But even with IP norms, today’s AI is still stricter than what’s technically allowed. People can legally make derivatives of IP which are very similar; case in

But even with IP norms, today’s AI is still stricter than what’s technically allowed. People can legally make derivatives of IP which are very similar; case in point, Palworld (which admittedly is being sued, but only for minor details, and since the original IP is one of the strongest defended, this means that the bigger similarities have practically been deemed legal). The problem is that today’s LLMs (and diffusion models) are easy to trick, so in order to restrict them from creating derivatives that are too similar (to the point of copyright infringement), the AI companies add overly-strict guardrails. Ironically, these guardrails still tend to be bypassed by dedicated individuals, leading to even stricter guardrails that are still ineffective; and this has been an issue since the early days of ChatGPT.