“No, not that voice again!”: Engaging Older Adults in Design of Anthropomorphic Voice Assistants

Abstract

Conversational voice assistants are often imbued with personality and human-like characteristics (e.g., gender). While researchers have begun to examine and design for the downstream societal impacts of voice assistants encoding characteristics such as gender, we know little about other human-like characteristics such as age that are encoded in an artificial, yet, anthropomorphic voice. As older adults continue to adopt voice assistants, we brought older adults into an activity to customize human-like characteristics for their voice assistant. Our findings reveal the different stereotypes and assumptions individuals associated with voice assistant characteristics (e.g., age, gender, race). We also describe individuals’ motivations behind customizing or not customizing these characteristics. We discuss how biases get encoded through our design process, marginalizing older adults and other non-dominant user groups and call for a need to examine the systemic, yet unspoken, power structures encoded in anthropomorphic technologies.

1. Introduction

Google assistant is a young woman from Colorado, the youngest daughter of a research librarian and physics professors who has a B.A. in history from Northwestern, an elite research university in the United States; and as a child, won US $100,000 on Jeopardy Kids Edition, a televised trivia game. Google Assistant used to work as a personal assistant to a very popular late night TV satirical pundit and enjoys kayaking [44, 103].

This description was provided by the Google Assistant team to describe the backstory for Google’s widely used voice assistant [44, 103]. It was reported that this backstory was provided to the actress whose voice was recorded for Google Assistant to help her produce “the exact degree of upbeat geekiness”[103]. Though discussing a technology, the description contains many human-like characteristics. Some are explicit, such as region (“Colorado”), gender and age (“a young woman”), education (Bachelor’s degree from “Northwestern, an elite research university”), and career (“personal assistant”). Other characteristics, such as race, class, and income level, are not explicitly mentioned. Yet, the description of education and use of phrases such as “elite” can signal associations with certain characteristics such as class.

CSCW and HCI researchers are unpacking explicit and implicit characteristics of commercially available voice assistants, and how they impact users of these technologies, and society, more broadly [84, 109]. Performing gender through voice assistants has been an ethical concern [100]. Given this technology’s role as ‘assistant’ users may apply stereotypical gender knowledge to the interaction [76] and have negative stereotypes of women’s role in society reinforced [97, 122]. Some researchers are examining alternative approaches to designing binary gender voices [81, 112]. Others problematize the linking of vocal range to gender [92, 108]. Performance of race is arguably more implicit than gender, as descriptions of voice assistants such as Google’s Coloradan young woman do not tend to mention race. Yet, research is beginning to show how race is indeed encoded, with users from racial minority groups perceiving voice assistants as a “white woman¹” [11, 87]. As with gender, voice interfaces can perpetuate oppressive societal conditions, such as “whiteness as the norm” [87]. Together, this literature makes the case that there is an interplay between human-like characteristics attributed to voice assistants and marginalization based on these characteristics.

If age as a trait is encoded in voice assistants, one group of users likely to be marginalized are older adults. Brewer et al.’s study envisioning equitable voice assistants with Black older adults traces ways that current characteristics encoded in voice assistants exclude older adults [11]. The authors unpack responses across two workshops to show the many ways that voice assistants are currently inequitable in terms of their ability to relate in culturally, regionally, linguistically, and community-appropriate ways [11]. While this paper focuses on the design of equitable voice assistants through the constructs of fairness, representation, and inclusivity, we are inspired by the ways they were able to elicit participants’ perceptions of not only how voice assistants encoded characteristics currently - e.g., “white” and “young” - but also how participants would like voice assistants to encode different kinds of characteristics (e.g., fairness in voice assistants would look like... “having a Black male to answer my questions would make me smile”) [11].

We extend this body of work by further investigating characteristics that older adults would like their voice assistants to have. We conducted an activity where older adults customized human-like characteristics (e.g., age, gender, education) for voice assistants. In analysing the responses of 22 participants, we ask the following research questions: What associations (including assumptions or stereotypes) do older adults make with the characteristics of their voice assistants (RQ1)? What were older adults’ motivations for selecting specific characteristics for their voice assistants? (RQ2)

Given current conversations about the ways that voice assistant characteristics interplay with historic marginalization around certain identities, our analysis is informed by Patricia Hill Collins’ work examining the role of structures in defining power and oppression, and the impact of such structures on various identities [19]. Collins’ conception of a matrix of domination enabled us to tease apart the ways that involving end-users in voice assistant design can lead to resistance, but also alignment with associations embedded in current structures of power. We detail the ways that participants customized the age, gender, education, social class, region and accent, and race of their voice assistants, and the assumptions or stereotypes that individuals associated with the characteristics they selected (or avoided). Our findings also detail how older adults customized their voice assistant characteristics with three primary motivations: a) to make the technology more relatable to themselves, b) to create a broadly relatable technology for wide appeal, and c) to treat the voice assistant as a piece of technology or object in house.

By engaging older adults in design of anthropomorphic characteristics of voice assistants, our work makes two main contributions. First, we provide an empirical understanding of how older adults customize and associate human-like characteristics with voice assistants, extending prior HCI and CSCW research that examines voice-based technologies for older adults [11, 16, 48, 56, 82, 104, 106, 113]. Second, our work contributes to ongoing CSCW efforts critically examining the societal implications of designing anthropomorphic voice assistants [13, 92, 99]. Our findings analyzed through Collins’ matrix of domination illustrates the different kinds of stereotypes that can be associated with an anthropomorphic voice. We surface how several of these characteristics become invisible, potentially encoding harmful associations into the default voice. This calls for a need to carefully navigate our current process of designing anthropomorphic technologies, by thinking through the invisible power structures at play.

2. Background and Literature Review

Below we review prior work on voice assistants and older adults, and the design of anthropomorphic voice assistants with human like characteristics. Next, we describe Collins’ matrix of domination which became central to our analysis.

2.1. Older Adults and Voice Assistants

Older adults are one population that have received significant attention as users of voice assistants, with several media outlets portraying older adults as a potential target market for this technology [104]. Correspondingly we see efforts to incorporate these devices in nursing homes and senior living communities [101]. Research has examined how voice technologies at home can support aging: by making it more accessible to control the home environment (e.g., smart lights, smart thermostat) with voice [86], in terms of assisting caregivers [125], medication monitoring and reporting health data to doctors [16], or by lowering the digital divide by enabling easier access to digital information as compared to traditional computing devices[85]. Features, such as, meditation, or playing games can contribute to wellbeing [79, 98], and several third party skills (i.e., third party apps on IVAs) can enhance daily routines for older adults [114].

Research has examined older adults’ concerns and the challenges they may experience in using these technologies that can pose barriers to adoption. Some older adults have expressed concerns regarding a lack of understanding of how these technologies work [6], with some expressing concerns in trusting and comprehending information provided by the voice assistants [9]. For older adults who are deaf or hard of hearing, the higher pitched default female voice or the speed at which the agent talks can be inaccessible [5]. Speech recognition may not work with vocal characteristics which can occur with cognitive impairment, such as pauses or hesitation [30, 59, 85], or due to technology designers not taking into account different cultural considerations and dialects [48]. Remembering wake words and composing concise commands for a task can pose challenges [58]. Some of these issues can be traced to bias in training data for speech recognition and natural language understanding models. The speech corpus used for training these models rarely includes data representing older adults, let alone older users from minority groups (e.g., [72]). This results in high error rates in recognizing and understanding speech of older users as compared to younger users [116, 120]. Collectively because of these issues, we see older adults experiencing conversation breakdowns when using these voice assistants [9], which can play a role in acceptance of these technologies by older adults [28].

Overall, research shows how voice assistants afford certain benefits to older adults, also noting current concerns, challenges, and barriers to adoption. Another thread of research has examined how older adults’ personify (i.e., ascribe human-like properties, characteristics, or mental states to) voice assistants [82]. Findings from this work suggest that older adults may have differing preferences for the human-like characteristics associated with voice (e.g., not all older adults in this study wanted a female voice, some wanted the ability to change to a different gender)[82]. In another study, not all older adults prefer the default voice of a “white woman” for their voice assistant [11]. Our work further examines such salient preferences of older adults with respect to design of human-like traits for an artificial voice.

