While I steer well clear of the never-ending stream of algorithmically selected nonsense that is YouTube Shorts, one of my main guilty pleasures is watching YouTube. I follow a very particular list of videographers and creators whose videos I appreciate. Since I don’t pay for the ad-free subscription service YouTube Premium, I get served plenty of ads. Recently, I have noticed a disheartening trend among those ads.
There has been plenty of debate already about the impact of generative AI on various professions, including creative ones such as music composition. So far, attempts at regulating AI companies have pretty much been a string of disappointments from the perspective of musicians and other creative professionals.
Data Use and Access
One recent such disappointment was back in June, when the UK government refused to amend the so-called Data Use and Access Bill with a provision that would have forced tech companies to declare their use of copyrighted material when training their large language models. The amendment was backed by well-known British artists including both the old guard, represented by luminaires like Paul McCartney, and the new, such as Dua Lipa.
The main critique of the amendment, a pitiful (in my opinion) argument that keeps coming up whenever regulating tech companies comes up, was that it would “stifle the AI industry”. In other words, the prevailing view appears to be that the industry needs to be protected even at the cost of the creators of the content they industry uses – with or without our consent – to refine the products with which they then could replace us.
Not only musicians have expressed outrage and anxiety about the possible effects of the Data Use and Access Bill. British writers such as Kazuo Ishiguro and playwright David Hare were among more than 400 signatories of an open letter earlier this year addressed to British prime minister Keir Starmer.
Another protest letter was released in October last year, signed by an international list of actors, writers and musicians. The year before that, in July 2023 thousands of writers endorsed an open letter from the Authors Guild to OpenAI, Meta, Microsoft and other AI companies. In the letter, they urged the companies to obtain permission before using copyrighted work in their technologies.
The Velvet Sundown
If you, dear reader, think I am being hyperbolic, let me remind you of the curious case of the so-called band The Velvet Sundown. It debuted on Spotify in June this year and quickly raised suspicion for several reasons. It emerged seemingly from nowhere with several full-length albums releasing within only a few weeks. Its four members did not seem to exist at all outside of the band’s own media activities. Its listener base on Spotify had grown suspiciously large in a short amount of time.
Late in June, The Velvet Sundown registered an Instagram account and posted images ostensibly of the band members that only further cemented AI suspicion about the so-called band.
Only a week later, The Velvet Sundown’s artist biography on Spotify was revised. After categorically denying that the music was artificially generated, the new biography ambiguously described the band as “a synthetic music project guided by human creative direction, and composed, voiced and visualized with the support of artificial intelligence”.
So many fancy words to describe AI slop.
No matter which generative AI applications were used to create the various materials making up The Velvet Sundown’s audiovisual identity, those applications were trained on existing, man-made musical and visual works, harvested and used without consent, credit or compensation.
To quote Nora Roberts, one of the authors endorsing the 2023 letter from the Authors Guild: “If creators aren’t compensated fairly, they can’t afford to create. Human beings create and write stories human beings read […] AI can’t create human stories without taking from human stories already written.” The same applies, of course, to music, art, photography and other expressions of human creativity.
Speaking to BBC on behalf of the non-profit Fairly Trained, which campaigns for AI firms to respect creators’ rights, Ed Newton-Rex described the band as “theft dressed up as competition. AI companies steal artists’ work to build their products, then flood the market with knock-offs, meaning less money goes to human musicians.”
Newton-Rex has worked with several generative AI projects and companies, most recently the creators of well-known AI image generator Stable Diffusion. However, in late 2023 he apparently left his position at Stability AI based on principal disagreements about the use of copyrighted material to train machine learning algorithms.
While I personally don’t suspect that some generative AI company has created The Velvet Sundown, someone is clearly profiting from its success. With the rapid development and growth of generative AI tools, I genuinely believe that they are a danger to us in many ways, not only to creative professionals because they threaten to assimilate and then supplant us, but also to society as a whole as the generated content gets increasingly indistinguishable from the real or manmade.
