Silicon Valley's New AI Motto: Train Now, Pay (Maybe) Later

My thoughts on Suno/Udio lawsuits (and really any Gen AI model trained on ‘publicly available data’)

Let's start by acknowledging that publicly available data is a euphemism for scraping pretty much anything you can access via an internet browser – including YouTube, podcasts, and clearly commercial music libraries, and as a consequence pissing off record labels and celebrity artists.

The crux of the issue? What is ethical is not always legal, and what is legal is not always ethical.

Silicon Valley's Modus Operandi

"Our technology is transformative; it is designed to generate completely new outputs, not to memorize and regurgitate pre-existing content. That is why we don’t allow user prompts that reference specific artists," said Suno CEO Mikey Shulman in a statement.

In other words, this practice will absolutely continue until courts rule that training generative models on copyright material is not considered fair use, that the output is not transformative, and/or restrict it through a new regulatory framework.

Otherwise the Silicon Valley dictum applies: don’t bother asking for permission, ask for forgiveness later. Keep making the case that your usage is transformative, and then as your startups raises more money, retroactively strike licensing deals with the parties whose content you trained on, much like OpenAI.

Of course, this doesn’t always work out as expected.

A single tweet by ScarJo unleashed a visceral shit storm upon OpenAI, ultimately resulting in the eerily similar voice of Sky being removed from ChatGPT. And this is despite OpenAI maintaining the voice wasn’t based on Scarlett, and was instead a voice actor, who for some strange reason is choosing to forgo the free publicity and remain anonymous. More than a little bit sus.

And of course, the New York Times is pushing on what could be a seminal lawsuit against OpenAI - making the case that OpenAI allegedly used its copyrighted articles without authorization to train AI models, which they claim diverts web traffic and reduces their subscription and advertising revenue.

Mixed Media Response

Now, NYT has the war chest to take on this legal battle. But most other media companies see the writing on the wall — and are taking their pennies on the dollar now, with some vague hope of harnessing AI to further feed the beast that is the modern attention economy.

It is thus unsurprising to see OpenAI strike licensing deals with the likes of NewsCorp, The Atlantic, Vox Media and ShutterStock. Similarly, Google and OpenAI are rumored to be forging deals with Hollywood studios. These deals would allow them to train their Veo and Sora video generation models on the studios' libraries of audio-visual umami.

The Music Industry's Dilemma

Suno and Udio on the music end will likely follow these bigger players. They'll probably strike licensing deals with record labels in the future. But the question is, does the music industry today feel threatened enough? Do they believe that generative AI music represents a clear and present danger to their roster of human artists?

In the short-form video era, spurred by the rise of TikTok, Instagram Reels and YouTube Shorts, the music industry struck a different kind of deal with these tech giants. They provided up to 60 second excerpts of their music, licensed with favorable terms. This made sense because these videos going viral with the song attached drove streams and sales for their music catalog.

Such is not the case for generative AI, where attribution is largely a feature relegated to search engine products like Perplexity, and pretty much nonexistent everywhere else – especially in audio/visual generation models, whether it’s Suno, Udio or DALL-E 3.

The Current Workaround

So in practice, startups and tech giants alike enforce a policy on the prompting and inference layer. Essentially, they continue to train models on copyrighted material, but they prevent users from putting in names of artist or copyrighted material in the final prompt.

So for instance, I can’t ask for a photo of Godzilla from DALL-E 3, but I can ask for a Kaiju monster that looks like a dinosaur, and I’d pretty much get the same thing.

If you want to be more sophisticated, you might run some classifier on the final output itself. You might say it looks too much like Mickey Mouse or <insert Disney IP>, and you might not share that generation with the user.

This opaquely forces the user to re-prompt (without any useful feedback) until they no longer trigger that classifier, and get some generations close to what they intended.

But some might argue this is a rather hypocritical stance by tech companies. After all, their neural nets are indeed trained by this copyrighted material. It's only crudely being prevented from generating such output.

But at that point of scrutiny, they've usually raised enough capital to license key parts of the training corpus needed to quell any legal action. Combined with public domain content, this is sufficient to unblock their product. That seems to be the strategy playing out before us, anyway.

The Real Losers

In all of this — it’s really the indie artist who is royally screwed. The creators who aren’t big enough to drive a collective bargaining effort with tech companies and get paid off for their content libraries.

I’m thinking of the millions of creatives who tirelessly uploaded content on YouTube, Instagram, Artstation, Soundcloud, Spotify, Substack, Medium, X, and beyond — whose work has been slurped into these gigantic training runs and infusing these neural networks with abilities we on social media deem to be 'magical!' All the while forgetting the faceless army who created the knowledge foundation — this proverbial library of Alexandria scrubbed of all author names.

So where does this leave us? We can be very sure that as long as courts don’t intervene - the practice of training on ‘publicly available data’ will continue. Robots.txt be damned. Copyright be damned. Proponents of this practice will say that we all learn from each other, and that “everything is a remix.”

But machines are different from humans — a human can’t watch all of YouTube, and can’t listen to every song on Spotify, read all the text on the internet — but a machine can. And that makes it feel fundamentally different from a human learning from copyrighted material to produce transformative work.

Moving Beyond Human Limitations

Proponents will ask why machines should inherit the same limitations as humans? Why should they lobotomize the greatest record of human knowledge and creativity for outdated copyright law? And these are fine questions to ask, and I certainly won’t pretend to have all the answers.

But there’s one thing I feel strong about — no matter who produced the content, whether it’s a label-repped artist, YouTube creator, or a Substack writer — you’re at least entitled to an acknowledgement for your contribution to the greatest record of human knowledge and creativity.

So yes, perhaps it’s true that we shouldn’t infuse the limitations of humans into machines. It is also true that humans are incapable of outlining all of their influences — we assimilate them deeply, often unconsciously. But that's our flaw. Why should we settle for machines that possess the same shortcoming?

If you enjoyed this post, consider sharing it. You might also enjoy my LinkedIn feed on all things creative, tech and society: Bilawal Sidhu

-Bilawal Sidhu