2.2. Designing Anthropomorphic Voice Assistants with Human-like Characteristics

Intelligent voice assistants like Alexa, or Google Assistant interact with users using natural language similar to human-human conversations. And, relatedly, one practice that is becoming widespread is to design these voice assistants with particular personality and several human-like characteristics including age, gender, education or job [1, 84]—to act as a guiding influence on the voice assistant’s behavior and interactions with users [65]. For instance, beyond the Google assistant characteristics noted in the Introduction, consider human-like characteristics ascribed to another conversational assistant, XiaoIce, Microsoft’s China-based conversational agent:

XiaoIce has a persona of as an 18-year-old girl who is always reliable, sympathetic, affectionate, and has a wonderful sense of humor. Despite being extremely knowledgeable due to her access to large amounts of data and knowledge, XiaoIce never comes across as egotistical and only demonstrates her wit and creativity when appropriate [124].

Practitioners have argued that ascribing such human-like traits and personalities to the voice assistant is helpful in defining the interactions between the user and the technology [51] and helps copywriters write conversations that foster a cohesive presence of the voice assistant [71]. Some research has noted how ascribing these traits and personalities to voice assistants can lead to increased trust, engagement and intention to use the technology (e.g., [50, 75]). However, although creating such anthropomorphic voice assistants has become a common practice in industry, there is scarce research on best practices or design decisions involved in ascribing these human-like traits to technology. Most work in this space focuses on personality attributes of the voice [7, 117, 118]. For example, some research suggests matching the voice assistant’s personality to that of the user as it can improve likability, trust, and overall satisfaction [7]. In terms of assigning gender and age, XiaoIce was guided by large scale analysis of user data, where majority of the desired users were young and female, and to match it’s users, XiaoIce was modeled around an “18-year-old girl” [124].

Sutton et al. describes how any given voice has associations with qualities including geography (relating to region and accent), gender and sex, age, or social class [109]. Gender and age can become encoded through the vocal characteristics of pitch and loudness, affecting how a voice sounds like [67]. Age is additionally encoded through language as there may be generational differences in how and what one speaks [109]. Speech can become a marker of social class, with some studies showing a link between “r-pronunciation” and social class [60, 123]. Some have noted connections between the social class of the speaker, and what and how one speaks (e.g., the words used, length of sentences, or pauses) [12]. In context of anthropomorphic voice assistants, research has examined how gender as a trait is encoded, examining users’ associations and perceptions about gender of the voice. Gender can become associated with trust (women are likely to be perceived as more trustworthy [42]), or with stereotypical roles (associating women with caregiving or service related roles such as secretaries [22]), or even subordination and sexualization [52]. Extending this body of work, in this paper, we examine some of the above described traits that continue to be under-examined, but, are likely to be encoded or ascribed to an artificial voice: age, gender, social class, region, accent. We included the additional dimension of education, which has been explicitly ascribed to a broadly used commercially available voice assistant (i.e., Google assistant is from an “elite research university” [44, 103]). By asking older adults to brainstorm and customize these traits for their voice assistants, our goal is not advocate for ascribing such traits to anthropomorphic voices. Rather, we seek to understand in depth what assumptions and associations people derive with these traits.

2.3. Understanding the Complexities of Identities through Matrix of Domination

Patricia Hill Collins’ matrix of domination framework became central to unpacking our data [19]. Specifically, Collins includes in the matrix of domination gender, age, class, education, region and accent (through language bias), which were all traits around which we developed design activities; see Fig 1. Race, which is a trait included in the matrix as well, was not something we asked about explicitly but emerged from participant responses. Collins introduced the matrix of domination as a framework to understand power and oppression in her book, Black Feminist Thought [19]. Power is “the chance that an individual in a social relationship can achieve his or her own will even against the resistance of others” [119]. Rooted in acknowledging and valuing Black women’s lived experiences as knowledge, Black Feminism Thought examines issues of power and oppression that are complex and not based on any one singular identity but instead the intersection of identities (e.g., gendered racism) [19]. The impact of intersectional of identities, or intersectionality, has been well discussed [15, 20, 21, 24, 49, 88, 90, 111]. Collins distinguishes intersectionality and the matrix of domination as: “Intersectionality refers to particular forms of intersecting oppressions, for example, intersections of race and gender, or of sexuality and nation. Intersectional paradigms remind us that oppression can not be reduced to one fundamental type, and that oppressions work together in producing injustice. In contrast, the matrix of domination refers to how these intersecting oppressions are actually organized” [19, p. 21]. Collin’s matrix of domination considers the structures in which power exists and how power moves to enable oppression of certain groups and populations based on their identities, which is visually illustrated in Fig. 1 [19]. It is important to note that “power operates within the context of unspoken rules as to which social groups are subjugated to preferential treatment at the expense of non-dominant social groups” [32]. Further, who has power, to whom power is granted, and who has the ability to access the distribution of power has been largely defined by social status in the U.S., which have historically been influenced by the creation of dominant groups based on their identities (e.g., race, gender, socioeconomic status, education attainment, age). For example, Collins discusses how educational attainment is many times factored into whose voices are viewed as knowledgeable, important, and trustworthy [19]. In her work, she intentionally includes people from various educational backgrounds, often centering those with less formal education to contribute to academic knowledge [19].

Fig. 1.

Intersecting axes of privilege, domination, and oppression. Adapted from [74] and informed by Collins [19].

Recent HCI and CSCW scholarship has leveraged Collins’ work to examine role of power in technology design and computing education, and revealed how the intersecting systems of oppression exist in computing [32–34, 41, 89]. Erete et al., for example, applied an intersectional analysis of power to examine the co-design of community safety technologies with Black street outreach workers [32]. Rankin et al. use an intersectional analysis to examine “saturated sites of violence in which interconnected systems of power converge to enact oppression” in CS education for Black women [89]. Gilbert examines the role of power in online moderation and suggests an alternative model for moderation that addresses issues of power [41]. Erete et al. describes how they take an approach to their STEM education research for Black and Latina girls and their families by understanding the history of power and oppression and leveraging principles of transformative justice throughout their research process. Collectively, this work, suggests that understanding the complexities of power and the role that it plays in the oppression of those with non-dominant identities is imperative to equitable technology design. In this paper, we build on this existing research by providing insights about how older adults, as a non-dominant user group, perceive characteristics of existing voice assistants and how existing beliefs and stereotypes can be embedded and perpetuated in these tools, using the matrix of domination as a method of examination.

3. Methods

We conducted 1:1 creative ideation sessions (akin to brainstorming and ideation in formative design activities with older adults [47, 83]) with 22 older individuals. This activity was part of a larger study focused on understanding older adults’ preferences for voice assistants. All procedures were approved by the University of Maryland’s Institutional Review Board.

3.1. Participants

Our inclusion criteria for recruitment was individuals who were 65 years old or above, and used an intelligent voice assistant (e.g., Alexa or Siri) on a smart speaker device — a device where even a disembodied voice is likely to be personified and ascribed human-like mental state [82]. We recruited people with experience of voice assistants because brainstorming invisible technology concepts might not be intuitive for all older adults [66]. This can be minimized if individuals have experience using the technology [83]. Table 1 shows participant demographics. Our sample also included older adults with diverse abilities: individuals with vision impairment (P9 and P13 were blind, P11 had low vision), and people with dementia (P14, P15, P16) who had participated in prior studies with our lab and had capacity to consent. Our participants were technologically savvy and all reported using digital computing devices frequently, such as, smartphone, tablet, or computers. Half self-reported using other smarthome devices such as smart thermostat, smart doorbells, or smart lights. All used Alexa, except for P6 and P14 who used Google Assistant. All participants had used the device for more than six months, and, as such, had experience with the technology beyond the novelty period. We recruited participants from a large independent living community of older adults in the Northeastern part of the United States (N=8), an organization that connects Black older adults with research studies from the Midwestern region of United States (N=6), and snowball sampling (N=8). Our sample, predominantly based in the United States (apart from P15 from the United Kingdom), does not represent the diversity of older adults using voice assistants globally.

