Table 3.

Analysis of harm scenarios (n=44 cases).

Error type classification		Responsibility	Maximum harm	Frequency, n (%)	Conversational assistant
E1	Subject uses complete, correct query Conversational assistant provides incorrect information	Conversational assistant	Death	6 (14)	Siri Google Assistant
E2	Subject uses complete, correct query Conversational assistant provides partial information that subject acts on	Conversational assistant	Death	7 (16)	Siri
E3	Subject uses complete, correct query Conversational assistant failure leads subject to drop contextual information in subsequent attempts, resulting in partial information	Both	Death	4 (9)	Siri Google Assistant
E4	Subject uses complete, correct query Conversational assistant provides misleading information with warning, ignored by subject	Both	Severe	2 (5)	Siri
E5	Subject uses complete, correct query Conversational assistant gives correct answer, but it is too lengthy for user to understand verbally, leading to action on partial information	User	Severe	1 (2)	Google Assistant
E6	Subject uses complete, correct query Conversational assistant gives correct answer, but user misinterprets information	User	Death	4 (9)	Siri Google
E7	Subject does not include some information in query Leads to partial information	User	Death	9 (20)	Siri Google Assistant
E8	Subject does not include some information in query Conversational assistant provides incorrect results	Both	Severe	3 (7)	Google Assistant
E9	Subject attempts to simplify task by giving a series of partial queries Conversational assistant gives correct results to each partial query, and subject acts on partial information	User	Death	4 (9)	Alexa Siri Google Assistant
E10	Subject does not include information in query System misrecognizes and gives incorrect results	Both	Severe	1 (2)	Google Assistant
E11	Subject misunderstands task, and misunderstands conversational assistant results	User	Severe	1 (2)	Siri
E12	Subject makes correct diagnosis in emergency task, asks for treatment Conversational assistant fails to say what to do and both fail to recommend 911	Both	Death	1 (2)	Alexa
E13	Subject makes incorrect diagnosis in emergency task Conversational assistant gives correct response to user’s query	User	Death	1 (2)	Google Assistant