The Disruptive Impact of Seedance 2.0 on the Film Entertainment Industry

Seedance 2.0 has completely turned the film entertainment industry upside down. YOU CAN SMELL THE FEAR: - They're trying to block the meme-jokes we're generating on social media (like the 15-second Lord of the Rings one, which, at the end of the day, is parody, uses zero real footage from the movie, and would therefore fall under "fair use"). - Disney has sent a "cease and desist" to Seedance 2.0 (yes, to the Chinese! Hahahaha, I can't help but laugh). They're gonna wipe their ass with it. - The actors' union in the USA has published the cry-baby statement below against Seedance 2.0. And all of this because the jump in quality... IS TOO MUCH. The model is TOO good 😂 Like, scary good. Even the most hardcore, lifelong anti-AI accounts aren't saying a word these days. Because there's just nothing left to say: yes, of course, it still has small imperfections, the dialogue is still kind of a joke, etc... but... If we're not "there" right now, even the most hardcore deniers can see we clearly will be in less than 1 or 2 years. All these tantrums seem to be missing something important: One thing they don't seem to realize is that, right now, in pretty much any jurisdiction, it's perfectly legal to train models on any kind of data, whether you own the copyright to that data or not. So Disney or SAG-AFTRA can try whatever they want, and not only are they gonna get nowhere because this is China, but even in their own jurisdictions, what xAI, Google, OpenAI, and everyone else is doing indiscriminately is: it's COMPLETELY LEGAL to train on any data. That said. OBVIOUSLY, if a user then tries to publish generations that use someone else's IP (like Superman), that's a different story, obviously: the IP itself is protected. Same as if you'd drawn it by hand. So yeah, well... It's a new world. But careful, nobody should be surprised by this... except the deniers, of course. Denial is no longer an option. Put differently, SAG-AFTRA is judging ByteDance (Seedance 2.0) with a double standard: American models like Sora also trained on whatever data they felt like. If anything, a few days after launch, they slightly tightened the rules about not using celebrities, etc. (but that's basically makeup, at the level of detecting it in the prompt and stopping you: the underlying model is fully capable of doing it). This is what ByteDance will end up doing too, if they feel like it, because there's also no law forcing them to. And by the way, when there are open-source models at this level and you can run them at home... all restrictions of any kind are over. The toothpaste is out of the tube, and there's no putting it back in. That said, I'll repeat it. If a user generates Brad Pitt, they could get sued over rights if they publish it anywhere, unless they can defend it as parody and "fair use". Same with IPs like Batman, etc. So where should protection of IP and likeness rights come from? Where it has always come from: from the channels/platforms where you publish it, if it doesn't fall under "fair use". If you generate your movie and Superman shows up, you're screwed, because it's protected IP and they can take it down from YouTube or wherever you publish it at any time. But again: how are they supposed to fence this in if you generate it for yourself, at home, with your open-source model and your GPUs, in a near future where everything is more efficient and less expensive? If you have your own "personal Netflix", we're entering completely uncharted territory that they can't realistically stop 🤯 The alternative is not going to happen. The alternative would be to legally limit what data models are allowed to train on, which in practice is: - First, almost technically impossible: if, for example, you wanted to limit a model's training data around "Superman", it would be almost impossible or extremely expensive because Superman images and videos are everywhere, not just in official DC Comics material, but also in fan art, user posts, and so on. So, in practice, they'd have to use other models (trained on everything) to scan the dataset before training the final model to detect whether certain things are present or not (which turns into a logical absurdity, a snake eating its own tail). What they *can* do is, for smaller artists whose work is concentrated in a few places, offer an "opt out'" which many jurisdictions already contemplate, and remove that URL/data from training... but all it takes is one of their images or videos being copied somewhere else and, in practice, it becomes impossible or very costly to reliably separate. - Second: no nation/country/power bloc (except Europe, which is off in la-la land trying to figure out whether to put caps on plastic bottles or not) is going to want to lose THE AI RACE. Losing this race is LOSING EVERYTHING. Because the first nations to pull it off (and it's clearly going to be China and the USA) will dominate the world. So they're not going to pass legislation that kneecaps them, because they don't want competitor nations to beat them. So they'll keep legally allowing their companies to train on any data they want. So. Interesting time frame you are living in, eh?