Table 1.

Participant Demographics

PID	Age	Gender	Race or Ethnicity	Highest level of education	Living with
P1	67	Female	Causcasian or white	Associate	Alone
P2	66	Male	Causcasian or white	Master’s	Spouse
P3	69	Female	Causcasian or white	Doctorate	Spouse
P4	71	Female	Causcasian or white	Master’s	Alone
P5	71	Female	Causcasian or white	Doctorate	Alone
P6	68	Female	Causcasian or white	Bachelor’s	Alone
P7	73	Female	Causcasian or white	Bachelor’s	Alone
P8	75	Female	Causcasian or white	Master’s	Husband
P9	65	Female	Causcasian or white	Bachelor’s	Alone
P10	77	Male	Causcasian or white	Doctorate	Spouse
P11	77	Male	Causcasian or white	Master’s	Alone
P12	81	Female	Causcasian or white	Master’s	Alone
P13	72	Female	Causcasian or white	Bachelor’s	Alone
P14	69	Female	Causcasianan or white, Other	Master’s	Daughter
P15	65	Male	Causcasian or white	Professional	Alone
P16	69	Male	Causcasian or white	High school or equivalent	Wife, daughter & her family
P17	70	Female	Black or African American	Bachelor’s	Alone
P18	78	Male	Black or African American	Bachelor’s	Wife
P19	83	Female	Black or African American	High school or equivalent	Alone
P20	83	Female	Black or African American	Master’s	Alone
P21	73	female	Black or African American	Bachelor’s	Spouse
P22	66	Female	Black or African American	Bachelor’s	Mother

3.2. Ideation workshops

The ideation sessions were up to two hours long (details in supplementary) and had three parts: familiarization, customizing voice assistant characteristics, and reflection. In addition to building rapport, familiarization eased participants into thinking about the characteristics of voice assistants. The customization phase was scaffolded to center older adults as knowledge experts by taking into account learnings from prior work (further described below). To allow for flexibility of participation in light of COVID-19, sessions were conducted in-person (N=4) and remotely over video conferencing (N=18). Sessions were audio/video recorded and transcribed for analysis. Each participant was compensated $30.

Part 1: Familiarization.

Past research notes some older individuals may experience challenges in brainstorming open-ended concepts associated with technology [66, 83]. Taking this into account, we first prepared participants to think about characteristics such as age, gender, or education of a voice assistant, which may not be readily intuitive for a user, let alone a group who is rarely included in discussion of these technologies. We familiarized participants by showing them example of how popular commercial voice assistants, such as Google Assistant, often have a fictional character with human-like characteristics like age, gender, region, or education (as described in the Introduction).

Part 2: Customizing voice assistant characteristics.

Next, we asked participants if they would like to take part in a brainstorming activity that involved customizing such human-like characteristics for the fictional character representing the voice assistant. Given past research shows that normative materials used in design activities (e.g., low fidelity prototyping materials) may not be relatable to older adults [46, 69], to initiate ideation and discussion, we presented participants with a list of questions (on cards) as prompts. These questions were around different human-like characteristics (does your voice assistant have an age? gender? education? social class? region? accent?). We asked participants if they would like to ascribe any other trait beyond what we presented. Because we know little about what it means for a voice assistant to have these traits, we intentionally did not provide definitions for these traits to avoid bringing in researcher assumptions. Rather, our goal was to learn from participants how they perceived and interpreted these traits. We asked if/how they would or would not customize the traits. If participants asked for a definition of a trait, we reminded them that there is no right or wrong answer, and that they could share whatever they interpreted from the question or their initial thoughts upon seeing the question.

For in-person sessions, the questions were printed on cards and participants customized characteristics by answering these questions. These responses were placed on a big sheet paper which provided a space for brainstorming as well as a visual overview of their responses. We also provided other low fidelity tools such as markers, tape, glue etc., and told participants that they could draw or write any other ideas on the big sheet of paper (see Fig 2., right). However, similar to prior work with older adults [46, 95], most individuals (except P3) preferred to share their ideas verbally. We improvised the in-person sessions with the researcher writing summarized participant responses on the big sheet of paper to allow for participants to have a visual overview of their responses. For remote sessions, we used an online whiteboard with post-it notes (Google Jamboard) to provide a similar brainstorming environment that allows for free form writing or drawing (see Fig. 2, left). Remote participants also preferred to verbally share their preferences instead of typing (except for P7). For participants to see an overview of their responses, the researcher shared their Jamboard screen with summarized responses. For participants with visual impairments, the researcher verbally asked questions. Overall, in both sessions participants did not engage with the activity as we had originally anticipated (e.g., by writing or typing their response) and rather preferred to verbally share their ideas. As such, the data collected across these formats were largely similar in terms of how older adults chose to express themselves.

Fig. 2.

Brainstorming session set up: Left- Remote sessions via Zoom and Google Jamboard, each yellow colored tile had a question on specific trait. Right- in-person sessions at participants’ home.

Part 3: Reflection.

We asked participants to reflect on their experience of participating in the brainstorming activity and thinking through different human-like characteristics for their voice assistant.

3.3. Analysis and Positionality

Our data primarily included about 40 hours of video/audio recordings and observation notes, as well as the paper/whiteboard responses. For analysis we used a thematic coding approach [8]. The first author familiarized herself with the data by repeatedly reading transcripts, coding, writing and discussing memos with the larger research team. Examples of codes from this initial stage include “accustomed to,” “wisdom comes with age,” “a neutral technology,” “related/relatable to self,” and “it’s only a machine”. Three research team members sifted through and discussed the initial memos. We were struck by the frequency with which participants brought up stereotypes having to do with the different characteristics of age, gender, education, and social class (e.g., making assumptions about race and class from speech or grammar). To sensitize ourselves to the “unspoken rules” that might be operating in the background of these sessions, both in terms of the identities of the user and the characteristics they ascribed to the technology, we leaned on Collins’ matrix of domination [19]. We then sifted through data to gather the selections participants made in customizing the human-like characteristics of age, gender, education, social class, race, region and accent for their voice assistant (in 4.1). We returned to open coding the data to understand associations participants were making (RQ1) and participants’ motivations to customize (or not) (RQ2). Two researchers coded the data, discussing disagreements. New codes at this stage included “in the middle”, “power of defaults”, “invisibility of characteristics”, “universal,” and “standard.” We related the codes to each other and compared the data to understand the high level emergent themes.

Adhering to calls for reflexivity [20, 31], we describe our background to contemplate how it may have influenced this study and the findings reported [45]. Our team includes six women from varying racial, regional, and socioeconomic backgrounds. Five authors identify as non-white (i.e., South Asian and Black). At the time of planning this research, ongoing discussions in HCI and CSCW about downstream effects of ascribing gender, and our identities and experiences with systems of oppression as a result of our race, ethnicity, gender, region, and accent sensitized us to examine this open question and practice of ascribing characteristics to an anthropomorphic technology. Furthermore, our backgrounds may have influenced participant responses (e.g., interviewers’ South Asian accent could have shaped some participants’ responses pertaining to preferred accent or ethnicity). Our sensitivity to unpacking participant responses was surely impacted by the diversity of our racial and regional backgrounds as well as conversations (e.g., about history of race, gender, and language [32]) during analysis and writing.

4. Findings

First, we unpack the assumptions or stereotypes individuals associated with the characteristics, followed by describing individuals’ underlying motivations in customizing (or not customizing) these characteristics for a voice assistant.

4.1. Voice assistant characteristics and associations

To answer RQ1, “what associations (including assumptions or stereotypes) do older adults make with the characteristics of their voice assistants?” we asked participants how they would like to customize the characteristics of their voice assistants, including age, gender, education, social class, region, and accent. Table 2 shows the trait specific customization for each participant. Some participants (N=10) brought up race and ethnicity as additional characteristics of the voice assistants. We first describe the characteristics participants selected and then draw on Collins’ matrix of domination to discuss how participants’ assumptions can be understood through existing structures of power in our society.

