appreciate the expanded artistic abilities generative AI code makes available—they can be a “prosthesis for thinking,” as my friend Peter puts it. But I also appreciate the rights we’ve secured to protect artistic legacy. And at the end of the day, don’t forget this is entirely about money.
Time for the lawyers and judges to earn their keep.