Recently there's been some drama about copycat LoRAs and CivitAIs refusal to do anything about them based on mere output similarity. This made me realize something huge: if you are a traditional artist and you publish a LoRA made from your own works, a copycat LoRA can be legally created, and you might not have any rights over that whatsoever.
This might seem obvious to many of you, but it was a shocking realization to me. After all, copyright exists to grant authors the exclusive right to profit from their work, and it also grants moral rights such as recognition and integrity of the work. Outputs of the AI models are typically considered free of such restraints, and therefore new LoRAs trained from outputs of the original LoRA need not follow ANY terms, conditions, or wishes the original author may have had about the use of their LoRA. This means LoRAs, while technically protected by copyright in their file form, might not in practice enjoy the same protections in a meaningful way.
Which brings me to an interesting point... You know, CivitAI will take down LoRAs that have been trained on copyrighted work without permission. However, they will first ask if the artist would instead like an official LoRA to be published with a 70% revenue cut. See where this is going?
CivitAI doesn't bother to mention that once you publish a LoRA, anyone can just use outputs from that LoRA to train another one, and CivitAI won't care. Since outputs from a legally published LoRA are legally used, the original artist might not have any economic or moral rights over the copycat models, no say on how they're used, and no right to ever take them down. There's no backing out from this deal, you'll lose control of your style for good.
Obviously this has never been tested in the courts, and a judge might rule otherwise. But before a legal precedent exists, it seems awfully risky to publish LoRAs of your own art style, or ANY LoRAs you care about at all. At the very least, you shouldn't trust anyone that promises a revenue share from the use of your style. It literally feels like a scam.
The realization that artists currently cannot safely publish original LoRAs makes me believe there's a fairly good chance that AI model outputs will eventually be judged to be derivative works of the models used and thus also their training data, at least in some circumstances. If CivitAI hasn't gone bankrupt by then, they will finally have to address the issue of copycat LoRAs as well, but I guess they're too busy trying to turn a profit to care about the complicated matters of creator rights for the time being.