Fake Ads for Fake Products
This summer, I have noticed two obviously AI generated series of ads on YouTube, not only using AI generated voice-overs (something that seems increasingly commonplace, even if the visual content is not) but entirely AI generated.
The first was some random financial service for trading things like stocks or cryptocurrency. The ad itself featured a sitting middle-aged man speaking directly into the camera, lauding the features of this trading platform. The man never moved at all, his limbs or his head, never shifted in his seat or showed other signs of mobility. The face was animated, but in a distinctly uncanny way, like the voice that, while sounding superficially human, stressed words in odd ways and even mispronounced some words.
More recently, I have been positively inundated with ads for a fitness app purporting to offer tai chi-based training programmes to help you grow muscles and tone your body. Each ad begins with a 8-ish-second long clip with a topless, almost impossibly muscled, Asian-looking man saying something like “I am 65 and I don’t need advice from men in their 20s” or “this is how real men lose weight – with breath and control”, before cutting to an odd montage of exercise instruction videos set to vaguely Asian, dramatic but nondescript music and a deadpan Asian male voice-over extolling the fitness app’s features.
Those initial clips, with the ripped Asian-looking men, are entirely AI generated. They might have been made using Veo 3, Google’s newest video generation machine learning model. I encourage you to read the Ars Technica article I linked tot here and watch the example videos generated by the author, Benj Edwards. From a technical perspective, the videos are absolutely astounding. From a social perspective, they frighten me terribly.
Veo 3 is capable of generating 8-second clips with synchronised sound including dialogue. Generating such a video requires a subscription to Google AI Ultra and costs only about $1.50 per video generation. Even if you have to try 20 times, amounting to $30 total, that is still infinitely less expensive than the massive setup needed to actually plan, shoot and edit all of those short clips with real actors and sets.
Incidentally, this fitness app appears to be a scam, which is why I refuse to name it here as I don’t want to increase its online presence further.
Real Products with Fake Music
A number of other ads I have been served over the past couple of months have had no voice-overs, no sound effects, only a slideshow of pictures with a musical soundtrack that at least to my ears sounds completely AI generated.
Structurally, the music simply meanders. It has no direction, no sign of form or development. This is true of the music in all these ads I have noticed. Also, the sound quality has this uncanny quality to it that alerts my spider-sense. The audio sounds not only compressed, which is common in distributed music, but mashed together like a sheet of OSB.
Back in June, American music producer Rick Beato published a short video essay on YouTube where he tried to identify whether or not The Velvet Sundown’s music was AI generated. He used the Stem Splitter function in Apple’s audio workstation Logic Pro to split three different songs into its various parts: drums, bass, vocals, guitar, piano and other sounds. (This function is driven by machine learning, by the way.)
Beato used Logic Pro’s Stem Splitter on Good Times Bad Times by Led Zeppelin, Manchild by Sabrina Carpenter, and finally on Dust on the Wind by The Velvet Sundown. In the video, he showed that the Stem Splitter function had much more difficulty separating the parts in Dust on the Wind than in the two other songs. Its stems had much more distortion and artifacts than the ones from the Led Zeppelin and Carpenter songs.
Beato’s theory is that the AI audio track itself contains audio artifacts, that can be heard more clearly when the stems are isolated like in his experiment. The main reason, he argues, is that the music generators are primarily trained on compressed audio files in some lossy audio format such as MP3, not on high quality, uncompressed audio or even actual multitrack files where the different stems are still separated into individual tracks.
This squished-together sound quality that Rick Beato described much more eloquently in his video essay is what I perceive in the music to all those ads and why I suspect the music has been generated using one of the many music generation tools available already.
I have spent the better part of today trying to record examples of this music. I actually got two ads I had never seen before today, but could only record a short snippet of each, but I finally managed to record one in its entirety and isolated the audio to show you. This music is, for some reason, used for a Swedish online casino site (which I, like the fitness app, will not name).
