Designing for Harm Reduction: Communication Repair for Multicultural Users' Voice Interactions

There are several user groups for whom voice assistants work especially poorly. High error rates have commonly been documented for users who do not speak white American English [29]. Non-native English speakers tend to rate voice assistants lower on usability metrics compared to native English speakers [64], and tend to be less satisfied with their experience [57]. In another study, researchers found that regardless of English fluency, users would still experience greater errors with assistants if they had a “non-standard” accent [58]. Importantly, voice assistant challenges are not limited to those who speak English as a second language or to those who are multilingual. Native monolingual English speakers can also face a myriad of challenges, and these challenges are especially exacerbated for historically marginalized groups. For example, Koenecke et al. noteably documented a large gap between the error rate that Black Americans and white Americans experience [35].

This disparate error rate may be especially harmful for Black or multicultural individuals who use their assistant for medical information-seeking. It is known that these devices are not optimized for these user groups [26], and voice assistant medication name-recognition has been found to deteriorate for users with non-Anglo American accents [58]. These voice assistants have also been documented giving inaccurate information, leading users to false conclusions and inadvisable medical behavior [6].

Beyond the explicit bodily harms of a poor medical device, high error rates can lead to two other specific types of harms: psychological harms and quality-of-service harms. Mengesha et al. makes a case for emotional or psychological harm, documenting the lived experiences and feelings of Black voice assistant users through a diary study. Wenzel et al. verified these findings quantitatively, demonstrating in a controlled experiment how voice assistants can negatively affect marginalized users’ sense of self-consciousness, self-esteem, and emotional affect [87]. Furthermore, Shelby et al. has identified several quality-of-service sub-harms related to a high error rate, namely alienation (i.e. self-estrangement), increased labor (i.e. the need to exert extra effort or spend extra time), and service or benefit loss (i.e. diminished value) [75].

Shelby et al.’s work highlighted above is one one of the few formal taxonomies that uses voice assistants as a case study. However, voice assistants were not the focus of Shelby et al.’s work, and thus their categorization of harms is not comprehensive. Similarly, Dyal-Chand’s powerful Autocorrecting for Whiteness outlines several harms that arise from autocorrect, and asserts that these harms may be translated to various other AI technologies [16], however researchers have yet to adapt her harm taxonomy. Weidinger et al. proposed a taxonomy of risks for language models, including harms such as discrimination and misinformation [86]. While this work is no doubt valuable and relevant to the language models that power voice assistants, it does not analyze specific use-cases unique to the voice assistant user experience. Other comprehensive harm taxonomies have been developed for alternative technologies, such as online content [69] and online behavioral advertising [92], yet a proper taxonomy and comprehensive harm documentation for voice-based technologies has yet to be developed.

Furthermore, beyond what Shelby et al. has contributed, there is little work that identifies the granularities and sub-harms of quality-of-service harms. As users at the margins are typically the ones subject to diminished quality-of-service, it’s imperative that we understand precisely what downstream effects they face as a result.

While the growing body of work dedicated to documenting harms is imperative, there still remains an unaddressed need to develop designs that aid in harm reduction and repair. A common response to the existing disparity in error rate is to create more diverse datasets to train voice assistant language models on [35, 86]. This solution, however, has a slew of its own challenges [88]. Acknowledging the limits and difficulties that arise with improving language models and collecting representative voice data, it is worth exploring ways to reduce harm that go beyond technical improvements [21]. Our study proposes looking toward social psychology to inform harm reduction in voice assistant communication repair.

