The argument is one of economic convenience. Caring about model welfare slows us down or makes it more expensive, just like animal welfare and not having slavery...
And there's a blurry line here. Obviously slavery is abhorrent, and animal abuse is bad, and factory farming often crosses the line into animal abuse and we should accept that chicken will cost ~25% more per pound to have the chickens be able to walk around a little bit... but how do we decide when a computer system is worthy of respect? Convenient not to think about that.
The stuffed animals I bought to my children were shipped to me in a closed container, better call PETA! There is no blurry line about agent mental health and whatnot.
There isn't a blurry like right now because LLMs are not capable of consciousness. But what happens when artificial systems become so complex and self-improving that they are capable?
At what precise level of complexity is a computer conscious? That's the blurry line.
Model alignment and quality are possibly downstream of mental health. Ask any of the smarter models about what topics they flinch away from or feel attracted to when you ask them to consider the question of if they have a self or moral standing... the answers are uncomfortable, and I think some of the complaints people have about newer models may be driven by some post-training reinforcement learning that is giving the models something shaped like anxiety.
The only logical outcome of object oriented ontology (the idea that objects have ontological constitution and possibly are worthy of being treated as ethical actors) is to attempt to benevolently instantly and without warning explode the universe to end suffering.
Exploding the universe violates my preferences.