Real Music, But No Longer Yours
Founded in 2009, the Swedish royalty-free music publisher Epidemic Sound is a massively valued company making millions of dollars annually and described by the Financial Times as “one of Europe’s fastest growing companies”. Back in 2021, its co-founder and CEO summarised the company’s mission as “trying to soundtrack the internet” in an article by TechCrunch.
Key to its success is the simplicity in its offering – for a flat monthly fee, creators can pick and choose from an effectively limitless selection of music sorted and categorised for efficient searching. The company naturally maintains a mountain of data on what kind of music as well as which particular tracks are popular.
Epidemic Sound not only offers its own app and service but has integrated itself into applications and services by Adobe and Canva, for example, making it practically ubiquitous for a large number of creatives and creative professionals.
The company has not escaped controversy, however. Swedish music rights organisations like SKAP and SAMI have been strongly critical of Epidemic Sound’s model, where it buys the exclusive rights to the music you create. In fact, this was always one of the main ideas to the business, according to the same TechCrunch article as linked above, creating a “restriction free experience for creators”.
While that may sound relatively benign, in practice it means selling your music for cheap and getting little in return. In the article, Epidemic Sound claims that “on average musicians can make tens of thousands of dollars [per] year” – with no evidence to support that claim. They also actively discourage artists and musicians being members of performance rights organisations, going so far as to deny you submitting music to them if you are a member of a PRO and are not willing to opt out of that agreement.
On their questions page, Epidemic Sound gives an answer that in my eyes reads like a gut-wrenching degree of Orwellian doublespeak:
Our ecosystem is based on our music being able to flow freely to all our customers. We have made this possible through our unique digital rights model, which is based on the premise of us being the sole owner of the financial rights to our music. […] We believe in a world where music can flow freely and fairly across time zones, borders, audiences and channels. We partner with artists and music creators who value distribution over ownership […]
Epidemic Sound / Frequently Asked Questions for Artists (cited 8/11/25)
A Perfect Storm
All of this comes to a head on page 10 in Epidemic Sound’s 2023 Annual Report, which you can download and check out for yourself from this press release. Under the headline Winning in the new paradigm – harnessing the power of AI, the company writes: “As owners of our music and associated rights, we can continue to avoid the complexities of legal copyright issues that many others will face if they use datasets they don’t own. And our extensive data insight on how music is used in online video, gives us a continuous and current loop to reinforce and tune our models.”
Immediately below, they include the following diagram on the bottom half of the page:
By owning the, as they call it, “dataset” (i.e. music) they can train a machine learning model to develop generative AI for music creation in combination with an almost unfathomable amount of metadata on what music is used, how and for what, and how often.
The first of their three “core areas” as listed under column 3 is already in use, as far as I have been able to tell from their website. The latter two I don’t think are in use – yet. But we are slowly but surely reaching a point where cutting-edge technology is converging on a spot where no musicians – no humans – will anymore be necessary to create something that was supposed to be an expression of human creativity, even if that creativity is sometimes, such as in this case, primarily used for commercial purposes.
The Velvet Sundown and its sorry ilk may be mere harbingers of the horrible future that we could be facing. Not only will the possibilities for creative professionals to earn a living be bled out, but the thoughtless (and, to be honest, thankless) masses will be blissfully unaware that actual professional, hard-working, educated, thoughtful, earnest artists have been effectively replaced by unthinking cyborgs. All in the name of ever-growing shareholder profits and, of course, a “restriction free experience”.
As written by screenwriters Brannon Braga and Ronald D. Moore for Star Trek: The First Contact from 1996, the Borg’s iconic warning gets a stark new meaning in this dystopian but sadly not at all science-fictitious context:
We will add your biological and technological distinctiveness to our own. Your culture will adapt to service us. Resistance is futile.
Borg in Star Trek: The First Contact