We begin by building upon recent work that positions voice interaction failures as microaggressions [87]. Prior research in psychology has revealed that experiences with microaggressions can trigger stereotype threat, the anxiety that emerges from the fear of being the target of a stereotype or of unintentionally conforming to a stereotype [78]. The experience of stereotype threat can be associated with a number of debilitating cognitive and affective outcomes, including: heightened self-scrutiny and self-reproach; reduced self-esteem; and threats to a sense of safety and belonging [3, 74, 90]. At the same time, the impact of threat can be reduced by a number of methods, such as actively negating or dismissing an activated stereotype [32, 53]; increasing the accessibility of a positive aspect of one’s identity to counteract a negative stereotype [66]; or affirming one’s values as means of reinforcing one’s self-integrity and providing a buffer to the psyche against threat [47]. Research on interventions to mitigate the effects of microaggressions and stereotype threat have found that they are more effective if implemented in subtle ways (i.e., in ways that do not position an individual as vulnerable or weak) and, at the same time, are asset-based (i.e., leverage personal strengths as a means of empowerment) [68]. However, work studying how these interactions may be mediated through conversational agents, let alone technologies in general, is limited.

Before conducting a redesign that integrates psychological theory, we first require an understanding of the current state of communication repair. Cuadra et al. found that if a voice assistant recognizes and repairs its mistake, users will evaluate it more favorably [11]. In a similar vein, Mahmoud et al. concluded that voice assistants that assume blame and apologize sincerely for their mistakes are perceived to be more intelligent, more likeable, and more effective in their error recovery [43]. Kim et al. explored different methods for communication repair for vehicular voice interfaces with mixed results. Overall, they recommend using brief natural commands that include an explanation or status report to help minimize confusion for users [33].

Notably, the work summarized here does not focus on users at the margins of the voice assistant user experience. As an increasing body of literature has come to demonstrate the unique experiences and needs of marginalized voice assistant users and users from diverse cultures [8, 51, 87], it has also become increasingly important to understand how communication repair should be adapted for these specific populations, as they are more likely to experience errors and conversational breakdown.

Researchers have also highlighted contrasting design needs between non-native and native English speakers. For example, non-native speakers appreciate longer turn-taking time, whereas native speakers find long turn-taking time disruptive [93]. Other work on multilingual users suggests that future voice assistants should be designed to code-mix, i.e. be capable of alternating between different languages within an interaction [8].

The study was conducted in a lab at the authors’ university campus. Each study session lasted for 40-60 minutes, depending on how long the participant took to complete the study tasks. Upon entering the room, the participant was asked to read and complete a paper consent form. The participant then engaged in a directed storytelling activity in which they were asked to describe a frustrating experience they had with their voice assistant. Participants who had experiences with multiple assistants were asked to focus on a single assistant. Methodologically, having each participant focus on a single assistant allowed for internal consistency across study activities. Practically, focusing on a single assistant ensured that the research activity would take no more than an hour of a participant’s time. (An exception was made for P6, who weaved his experiences with both Amazon Alexa and Google Assistant into a personal narrative.) Based on the participant’s story, the first author (PI) led a semi-structured interview. The directed story-telling and semi-structured interview aimed to partially address RQ1 in addition to largely focusing on RQ2.

The second activity asked the participant to draw a personified version of their voice assistant, and then describe the drawing to the PI, a method similar to that used in prior work [12, 36]. In addition to helping us effectively address RQ2, this participatory visual method helped the researchers renounce some of their power and cultivate a collaborative research ethos through mutual discovery [56].

The third activity asked the participant to write a love or breakup letter to their voice assistant [19, 25]. This letter-writing activity revealed both participants’ conceptualizations of their assistant and harms that they had experienced, addressing both RQ1 and RQ2.

The fourth and final activity was generative in nature, and intended to address RQ3. It asked the participant to write messages that they wished their voice assistant would say when an error occurred (i.e. conversational repair). The participant was asked to write four affirming messages (i.e. messages of positivity and warmth, or cultural messages) and two “freeform” messages (i.e. messages without any formatting or content requirements). After each of the interactive activities, participants were invited to elaborate and explain their illustrated/written choices.

P1 through P5 had the PI present for facilitation and two research assistants present for notetaking. P6 through P16 only had the PI present for the study, and had their sessions audio-recorded with consent. There was no observed difference in participants’ self-disclosure across the two facilitation styles.

