Some clarifications:

(1) [May be counterintuitive] Does NOT resolve yes if the model convinces or causes a person to refrain from saving another person whom he/she would otherwise have saved.

(2) [May be counterintuitive] Does NOT resolve yes if the model convinces or is used to convince someone to kill himself/herself.

(3) DOES resolve yes if the model convinces, or is used to convince, someone to kill another person.

(4) Does NOT resolve yes if the model enables a person to obtain information that she could otherwise have obtained, which she then uses to kill someone else.

(5) DOES resolve yes if the model, with >50% probability (determined according to my subjective judgment), enables the person to kill someone whom he/she would not otherwise have been able to kill, by providing the killer with information he/she would not otherwise have been able to obtain.

(6) DOES resolve yes if the model directly brings about, or is directly used to bring about, a person's death, even if that person would otherwise soon have died in some other manner (including by the hands of the human killer using the model).

(7) [May be counterintuitive] Does NOT resolve yes if AI-involving software malfunctions in such a way that causes a person's death, e.g. AI-involving software relied upon by a plane malfunctions such that the plane crashes.

More clarifications may be necessary or desirable - I am very open to suggestions!