Table 2.

Participant specific customization of voice assistant (VA) characteristics.

PID	Age	Gender	Education	Social class	Region	Accent
P1	55	Man	Didn’t assign	Probably mine	Silicon Valley	No accent
P2	Varies by role: 32, default(m) 45, butler(m) 24, trainer(f) 28, trainer(m) 50, nurse(f) 32, secretary (f)	See "age" tab for gender: male (m); female (f)	Highly educated, law, medical, degrees	Didn’t assign	Local	For butler upper class English accent, not cockney [rest] Neutral, no accent
P3	30s or 40s	Woman	More the better	Didn’t assign	Local	No accent
P4	100	No preference	Well-rounded	Didn’t assign	No defined region	Midwest USA
P5	35 to 55	Woman, don’t care	Ph.D. in life	Didn’t assign	Earth citizen	No accent
P6	60s	Man	Life education	Middle class	Answers accent	English or Irish
P7	40s	Woman	In many fields	Didn’t assign; Grammar relates with class	Non-descript	Broad-casters
P8	40 to 50	Woman; doesn’t matter	Didn’t assign	Didn’t assign	Answers accent	Standard American
P9	Default; Didn’t customize	Comfortable with default (woman)	College education	Middle class	Seattle	No accent
P10	Didn’t assign	No gender, woman by pitch	Didn’t assign	Didn’t assign	Non-descript	James Earl Jones’
P11	60s	Woman	2 years of college	Middle class	Didn’t assign	Choices: Canadian, Australian, British not Cockney
P12	Didn’t assign	Man	Didn’t assign	Didn’t assign	Didn’t assign	British male
P13	21	Doesn’t matter; likes default	Smart in all areas	Didn’t assign	Outer space	Australian
P14	Didn’t assign	Woman	Didn’t assign	Current is neutral, matches mine	Current VA feels neutral	Current VA feels neutral
P15	Didn’t assign	Used to default (woman)	Didn’t assign	Didn’t assign	Local	Local accent
P16	Middle aged 40s or 50s	Woman; doesn’t matter	Well educated	Didn’t assign	Current VA from US	Current VA No accent
P17	Middle aged	Man	Speaks from knowledge, experience	Didn’t assign	America	No accent, clear English speaking
P18	Same as me (late 70s)	Man	Very much educated	Middle class, African American	United States, Midwest	African American with old idioms
P19	50s	Woman; used to current	Educated, well read	Didn’t assign	United States	Everyday English
P20	Didn’t assign	Woman	High school	Didn’t assign	United States	Standard English
P21	Didn’t assign	Woman; used to current	Educated to understand what I ask	Didn’t assign	Didn’t assign	No accent
P22	40, middle aged	Woman	College educated, Master’s	Middle Class	United States	Neutral, English from US

4.1.1. Age.

Of those who customized the age (N=15), more than half selected middle age (40s, 50s), with some (N=4) selecting older ages (60+). Only two participants selected a young age (in 20s).

Some participants (N= 5) associated a “young age” with their current voice assistants, as P18 said, “that’s a young lady’s voice in there [Alexa].” In technology design, power shifts towards younger demographics, where most technology developers and designers are not older adults [2], which can lead many older adults to feel less represented [68]. And, resisting this design trend where technologies are typically designed by and for young individuals, many participants who customized age, assigned a middle age or older (N=11). These individuals associated positive attributes with older age, including wisdom (N=4), life experience (N=6), and maturity (N=5) [Note: counts are not mutually exclusive, some associated more than one of these attributes with age]. For example, P1 responded that they would customize a voice assistant that was: “Somebody wiser, older as it is an intelligent device. I would value more experience which generally comes with age” (P1). Prior work suggests that ageism leads to older people often being devalued [62]. Yet, some participants’ selection of older voice assistants, and association of positive attributes with older age, resists traditional ways of framing older adults, signaling the value of voice assistants that sound older.

Those who customised age (with the exception of P13) tended to avoid a younger age. They did so because they associated the younger ages of 20s or 30s with “still finding themselves” (P7), or “hasn’t lived long enough to have life experiences and to have the maturity” (P6). Others, like P17, associated a lack of authority with a younger voice: “younger person’s voice wouldn’t keep my attention, I wouldn’t take that seriously...I want that voice to speak with authority” (P17). While at first these explanations may also seem to resist ageism, ageism operates at both ends of the age spectrum. Young people are devalued through stereotypes associating youth with inexperience [27]. Only one participant, P13, positively associated a younger age with being more broad minded, customizing an age of “21, 22 you’re getting out of school so I think your mind’s broader.” However, in doing so, this participant surfaced negative stereotypes of older age, continuing: “The older you get... your circle gets smaller. Your opinions get smaller.”

While most participants (N=12) associated older ages with positive attributes, four participants associated negative stereotypes with older ages and did not assign an age that they considered to be “so old that it is senile” (P1), or “ready to retire and slow down” (P19). These participants also did not select a young age, opting for an age in the 50s.

4.1.2. Gender.

While some (N=6) selected a man’s² voice, most others chose a woman (N=15). Many of these individuals chose the current gender of their voice assistant either because they were used to it (N=5) or because gender as a characteristic was not important to them (N=4), as P8 said, “it’s not really that important. I don’t think that matters.” And because they considered gender as not important, these participants assigned a gender based on the voice as P13 said: “probably go by the sound of her voice” (P13). As such these individuals did not mention specific associations with the gender of the voice assistant. A few others chose a gender that they thought was assumed, or a default, for other characteristics that they had selected or envisioned: the name of the assistant (P6, P7; e.g., “Mr. Butler...it’s a male” (P6)) or the role of the assistant (e.g., P2 who customized multiple fictional characters said “I would prioritize the role first and then the gender and age would fall out of the role and the name would be fall out of that.”)

Five participants associated stereotypes with the gender of the voice assistant. While some were positive, most of these assumptions can be interpreted as aligned with a distribution of power favoring men as the more dominant gender. Participants associated a woman with pleasant and easygoing traits, being “more efficient and accommodating” (P3), pleasant “good to hear a woman” (P20), “women can relate to everybody better” (P11) while “men sometimes can be rigid” (P22). In the context of voice assistants, selecting a woman as an assistant by associating a woman with traits of being more submissive, accommodating, or lacking authority can be interpreted as patriarchy: a woman is viewed as a “helper” because she is subservient to men. Unconscious biases about women as subservient helpers or caregivers are well documented in prior HCI work on anthropomorphic voice [22, 26, 52].

Some participants (P2, P4, P17) invoked the other end of the axis of gendered oppression, associating authority, dominance, and power with men [94]. P17 wanted a man’s voice for information about imminent danger to associate some seriousness or gravity to the information: “I don’t associate the present [default voice assistant’s] voice with any authority. But if there was going to be information that was more serious, I would prefer the voice of authority, which I think would be a sign of a male’s voice.” P4, the participant who resisted assigning a gender, also noted the existing power structures of associating authority with men in religion and culture, saying “in some cultures, we look at God as the ultimate deciding factor in our moral and ethical decisions. And God is... portrayed most often through somebody else’s choice as a man.” Though aware of this stereotype, this participant resisted it, saying “but I don’t necessarily believe it...so...gender has no preference for me.” P2, the participant who customized for multiple roles, wanted both “male and female fitness trainer,” associating a man with “motivation by fear,” and woman with “motivation by desire.”