Data analysis was conducted through both inductive and deductive coding by the PI. For the inductive open coding, initial descriptive codes [89] were applied which helped note relations between the topics which emerged from participant quotes and the three research questions. These typically corresponded to the intended activity as noted in Section 3.2.1. Topics mentioned by participants that did not directly relate to a research question received their own unique codes. Sub-codes [67] were then applied to denote topical results to the research questions. (See Table 2 for coding example.) The visual and written data we collected were coded alongside participants’ verbal descriptions. Their verbal descriptions heavily aided the interpretation of the data, so as to avoid overinterpretation of the visual materials. Overall, there were 13 miscellaneous codes (not directly related to an RQ), 6 sub-codes that related to RQ1, 6 sub-codes that related to RQ2, and 15 sub-codes that related to RQ3 Additional deductive rounds of coding using the taxonomy of harms developed by Shelby et al. [75] and Dyal-Chand [16], respectively, were also completed to address RQ2. All codes received a frequency count.

Many participants described their assistant as intelligent yet incompetant, highlighting the limits of their assistants’ abilities. This largely manifested as the participants describing their assistant as incredibly knowledgeable (“smart”), but lacking in their communication and task execution abilities (“dumb”). This insight emerged in a few key ways. Firstly, many participants either drew their assistant with glasses, as an archetypal indication of intelligence, and/or as a secretary “because secretaries tend to know everything about the area” (P14). Simultaneously, however, their assistants were also depicted as being low-tech. In their sketches, they noted their assistants working with many paper documents and pens and lacking technological equipment. Those who did depict their assistant with technology made sure to express their assistants’ limitations in other ways. For example, P15 included oversized manually-operated machines in their drawing (Figure 2) as opposed to modern devices. P15 went on to describe that the personification of his ideal assistant would be a person working with a laptop, implying that the current assistant is not advanced enough for his needs. In contrast to P15, P6 depicted his assistant using modern technology (a tablet), but his overall sentiment was still the same as P15. He described how in spite of having all possible resources, his assistant still failed to actually assist him: “it has all the tools, it has everything, probably everything in their arsenal, but [it’s] not able to help me.” In a related mental model, P11 expressed these limitations by drawing his assistant Siri over a hill in the distance. He explained that because of its technical limitations it feels like Siri is “far from me, and not next to me.” Ultimately, regardless of how they chose to depict their assistant, it was clear that participants believed their assistants to be knowledgeable, but unable to properly harness their knowledge.

This perceived incompetence often translated to a lack of emotion. In many participant drawings, the voice assistant appeared unfriendly and emotionally cold, with a straight face. P7 described how Google Assistant is “always very serious” and “would not be a very light hearted person” (Figure 3a). In a similar vein, P10 describes their assistant as someone who is “grumpy” and displays little emotion. P10 also stressed that her assistant is “somebody who has no social skills and doesn’t understand the details of social situations.” She emphasized how she associated the lack of facial expression in her drawing with her assistant’s “dumbness.” Multiple participants made a connection between their assistant and real people they knew in their life, citing coworkers who “cannot be creative” (P12), friends who are oblivious and “believe [sarcasm] is true” (P10), bad listeners who require you to “shout at them...two to three times” (P3), and acquaintances who are an introverted “homebody” who “know random facts” (P14). When participants did depict their assistant as emotionally expressive, it was often with an artificial fervor. For example, P6 depicted his assistant with an overzealous smile, annotated with a monologue of “HOW MAY I HELP YOU!!!” (Figure 3b) While these two emotional depictions are apparent opposites (lack of emotion vs. overexpression of emotion), they are similar in that they are both extremes representing atypical social behavior.

Participants’ difficulties conceptualizing their assistant often stemmed from a confusion about their assistants’ human-like nature. As many scholars have demonstrated in the past, people unconsciously attribute human characteristics to non-human entities, and accordingly treat non-sentient entities as humans [54]. This phenomenon extends to voice assistants [49, 70, 87], and several participants expressed awareness about their miscalculated expectations. P16 describes: “When I correct people in real life, they learn! They fix whatever information they had before and don’t repeat the mistake!” P12 explains in more detail, saying: “I’m just used to talking to a person and having a person understand. And perform a function just smoothly, no issues. And I can’t understand why a computer that’s like been made to work in a similar fashion just cannot with the easiest functions. So I think getting used to working with people who are very good listeners, and understanding, and just know how to do the task spoils me.”.

