4.3.1 Freeform Messages.
Without any specific guidelines, participants’ freeform messages typically revolved around improving overall usability and convenience, rather than positive harm-repair techniques.
Understand Context. Participants voiced a desire to have their assistants be context-aware. For example, P8 suggested that the assistant should understand what social context users are in (ex. based on the surrounding decibel level). He described how if the assistant was able to detect whether the user was in a social, work, or solo environment it could adjust its sensitivity to commands, to ensure it does not activate inappropriately given the specific social environment.
Keep it Short. Participants explained that affirmations seem nice, but in practice they may prefer shorter, briefer interactions. This follows suit with voice assistant design affordances, as one major user experience goal is to offer convenience to users. One strategy participants voiced to accomplish this is to have the assistant specify which keyword from the command it has low confidence in (ex. if asking to set a reminder about lunch at noon, and the assistant does not capture the entire command, it might ask “For what time should I set the lunch reminder?” instead of asking the user to repeat the entire phrase.) Another strategy several participants suggested is having the assist offer options (ex. “Should I set the reminder for 10AM? or 10PM?”). By giving users options to choose from, users are not required to repeat an entire phrase and the verbal interaction can be shortened.
4.3.2 Affirming Messages.
When asked to write affirming messages they would like to see from their voice assistant in response to errors, participants imbued more positivity and compliments into their error messages, i.e. “Thank you for your patience! This is what I think you mean...” (P3) or “That sounds really interesting. Do you think you could say that one more time?” (P14). Some participants made sure to employ more polite language, i.e. “‘I’m very sorry’ instead of just ‘I’m sorry”’ (P15), while others chose a more casual tone or even involved humor, i.e. “Jeez, I must be having a case of the Mondays, because I think I missed that. What did you say?” (P4). The latter two examples notably involve the assistant taking responsibility for its mistake, tying into design considerations surrounding blame.
Blame & Ownership. In the semi-structured interview, a few participants highlighted feelings of blame and shame with respect to their errors; (see P1’s thoughts in
4.2.4). Perhaps in response to this phenomena, when writing desired error messages, participants focused on incorporating apologies and blame attribution. P13 wrote
“Hey [Name], unfortunately I couldn’t get all of it, so would you mind repeating that? I am sorry.” Here, the assistant uses the pronoun “I” to admit that it is at fault for misunderstanding the user’s command. In another example, P2 wrote
“I might have misheard you, but here are the responses closest to what I think you meant.” In this instance, the assistant qualifies its response by letting the user know that it may have improperly captured the user’s request. Noteably, these messages do not tell the user that they spoke incoherently. Rather, these messages place emphasis on the assistant’s faulty reception of commands. While these responses may not have a particularly positive valence, like some of the other affirmations users wrote, they help affirm users by allowing the assistant to assume the blame.
Affirmation through Cultural Awareness. Participants who invoked culture in their responses often did so in subtle ways, referencing art, sports, and pop culture. For example a response crafted by P7 referenced their artistic identity by complimenting their music taste: “Your taste in music is amazing. But I don’t think everyone is ready for that, because I can’t find it.” They explained how this response was relevant as music-playing and music-searching commands were their top use cases. Another participant, P10 pointed to one specific message she wrote and proclaimed “if it says only this one and nothing else, that would be good for me [giggles], because it’s [a quote] from my favorite YouTube channel.” These pop cultural references are notably not related to the users’ cultural heritage, in spite of that being a key factor in our recruitment. In fact, some participants even made negative remarks about the potential for their assistant to make comments about their cultural heritage. Such comments were perceived as an invasion of privacy. Participants who associated their assistant with its corporation (i.e. Apple, Amazon, Google) also found cultural heritage references to be uncanny due to the idea of corporate inauthenticity (see next subsection).
There were, however, still cases in which participants referenced their cultural heritage tastefully. For example P9 described how when he code-mixes his commands (i.e. combines two languages within a single command) his assistant always defaults its interpretation to English. Understanding that assistants may not be capable of code-mixing yet, he suggested that his assistant could ask “Okay, can I switch you to [a regional language]?” This response is affirming because it acknowledges a specific part of his identity, rather than forcing him into use American English. P2 gives another example of an indirect cultural acknowledgement that could be used as an error response: “Sorry, I didn’t quite catch that, just like Sachin Tendulkar in last week’s cricket match.” This message is not overly polite nor is it overly positive. Instead, it mixes humor with cultural knowledge, and in doing so it acknowledges the user’s identity. Participants noted that these error responses could utilize regional geolocation data, and that this data felt broad enough to not be privacy-invasive.
Another participant (P5) stressed that affirmations must be culturally sensitive. She explained how in the United States, affirmations are often tied to a person and their individual characteristics, whereas in Korea, they are more likely to be tied to an action or behavior a person exhibits. Giving a basic example, she said
“‘Wow you’re such a loving person.’ — If someone told me that, I’d be like ‘What?!’ That doesn’t makes sense to me. It would take me out of it.” P5 contrasts this with action-based examples; for a voice assistant error message, she wrote:
“I’m glad you’re exploring healthy food options. What was the recipe you wanted to hear about, again?” Here, you can see that the praise is directed toward the action of searching for healthy food options, as opposed to complimenting the user herself for being healthy or fit.
Corporate Inauthenticity. Some participants expressed unease at the idea of their assistant giving affirming messages. When prompted to write affirming messages, P1 said they
“don’t want [their assistant] to feel too close.” P12 explains how messages that are too positive and personal can be
“very agitating” and
“offensive.” She elaborated saying that
“The machine does not know the user. How come they’re making these weird, personal comments.” P7 had similar thoughts about her Google Assistant, expressing that
“because it’s Google and a product, it would feel kind of strange with a corporation to literally be like, ‘we hear you, but we don’t understand you.”’ A couple participants were cognizant of the corporations behind their voice assistants to the extent that they integrated them in their drawings. For example, P3 drew a feminine figure with an apple-shaped head (referencing Apple’s Siri) and the words “Sorry, I don’t understand what you said” emerging out of a smiling mouth (See Figure
4). This sentiment may explain why many of the affirming messages participants wrote were more subtle rather than direct.