I can believe that – detecting images generated by their own image generator feels like a more tractable problem to me than detecting text from their LLMs Also model collapse for images feels more likely to me than for text, though I'm not sure I could explain that intuition

→ View original post on X — @simonw