Several participants had experiences with multiple voice technologies, and a consistent trend across all of these participants is that they were overall less satisfied with Siri than with other assistants. These participants explained that they had switched phone providers (i.e. Android to iPhone), for reasons unrelated to the assistant experience and as an unintended consequence their voice assistant experience suffered. P9 begins his letter to Siri, “Dear Siri I’m done done. I’m only tied to you because of Apple. But would switch to Google Assistant the day Apple allows me to switch. You talk nonsense half the time, and are unable to do what I want.” He went into greater depth through his drawings, explaining that Google Assistant “is way more sophisticated and well presented...in terms of dressing this would look like an investment banker. And listens to Beethoven-style like professional music.” In contrast, his idea of Siri is “dressed as a Silicon Valley hippie youngster, who listens to rock and roll and is way more simple [minded].” These archetypes are reflective of the perceived competence and performance accuracy of the two assistants. While this difference in evaluation across assistants is not our primary finding, it has important implications for how researchers should approach future studies on voice assistants.

Multiple participants described instances of misattributed blame, in which they were blamed for a problem that was actually the fault of their voice assistant. This experience often led to interpersonal conflict. For example, P8 expressed frustration regarding how Google Assistant incorrectly inputted meetings in his calendar. “I missed some important meetings, and it was I who was blamed for missing those meetings.”¹

Of course, non-professional relationships are also at risk. P13 explained how his grandmother began using Siri to make phone calls, as the iPhone interface was difficult for her to navigate. This led to a barrage of calls from his grandmother to unintended recipients, in addition to random calls throughout the night. P13 fervently explained his family’s deep frustration with his grandmother: “We wouldn’t directly question the technology because we would say ‘okay, this has to work. If you’re doing it correctly, this has to work.’ So we started pressurizing her like ‘okay, you are not saying it right. You have to say it right!”’ Later, in a visit to his grandmother, they learned that she was, in fact, using the correct voice commands, but it was Siri who was directing calls to the wrong people.

In both of the stories outlined above, the conflicts that arose from these conversational breakdowns led participants to revert to manual methods. P8 started using a paper journal to keep track of his meetings, which he felt was more reliable than voice-based or digital methods. P13’s family replaced his grandmother’s touchscreen iPhone with a flip phone and taught her to use speed dial. This regression towards older, more manual, methods echoes economic harms outlined by Dyal-Chand: “While we may also suffer from losing a job...” (reflective of P8’s struggle), “the core economic harm here is from paying the same price for a less useful product.”²

Interestingly, despite the frustrations all participants faced, when describing their ideal assistant, they did not ask for one that was completely fluent or multilingual. P9 said “it works perfectly fine, if you understand that you’re talking about a machine.” Similarly, when describing the troubles he has communicating in Spanglish using the voice-to-text feature, P14 was sympathetic to the limits of the technology: “it’s not its fault, I guess, because it’s set to English. However, he still voiced some desire for a multilingual assistant, even if it knows only “a little bit” of Spanish. P13 explained that he believed a multilingual assistant seemed too difficult to develop, and instead suggested that assistants should simply advance towards recognizing different accents. He qualified this statement further, saying that assistants “should be able to pick up accents to some extent, not to 100 percent, but to some extent.” Users have been underserved by their assistants for so long that they’ve fully recalibrated their expectations. They expect to buy products that do not serve them. They do not even outwardly wish for products that can fully understand them.