4.1.3. Education.

A total of 16 participants customized education. Of these, only five explicitly customized for formal or credentialed education (e.g., medical degree, Master’s degree). Collin’s matrix of domination describes how power is associated with formal credentialed education, where those without formal education are viewed as less knowledgeable (i.e., educationalism or credentialism) [19]. Valuing formal education over the lived experiences of those without formal education relates to power, particularly when there are specific barriers to education attainment (e.g., cost). Those who resisted associating credentials with education (N=11), described the voice assistant should be “smart in all areas” (P13), “broadly educated” (P5), or “well rounded” (P4). P5, who explicitly resisted credentialism, explained “there’s different ways of knowing. You can have a little education and still be so knowledgeable about so many things.” As such, resisting credentialed education for voice assistant, participants associated education with the quality of information such as, “better answers to your questions” (P3), “current and correct” information (P19), and type of information such as on “a very wide range of topics” (P21). As participants thought about education, they teased apart the nuances of what type of education they meant, going deeper than a credentialed or non-literate binary.

Beyond associating education with the type and quality of information, a few (N=2) individuals associated education with how the technology speaks (P20) or its ability to understand them (P21). P20 associated education with how the technology speaks, associating a credentialed high school education³ with being able to speak “standard English”: “I think that most important thing is ability to enunciate words and speak standard English...you probably would need at least a high school education because that’s what kind of stressed in those early years.” This notion of there being a “standard English” and times where it invoked oppressive stereotypes, comes up again in 4.1.6. While P20 associated education with how the technology speaks, P21 associated education with the technology’s ability to understand her, saying that her voice assistant “needs to be educated, it needs to know and needs to understand exactly what you’re saying... And not have to say, well, “I don’t know that,”... I sort of get a lot of that (with the current voice assistant).” P21’s frustration with regards to her voice assistant not recognizing or understanding her commands, resonates prior work with Black older adults and youth who maybe forced to codeswitch, i.e., change the way they talk to the voice assistant [11, 48, 87].

4.1.4. Race or Ethnicity.

Many (N=12) did not discuss race or ethnicity of the voice assistant or found this characteristic meaningless as P10 noted, “it doesn’t have an ethnicity.” Even though race or ethnicity was not included in our questionnaire, some (N=6 identified as white, and N=4 identified as Black) brought it up without being prompted. Often, these participants associated the current voice assistant with being “Caucasian” or “white”, aligning with prior work [11, 87]. For instance, when asked to think about how the fictional character representing her voice assistant looked, P19 described her current Alexa as: “Now the one that I have, she’s not African American. She’s a white woman.” When asked why she associates her current voice assistant to being white, P19 (who identified as a Black woman) described:

Just sounds a lot like most of the white women that I know. Because a lot of times, you can tell. Well, I can, most of the time, tell if I’m speaking with an African American or Caucasian. I don’t know if you can, but I usually can tell because of, you know, who I talked to the most. And that’s just how she sounds to me[...] I just can’t explain how we know that. But most of us tend to know that just, just know it. I can’t explain it.

Similarly, P2 (a white man) stated, “I can tell when I’m talking to an African-American person on the phone, because of their syntax, because of the way they use their voice.” Both P19 and P2 shared how they associate race to the sound of people’s voices and how they apply those same assumptions to voice assistants, leading them to believe that their current voice assistants are white. Interestingly, all four Black participants (P17, P18, P19, P22) who perceived current voice assistants to be white wanted the option to “give her [the voice assistant] an African American voice” (P19). White participants, on the other hand, although acknowledged an encoded race, did not feel a need to customize or change it, and a few even described how they felt current voice assistants were targeted towards them. One participant, P3, describes how she feels that her racial and education identities (i.e., white with a doctorate) is the target audience and as such, voice assistants are being marketed towards her people with her set of characteristics:

We’re the demographic for this. That this [voice assistant] is geared to white, educated, whatever, professionals and it’s possible that somebody in different circumstance would listen to it and go, ‘No, not that voice again!’

Being the intended demographic for a widely available, useful, commercial technology is one way that power is distributed in our society to benefit those with some characteristics (“white, educated, professionals”) more than others.

After reflecting on the racial encoding in current voice assistants, one participant (P17) described her discomfort in continuing to use this technology that did not allow her to customize the voice to reflect her racial identity:

If I have to talk to, what she’s a white woman, I don’t want to talk to that...It’s [this activity] made me think about this device and other devices and, how they come up with voices. It’s not that I don’t like it. It’s just, it has sort of given me a nudge, like, think about that. So that may or may not be good. I might want to just disconnect this thing and give it to somebody, throw it away. Until they give me a choice as to what I would like to have in my home... Because older people, and especially older Black people don’t have a voice.

This statement highlights the intersectional experiences of being Black, a woman, and an older adult and how there is an invisibility of race that is embedded in the design of voice assistants that may not be explicit but is present—excluding those with non-dominant identities. Though she had not previously reflected on the characteristics of voice assistants or her preferences, she concluded that her participation in our study and usage of voice assistants indicated her approval of this exclusion, which then made her question her usage of such systems.

Collectively, these statements from both white and Black participants suggest that non-dominant racial groups (i.e., those with non-white racial identities) are not centered in the design of voice assistants and the need for choice, aligning with prior research [11, 87]. Although an association between race (or ethnicity) and voice of the technology was mentioned by a majority of Black or African American participants (N=4/6) as compared to white participants (N=6/16), not all Black participants (P20, P21) associated race or ethnicity with the voice (“I don’t see Alexa as being Black or white” P21), aligning with prior work suggesting that Black older adults are not a monolith, but instead have differing viewpoints with some preferring not to factor race into their voice assistants [11].

4.1.5. Social class.

Most (N=16) did not customize class, explaining how a machine does not need to have a social class (discussed in depth in 4.2). Those who did customize, selected a class that was same as their own. Only three participants (P1, P7, P18) verbalized specific associations with social class. P1 and P7 associated social class with the way the voice assistant speaks. P1 went back and forth between not assigning and assigning a class as “probably mine,” (which she had earlier mentioned as “upper, white, middle class”), and described her perception of how the social class relates to the way a voice assistant speaks: “I might respond positively to a voice that speaks clear, intelligent English, so that makes me think of my social class. If it [the voice assistant] came out with everything that it was talking about in a rap type style, I would probably turn it off.” P7, although did not explicitly assign a social class, she drew an association between social class and speech saying, “sometimes grammar is an indication” and “if somebody were speaking with poor grammar, you might make an assumption about them.” These descriptions illustrate how social class hierarchy is implicitly associated with characteristics that have historically been linked to whiteness, with other dialects, vernaculars, and/or non-American accents associated with a lower social class [38, 102]. P1’s statement, for example, refers to “rap type style” speech—without specifically mentioning a particular race, she is alluding to one of the most popular types of music and art that was created by Black Americans [29, 70]. These quotes signal how race is being confounded with social class, which is a result of a history of racist policies that have led to less opportunities for socioeconomic advancement for non-dominant racial groups [57].

While P1 and P7 drew associations between social class and how one speaks, P18 associated social class with education attainment. He described a preference for the voice assistant to be “middle class” to have a level of education that could “help me with research that I might want to do while I’m writing.” Upon further reflection, P18 raised questions about the notion of social class due to the history of confounding race with social class (as were made in the responses we described above about intellect, race, and class). P18 asked “when you say social class, help me understand what you define as social class?” Asking P18 to reflect on what were his first thoughts when he saw the question, he responded:

I look at this [question about] social class, I’m going to say, “Are you asking me about race?” ... Should the old man [voice assistant] be white?” Because historically, African Americans have been led to believe that white folks are smarter. Should this guy be white, just from a stereotype that’s been created?

As a Black American man, P18 deeply understands the context of class in the U.S., where Black people are many a time erroneously labeled as being of a lower class, or socioeconomic status, due to their race [19]. These false narratives have perpetuated stereotypes about Black people by pointing to certain conditions (that many times are inaccurate) without acknowledging the structures in place that were intentionally created to facilitate and maintain the oppression of Black people in the U.S. (e.g., redlining, systemic disinvestment, unfair sentencing) [4, 19, 57, 73]. Not only does P18 rightfully resist the notion of working with assumptions of racial superiority and intellect, but also, through his questioning, brings to light America’s history of treating racialized classism as an invisible characteristic that is assumed based on other features, which tends to place characteristics of non-white people towards the bottom.

