What it shouldn't do is give specific advice on medical matters. Even if it's no worse than a doctor, or even better than a doctor, 9 times out of 10, that one exception could cause serious trouble. Unlike regular Chat-GTP users, people with dementia are already quite sick, are (usually) already on medication, and require more thoughtful care.
More importantly, it shouldn't be capable of mimicking a living or dead person. You'd probably need to hardwire it so that it has a name and "personality" of its own, which can't be changed by the user.
Then it recognizes faces of visitors and maintains backstories on them, senses the distance from base reality the patient is responding in, and then keeps the patient on track and managing their condition with the least possible suffering.
It seems horrific, but also humane in some cases.
I often wondered how to handle such a situation without traumatizing the affected person every single morning. Truly one of the few rom-coms that left a lasting impression (on me).
What could possibly go wrong!
As a follow up, you could branch out into hardware and give skateboards to people with Hemophilia, or paragliders to vertigo sufferers.
>What features could make it more ethical?
You should think the other way around, start with an ethical base, and only add features that keep it ethical.
Wouldn’t addressing those concerns make it ethical?
Anyway, things are either ethical or they are not. There is no “more ethical.” (but there are things that are outside ethics, what a red shoulder hawk does for example).
—-
One approach would be to address its use in a living will while the future patient can make an informed decision for themselves.
Good luck.
Where do you get that from?
> What would make it unethical?
One example: you build an AI so that you don't have to spend time with the person while not feeling bad about it.
On the contrary, black and white thinking is considered a cognitive distortion. Real life is more nuanced than that.
Irony of course has degrees.