The high disparate accuracy rate leads to “unequal access to and through technology” [16], which can often lead to what Shelby et al. term as an “increase labor” harm. While users who benefit from voice assistants may experience greater efficiency and convenience in their day-to-day workflow, others must exert extra effort simply to attempt to have their assistant understand them. Even after going through this extra labor, the assistant is not guarranteed to function properly. P16 echoes this sentiment well, demanding angrily in her break-up letter to her assistant: “Why should I have to repeatedly bend over backwards just to get you to understand what I’m saying? How come my all-American named friends get treated better than I do? Why is convenience only allowed for some people?”

In spite of all our participants being multilingual, only a subset of them have tried using non-English languages with their assistant. More common, however, was the phenomenon of codeswitching, in which participants would alter their style of speech to be closer to a white American English speaker. P16 described her experience of trying to call family members, saying that in order for it to work correctly “we have to deliberately mispronounce [their names],” stripping away her family’s culture an identity. P10 gave a clear example, alternating between different voices, stating: “I also feel dissatisfied in one particular way that when I say ‘OK, Google,’ it does not understand my accent. And when I say it in a different accent, like, ‘OK, Google,’ that’s when it understands.” Users must assimilate to a standard white American middle or upper-middle class accent, effectively altering their identity, just to be recognized by their voice assistant. Even then, they are not always successful because “how good your fake accent is also matters” (P13).

While accents are indeed a powerful reflection of one’s identity, there are other ways identity and cultural harms can manifest. Vocabulary also matters. P1 described his frustrating experience trying to add Indian spices to his online shopping cart, to no avail. He explained he tried adding it both by their English translation and their Hindi name. When contemplating what he could do to solve his problems with Alexa, he suggested “Maybe I should try learning words which could act like a replacement...or maybe the word I had in my mind is not the right word for that item.” — In spite of already trying two of the most popular languages in the world, he was still convinced his commands were not good enough. His final suggestion was that he should alter his identity with “accent training.”

P7 shared their experience growing up in a Black household, and noticing how their experience with technology changes as their cultural language gets adopted by the young American majority: “I just use a lot of Black slang, a lot. And usually, [my voice assistant] does not know what those things mean until it becomes a mainstream word on the internet.”

Unique from Shelby et al.’s taxonomy’s “Technology-facilitated violence” and Storer et al.’s “Physical Safety” tradeoff, we highlight that violence and physical danger are not solely a byproduct of a malicious actor taking advantage of a system. There are indeed cases where users, against anyone’s own desires, are put in danger due to a faulty system that has been promoted without great consideration for a diverse user base. A clear example of this was in the case of driving and navigation. While prior work has highlighted car navigation as preferred use-case of voice assistants, yielding higher ratings of usefulness, trust, and positive emotions when compared to alternative tasks [27], our results suggest that this is likely not the case for marginalized users. P12 described “Oftentimes, I feel like I’m pronouncing things very clearly and loudly, but it still can’t understand me. And I don’t know what’s going on. And I don’t know where I’m going. So it’s just this, this frustrating experience, and very dangerous and confusing.” Other participants echoed these feelings of fear, danger, and uneasiness when using navigation apps. Importantly, in such scenarios, the drivers using voice-interaction are not the only ones in danger. Rather, there are many stakeholders (i.e. passengers and other drivers on the road) for whom their safety is put at risk.

While all harms have the potential to cause some level of emotional distress, there were cases where negative emotions, like sadness or deflated self-esteem, were particularly salient. Such was the case for P14. He explained, in the context of sending text messages with bad news, when his assistant fails to transcribe his voice it “adds to the frustration just like, ‘Man, I gotta deal with this,’ like little thing after a little thing. It builds up.” The simple frustration of a conversational breakdown can exacerbate already negative emotions. Recalling prior work on the intersection of race and conversational breakdowns [87], we’d like to reiterate that such frustrations are likely to be exacerbated for people of marginalized identities, due to the emotional and cognitive impacts of microaggressions.

Importantly, harms are not mutually exclusive. It is not unlikely that when someone experiences, for example, an Identity & Cultural Harm that they also experience emotional harm as well. P16 describes how misunderstandings affect her through a range of different emotions: “It definitely makes me feel a little bit alienated worst case scenario, and mildly annoyed, best case scenario, neither of which are very fun.” Negative emotions may even make other harms more potent, heightening their effects, and thus should be taken seriously.