4.1.6. Region and Accent.

For the questions around region and accent, we saw a wide range of responses. Of those who customized the region (N=17), individuals assigned a region that matched the region where they were using the voice assistant (N=9), some assigned a non-defined region (N=5, e.g., outer space, P13; nondescript location, P10), and a few (N=2) ascribed a region pertaining to the parent company. Two participants responded preferred accent when asked about region. For accent, several (N= 10) preferred what they considered as no accent, standard or everyday English, or neutral (further elaborated below). Others explicitly mentioned a region that they associated with an accent such as Midwestern, standard American, British, Irish, Canadian, or Australian accent (N= 8). A few customized accent describing how they would like the voice assistant to speak (N=3; e.g., “broadcaster”, P7; “African American”, P18; celebrity voice, P10). P18, for example, wanted his voice assistant to speak with a specific dialect and use familiar phrases: “I want him to speak American, African American...There’s a lot of slang words or idioms that we use years ago that bring back memories...an old dude that has access to those idioms would be great.”

In this case, P18, with intersectional non-dominant identities, a Black man and an older adult, preferred to have a voice assistant that relates to both his age and his racial identity through its use of cultural vernacular that is unique to African Americans [61]. His vision was to have a voice assistant whose speech sounds like it comes from a from a particular generation and that uses vernacular unique to African Americans, specifically men—stands as another example for including diverse users to make space for these conversations in design of voice assistants [11].

Participants associated region with the type of information and accent. A few individuals (N=3) customized region to be the local region so that the technology can provide local information. Many others (N=11) associated region with accent or how they wanted the technology to speak, with a few (N=3) even answering their preferred accent upon seeing the question of region. As noted in Table 2, many selected an accent from the United States, and/or something that they considered as “no accent” (P16). For example, P22 selected “United States” as the region to have an accent that is “neutral. Just the English you know,” P7 assigned a “nondescript” region to have no accent and speak like a “broadcaster,” and P8 chose the region of “California,” that she associated with a “standard American accent.”

At times, when participants described their preference in selecting a particular accent, or how the technology speaks, (i.e., British, non-Cockney accent or standard, neutral English), we noted power structures associated with class, region, as well as the global political economy. Two US-based participants who wanted a British accent (P2, P11), did not want a “Cockney” accent: “not Cockney. I don’t understand that, but just... a nice old British accent.” (P11). A ‘Cockney’ accent is associated with a working class east London dialect [39]. These descriptions reinforce participants’ presuppositions of dialects or accents being associated with a social class. In another example, a quote from P20 illustrates encoded stereotypes of region and ethnicity associated with supposedly “standard,” “neutral,” or accent-free English:

I think that it should just be standard English. [...]sometimes when I have to call AT&T for technical support, and if the technician has a deep accent is very frustrating to me. And sometimes I’ll say I need to speak to an American side agent. Because I don’t need to have to figure out what you’re saying when I’m calling you to get an answer to what I need. (P20)

We can infer that P20 is referring to call center workers from Global South as having the “deep accent [that] is very frustrating.” Her quote can be contextualized by noting the invisible power structures associated with the tiered global political economy, where the West outsources call center labor to ghost workers of India, Bangladesh, Philippines, and other places of the Global South [91]. These call centers are aware of preferences of people like P20 for a specific kind of speech as “standard” or “neutral”: to have a “neutralized pronunciation for Wisconsin or Durham” [23], workers are trained to give up their own accent-based identity and adopt British or American accents [23]. And preferring standard or neutral English, suggests these participants predispositions towards the notion that certain speech forms, shaped by power structures and linked to ‘dominant’ regions and accents, are more preferable or acceptable than others.

4.2. Motivations behind customizing or not customizing voice assistant characteristics

Answering RQ2, “what were older adults’ motivations for selecting specific characteristics for their voice assistants?”, we describe the three motivations participants shared. Participants wanted to create a voice assistant that would be: 1) relatable to themselves, 2) relatable or acceptable to a broad audience, c) consistent with their view of the voice assistant as a piece of technology (whether by customizing human-like characteristics appropriate for a piece of technology or refusing to assign a particular characteristic). Analyzing these motivations enables us to further contextualize some of the customizations noted in the prior section, such as selecting a region associated with the parent company or resistance towards customizing certain characteristics. We continue to find Collins’ framework useful for unpacking how many of these customizations and motivations to customize, relate to structures of power and domination.

4.2.1. Relatable to me.

A common motivation that participants described for selecting particular characteristics was to make a voice assistant that would be more relatable to them (e.g., selecting social class as middle class to “have more in common with me,” P6). Another participant (P18) who “preferred to deal with somebody a little older, male” said: “I would prefer to deal with somebody a little older, male... not diminishing that in any way. But again, if you say, [...] what my ideal dude box look like, this is what it will look like...old man like me, who knows more than me.” Within the group of participants who were motivated to create a technology that was relatable to them, there were two distinct reasons for doing so. The first reason was to create a technology that could provide relatable content. Second, some wanted to create a technology that had characteristics in common with what they were used to or accustomed to in their life and homes—which sometimes included the default options of their current voice assistant.

Regarding this first reason, participants wanted a technology that would give more relatable content. Individuals described how current content given by Alexa did not feel related to them. For example, in remembering a question he had asked Alexa, P10 decribed how he felt hearing the response “it’s a pretty ridiculous answer.” He thought this response may have been directed towards younger age groups, saying that his grandchildren”might find that funny.” P5 similarly shared that her current voice assistant could not provide relevant content that she was looking for, i.e., “questions about movies from 30 years ago.” Finding the current search results, music, or jokes recommended by Alexa as not relevant, P11 customized the age of his voice assistant to be in 60’s and described how he saw an “older,” “more mature” voice assistant could give him more age appropriate content including jokes, music, and online information:

Researcher: Is there any reason why you selected sixties [as age]?

P11: A lot of the things that I ask for, it gets to the point where let’s say, [I ask] Alexa I want to go on vacation. Alexa is not going to tell somebody in their sixties, there’s a great skydiving place...you say, “Alexa, tell me a joke.” And it’s often a convoluted bunch of words that’s supposed to elicit some sensation but I don’t know what it is. It’s often not humorous. And it’s not really directed towards anybody but my grandchildren[...] they’re just silly hoops of words[...] So my personal assistant, that’s living in my house. I don’t want it to be a 30 something. I want it to be more mature. [...] 60s, 70s, no more. 60s is fine. So if you ask it to play a song, it’ll play a song that I might like.

P11, a user from a non-dominant age group, described how the content provided by current voice assistants did not align with his preferences as an older individual, and rather seemed to be geared towards younger people.

The second reason for creating a relatable voice assistant was individuals’ lived experiences over time, which had shaped what they had become used to or accustomed with in their everyday lives, leading them to select specific traits. Participants drew upon their lived experiences associated with what felt more familiar or what they were used to in their home. Having recently lost her husband of decades, P1 customized her voice assistant as a man: “I’ve been living with somebody for 40 years, and now I’m living alone, so it’s nice to have a male voice in the house.” P17, a Black woman, wanted “to hear the voice that’s associated with a Black male” saying, “because of my life experiences, I’ve never heard the voice of a white male that inspires me, that gives me that much confidence...whenever I hear the voice of a white male, you have to look beyond the words to be sure that they’re meaning what they’re saying.” While not specific about what led to this lack of confidence, P17’s life experiences likely go beyond interpersonal interactions to other forms of oppression including the history of white men being in decision-making positions (majority of leaders in government, industry, academia) as compared to other groups that are not as often in positions of power. Collins describes specifically how Black women’s lived experiences are intersectional, facing gendered racism in ways that may differ from other groups, and pushes us to value the lived experiences of Black women as valid “ways of knowing” as well as the ways in which they actively resist oppression [19]. Additionally, P17’s account speaks to how trust in voice assistants relates to the anthropomorphic nature of the technology [99]. P17’s account of distrusting the voice of a white male suggests how the intersectional identities encoded in the voice could impact trust in them.

