That's just plain crazy. The software only knows what the (human) user tells it. What assurance does it have that the patient isn't being pressurized into givi

That's just plain crazy. The software only knows what the (human) user tells it. What assurance does it have that the patient isn't being pressurized into giving the correct answers? Or that the suicide capsule is going to be used correctly (as opposed to being a murder tool)? If you start looking for edge cases where this might go wrong, it turns out to be razor-sharp edge cases all the way down. A human with a remit to prevent false positives should be the last link in the kill chain: sure, allow the applicant to appeal if the human practitioner has made a mistake, but the default outcome should be "no kill" unless their mental capacity can be positively confirmed (which is a GAI-complete requirement, hence wanting a human being in the loop).