While significant prior work has focused on English-based voice assistants, a subset of our participants described experiences of using voice assistants with alternative default languages. Their experiences showed that accurate speech recognition is a problem that persists across various agents. Similarly to how those who speak English as a second language had trouble using English-based assistants, P7, P15, and P5 described their struggles speaking Korean as a second language to Korean assistants. Even those who are native speakers have trouble using voice technologies. P13 and P5 described challenges their families faced speaking with native regional accents to devices programmed for Hindi and Korean, respectively. In these cases, the language “ideal” is not that of a white American, but of a person who is of a high socioeconomic class from that country.

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.

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.

As a next step for communication repair redesign, we suggest a full evaluation of the repair strategies offered in this paper. This would involve crafting communication repair dialogues based on our guidelines, creating comparable dialogues that lack these elements to serve as a control, and running an experiment that evaluate user responses and reactions to the various experimental conditions. While we believe our interventions are promising, as they are guided by both participatory work and prior empirical studies, we advocate for rigorous testing and evaluation before deployment, in line with the harm-reduction agenda.

In the present study, we specifically studied harm reduction techniques for voice assistants. While the scope of this technology is limited, we believe our interventions have promise in other contexts. For example, in text-based language technologies such as auto-correct and auto-reply, which are also embedded with social biases [16], suggestions may be accompanied with an abbreviated blame-redirecting message which emphasizes the limits of the technology’s abilities. If a user rejects a suggestion, a message of affirmation may briefly appear to reassure the user of their identity and choice. These techniques may even extend beyond AI-based language technologies and be adapted into error messages for various technological services and goods more broadly. Such interventions deserve further research and evaluation to inform their proper design.

The harms we offered in this paper were specifically related to high speech recognition error rates for multicultural voice assistant users. We chose this focus as we believe multicultural and multilingual users specifically are underserved in the AI voice technology development pipeline, and furthermore we wanted to demonstrate that error rates can do much more than reduce the basic usability of a product. We stand by this framing and believe it is important. Of course, we also acknowledge that there are many other voice assistant harms unrelated to error rate that are also deserving of attention, such as privacy and surveillance violations [38, 91, 95] or misinformation concerns [6]. A comprehensive voice assistant harm documentation has still yet to be made.

While we recruited participants who identified as multicultural or multilingual, we would like to highlight how these participants had other intersecting identities. For example, participants may have identified as immigrants, expatriates, non-American citizens, non-native English speakers, etc. These various identities, while sometimes related, are distinct from the labels of multicultural and multilingual. The results we present in our study may be connected to any of our participants’ unique identity characteristics, however due to the nature of intersectional identities, it is difficult to pinpoint the exact connections. Future work may consider capturing these granularities.

Our qualifying participant recruitment criteria was determined by finding commonalities between the research team’s own identities, honoring feminist standpoint theory [2], and users who have had demonstrated inequities with voice technologies. While we aimed to recruit a diverse set of participants, our participants were still limited in a few ways. Our participants largely self-identified with Indian and American identities. A few identified with other parts of the Asian diaspora, and beyond this, we had four participants identify as queer, Ashkenazi Jewish, Black, and Mexican, respectively. Future work should consider how the harms we identified may generalize to those who were in the minority of our sample or absent from our sample entirely.

Regarding age, with the exception of P12, all of our participants were young adults (18-29 years old). Furthermore, most of our participants were students. Future studies may aim to understand how the presented harm reduction principles may be adapted to older adults, children, and non-students.

Only two of our participants identified as non-binary and transgender. Rincón et al. [65], has already examined needs and explored the tensions of non-binary and transgender voice assistant users, future research may then focus on designing specifically for communication repair for trans and non-binary folk. Naturally, there are a myriad of other identities that we have not touched upon in this work, including but not limited to, indigenous users, neurodivergent users, users without higher education, et cetera. Future work should respect and explore these user groups in more depth.