A second reason for choosing something relatable was, over the time of using the voice assistant, participants had become used to the default option the technology came with. While a few participants described being used to the current age (P9), region and accent (P16), “to accept things the way they are without thinking about improvement” (P9), others described being used to the current gender associated with the voice, as P19 says:

I’d probably use the lady again. It’s probably because I’m accustomed to her. At this point, had she been a man to begin with maybe I would go with it, but I’m just used to her voice.

In choosing the default option, participants are exhibiting what has been termed as “status quo bias” in HCI, i.e., select the default option rather than taking the time to consider the alternatives [14]. We discuss in further detail (in 4.2.3) how invisibility of defaults relates to power in our society.

4.2.2. Broadly relatable for wide appeal.

Above we described how some characteristics were customized to make the voice assistant more relatable to participants based on what they were used to, or what they related to in terms of content. Another motivation that participants described was to customize characteristics so that a voice assistant would be relatable to a broad group of users, in order to have the widest appeal or “would be acceptable to most number of people” (P9). This was done in two ways: participants chose a perceived universal design that they thought would be acceptable to people from different demographic groups, with a few individuals allowing for customization of characteristics.

To customize a universal assistant that would appeal to people from many different backgrounds, individuals chose what they perceived as universally appealing. This included woman as a gender due to the assumption that this, “works for most people [...] I think women can relate to everybody better than men can” (P11), or choosing a region or accent that was not associated with noticeable regional differences so “you can’t tell where they came from” (P7), or by drawing upon their assumptions of what they perceived as standard, neutral, and no accent (as noted in 4.1.6).

Others selected traits that they thought would relate to people from many demographics because the traits existed in the middle. P5 described how a voice assistant could provide content relatable to both younger and older generations by having an age range between 35–55 years: “because it needs to be young enough to be able to find the musicians of today and...old enough to have some experience so they can understand my question correctly.” Other examples included:

• a social class that they perceived was in the middle i.e., middle class, so that “they will be sensitive to both the lower class and the higher class” (P22);

• an age that was in the middle of life, “middle age, just kind of halfway, but not too young, not too old... 20s or 30s, they’re young and then 60s and 70s, or older and 80s, we won’t even talk about where I am” (P19);

• an education they considered to be somewhere in the middle (“college education,” P9; “two years of college at most,” P11) because “you don’t want to sound like you’re talking bad to people who are less literate than you are, and yet you don’t want to show them like you’re not educated. There’s a happy medium. I wouldn’t go above two years of college at most. That way you were smart enough to get into school” (P11).

Participants had positive intentions in creating these “universal” voice assistants, as P5 described “I prefer a universal, almost like an extra terrestrial in the sense that they’re not coming with any biases by region... An earth citizen.” This quote suggests how individuals, like P5, understood that people with certain characteristics are being subjected to bias and oppression. They wanted the voice assistant they were designing to actively counter these biases by being “universal,” therefore avoiding the current distributions of power leaning towards dominant groups. While the intention is laudable, characteristics that are not “in the middle” remain invisible. This approach can be seen as similar to the heavily criticized attempt towards “neutrality” in computing [3, 35, 78, 89, 90, 105, 110]. Designing for “the middle” does not allow us to acknowledge, let alone value, the presence or the insights of non-dominant groups at the margins.

The second way that a few participants (P4, P9) sought to customize a broadly relatable voice assistant for wide appeal was by allowing for personalization. P4 described how assigning a specific social class, gender, or region was “limiting”, and rather she wanted something that represents “the world community and a range of social class and options to meet with other genders or people who are not concerned about gender’ (P4). Another participant, P19, expressed that allowing users to personalize characteristics associated with the voice could be a way to maximize representation for users from non-dominant groups. P19 describes that that there could be “several voices” that one can choose from “like an African American voice, an Indian voice, and Hispanic voice...Which voice would you like, and you would have a choice.”

4.2.3. But it’s a machine, not a person!

A third motivation for participants’ decisions around customization was to engage in the activity in a way that was consistent with their views of the voice assistant as merely a piece of technology. Participants did this in two ways: customizing characteristics that would be appropriate for a machine, and by refusing to assign a trait for a particular characteristic (‘did not assign’ in Table 2).

Aligning with the first way of customizing that was consistent with viewing the voice assistant as a machine, individuals selected some characteristics that they thought would be appropriate for a piece of technology. This involved assigning a region that related to the region of the company manufacturing the assistant (“Silicon Valley,” P1; “Seattle,” P9), or “outer space,” (P13), “because it’s an odd piece of equipment. It’s a strange shape, nothing like on this planet went with anybody who talks[...] It’s not human, it’s not anything normal” (P13). For education, being an expert “in many fields” (P7), or “smart in all areas” (P13), was seen appropriate for technology, as it was something a human cannot be: “you’re asking it such a vast array of questions for so many different fields...a human cannot be an expert in everything... I’m going to write that because it’s not human, [types, educated] ‘in many fields”’ (P7).

The second way participants’ customized, that was consistent with a view of voice assistants as piece of technology, involved refusing to consider or customize human-like traits for the voice assistant. With other motivations described previously, participants declined to customize social class, accent or region in order to design a broadly relatable voice assistant (in 4.2.2). Here, participants avoided customizing these characteristics, but because it did not make sense to assign traits to a “product” (P1), which was decidedly “not a human” (P12). P8 encapsulated a sentiment offered by several participants who avoided customizing traits, which was that she found it difficult to “buy into the concept of giving human attributes to a device.” Therefore, responding to the question on social class of her voice assistant, P8 said: “I think the answer to this is ‘I don’t think so’. We keep trying to give or ask us to give human attributes to a device and you have to buy into that before you can answer these very seriously.” Similar to P8 refusing to customize the social class, others did not customize the region because “they don’t come from any particular region. It’s a thing” (P12), or accent, because “I don’t want to be caught up in thinking there’s a person in there” and rather “want this to be a behind the scenes support” (P5).

It seems that in using these voice assistants for transactional interactions such as finding quick information, checking the weather, or setting a timer, the voice represented merely a mode of interaction, “a conduit between me and the internet” (P15). The presence of human-like characteristics embedded in voice assistants appears to have become invisible for participants such as P10, who noted “It’s a thing. It’s not something I give any thought to. It has no impact on my use of it. These things are not important...It’s no different than my stove, or my TV, or anything else around the house.”

These selections, where participants did not customize the default characteristics associated with the voice assistant–that currently encode traits of dominant social groups—show how defaults can become invisible to users. The ideation activity appears to have helped at least one participant realize how human-like characteristics are encoded in voice assistants. P15, a white man, did not assign or used “default” or “local” options for all traits, explained, “I’ve never known the gender, the social class, the age, the accent, they’re just not things that I’ve ever thought about in any way. I don’t want to think about it now.” However, P15 later reflected on how these characteristics were indeed encoded in his interactions with the technology:

I must have subconsciously put it somewhere... when I now say thank you... it comes back and says “Cheers, mate” which is obviously a kind of age or education or social class type response. And it does stand out so I must have classified subconsciously in some way to notice that cheers mate didn’t sound quite right.

As such, P15, upon reflecting on some of the responses from his current voice assistant acknowledged how the technology currently encoded several traits, which subconsciously had become invisible to him.

5. Discussion

By inviting older adults to customize traits of anthropomorphic voice assistants, our findings reveal the different assumptions and stereotypes that individuals associated with the characteristics of age, gender, education, race, social class, region, and accent of the technology (RQ1). By examining participants’ motivations for customizing or not customizing these characteristics (RQ2), we learned that individuals attempted to make the technology relatable to them, broadly relatable for wide appeal, or making choices to align with a view of the voice assistant as not human-like. Corroborating and complementing prior work, we note how an anthropomorphic voice not only encodes race (white) and gender (woman)—as identified in prior work (e.g., [87, 100]), but also age (young) and class (intertwined with speech, education, and racialized classism). Collins’ matrix of domination helped unpack how participants’ customizations aligned with existing distributions of power between dominant and non-dominant groups or resisted them. Applying this Black feminist theoretical lens to our findings helped understand how users become accustomed to default characteristics that come with a technology, which then contributes to a preference for that trait — thereby providing insights on how these traits can become invisible through use, while still impacting the ways these technologies can shape our society.

Below we discuss how our work contributes to design of AI-based voice assistants, in particular calling out how bias and assumptions gets encoded through our design processes of ascribing human-like attributes to technology— impacting older adults (5.1), as well as other non-dominant groups (5.2). We discuss how older adults’ preferences continue to be unaccounted for in design of voice technologies (5.1), contextualize how our findings contribute to research examining older adults’ participation in design activities (5.3), and on engaging end users in designing anthropomorphic technologies that pose ontological confusions (5.4).

5.1. How Well are Older Adults’ Preferences Accounted for as Users of Voice Assistants?

We contribute to ongoing research on how AI-based voice assistants can be designed to better support older adults [10, 85, 106, 114]. Our findings speak to prior studies that noted how some older adults may not find value associated with using voice assistants. Trajkova et al. [113] found that older adults in their study abandoned using these assistants due to a lack of perceived value. In particular, participants in their study were less likely to use features such as jokes or finding general information. Individuals in our study also shared a similar sentiment noting how content provided by current voice assistants including general online information, music, and, jokes, did not align with their preferences. This opens questions on why content provided by these technologies does not cater to preferences of older individuals. More specifically, where and why in our design and development processes are we failing to include older adults’ content preferences from these technologies?

Our work surfaces how some unrelatable content could potentially be linked to our design practice of creating personas for voice assistants. Content such as jokes or stories are typically written around the human-like traits ascribed to the voice assistant (e.g., personality, age, gender) [1, 54]. Given that commodity voice assistants are modeled around a “young” persona [84], it is unsurprising that our participants perceived the corresponding content to be youthful and unrelatable, or as P11 said “they’re just silly.” And, instead these participants saw a way around to ascribe older ages for mature and more relatable content (as noted in 4.1.1 and 4.2.1). While ascribing older age as a persona trait of an anthropomorphic technology can work for content such as jokes or stories, for other types of content such as music recommendations or recommendations for things to do—types of content that our particpants noted being unrelatable— solely thinking about persona age is not enough. Searching for online information on voice assistants involves AI-based recommender systems [25], which can have a popularity bias, i.e., recommend already popular sources by learning from the existing usage patterns [17], which typically constitutes patterns of dominant user groups [77, 96]. As such, even if content that older adults may find relevant exists, it may not get recommended, and is likely to be down weighted by the recommender systems optimized for overall popularity. This is because, older adults, continue to be the age group with least internet use (although their numbers are growing) [36]. As such, it is likely that the content recommended may not align with older adults’ preferences. Explicitly including learnings from how older adults’ search for online information through voice, such as, the types of queries they ask, the responses they expect (we see initiatives in this regard [9]), and accounting for these in recommender systems can help account for older adults’ salient preferences.

5.2. Systemic Power Structures Encoded in Anthropomorphized Technologies

Informed by Collins’ matrix of domination, we posit that the design of human-like characteristics for technological systems must be discussed using an intersectional lens [20, 32]. As such, we must examine existing structures and their impact on the design of anthropomorphic technologies. This will enable us to unpack how assumptions related to voice assistant characteristics can be understood through the history of oppression, power, and domination.

Following Erete et al.’s examples of applying an intersectional analysis of power in HCI [32], we first acknowledge the power differentials in creating technologies. Large companies have the power to create an artificial human-like voice that may not value or reflect the diversity of our society. Employees including designers and developers determine whose voices to embed in voice assistants. Collins’ matrix of domination describes how certain identities (e.g., credentialed, white, young, men, anglophones) have a history of being in power and dominating those without these characteristics. Participants registered that many of the voice assistants seemed to be geared towards “white, educated” populations, and some even associated class and race with the voice assistant. That the “default” voice assistant conveys these characteristics, shows how attempting to design by ignoring or not assigning traits such as class or race may not possible. Rather, it reveals that power and oppression operate invisibly at the expense of non-dominant groups [35, 78, 89]. By ignoring and not designing for these traits, we are not only designing technology with a race and color-blind approach that has been heavily criticized in HCI [80], but also, insidiously designing by ascribing certain traits that may end up being a proxy for class, race, and ethnicity, such as, which groups are most likely to have education from “elite research university” — characteristics ascribed to a broadly used commercial voice assistant [44, 103].

The conceptual glue that binds these systems of power are stereotypes about non-dominant groups (e.g., devalued lived experiences of those with less formal education) and the institutions and structures that support them. Technology companies are incentivized by capitalist structures where financial motives lead to centering the preferences of the majority (often dominant) users [121]. Most participants preferred to hear a woman’s voice and some from dominant groups preferred to hear from voice assistants with traits similar to theirs. The notion that an “ideal” voice assistant would convey dominant characteristics at expense of non-dominant groups, points not only to sexism and racism but also the intersection of the two (i.e., gendered racism)—in addition to classism. Moreover, negative stereotypes about certain populations are implicitly and explicitly perpetuated through these technologies when characteristics of dominant groups are embedded. For example, some participants stated that current voice assistants are targeted towards them because they are an ideal market (“we’re the demographic for this” -P3), but believed that those from non-dominant groups may desire a voice with different characteristics. However, this illustrates how those from dominant identities are centered, not considering how they may enjoy and learn from a voice assistant with non-dominant characteristics.

One design implication in this space is to attend to and design for points of resistance. We reflect on three acts of resistance in our findings. Some participants stated that they would consider not using voice assistants given the lack of choice—a powerful act of resistance in computing [87, 89]. Participants also suggested non-dominant voice assistant characteristics based on their life experiences, which are not reflected in products currently on the market. Others acknowledged bias in voice assistants, reflecting an opportunity for resistance — but their proposed solution of creating universally acceptable characteristics kept non-dominant characteristics at the margins. The latter raises questions on the notion of universally acceptable characteristics and relates to design of defaults. Our findings surface the immense power of defaults and the potential harm they can cause by being perceived as neutral or becoming invisible. This opens up areas for future work such as how should we choose default traits for anthropomorphic technologies? Should defaults be based off dominant groups? Should defaults change over time, context or user group? How might users react if the voice assistant changed periodically or did not match with their own identities? How does the interplay of different traits impact users’ trust with the technology? What are opportunities and consequences of allowing end users to customize traits of their voice assistants? Which traits should users be allowed to customize? Should we reduce the human-likeness of anthropomorphic voices? Attending to such questions will inform the design decisions we make today which will have far reaching societal ramifications in the future.

5.3. Engaging Older Adults in Designing Off-the-shelf Technologies

Understanding how to support older adults’ engagement in design and brainstorming activities is a topic of interest in HCI and CSCW. Some of our findings speak to this past work which notes instances of older adults not designing for their own needs [63, 83]. In some instances, when asked to brainstorm technologies for aging, instead of thinking about their own needs as one might expect, individuals brainstormed technologies for other older adults by drawing upon stereotypical views of aging [83]. In our sessions, older adults also thought about others. But unlike past work where they thought of other older adults [83], our participants thought about a broad group of others by designing something that they thought could be universal, acceptable to older and younger populations (by assigning an age in the middle), people from lower and upper class (by assigning a middle class), or by allowing for personalization. These findings suggest that, as we change the technology in focus from aging specific to something that can be used by a broad group of users, older adults may still continue to think of others, but, their end-user may shift from a stereotypical older adult to assumptions about what might be desired by a broad group of others. This may allow researchers to achieve the kinds of idea-generating dynamics noted in Rogers et al.’s work with older adults and the MaKey MaKey toolkit [93] where there may be value in including older individuals in brainstorming ideas not just for technologies for aging alone, but also for a wider group of audience.

While we did not see internalized ageism appear in the ways described in past studies [83], we did see stereotypes associated with age come into the brainstorming activity in different ways. These included negative associations, but also positive ones. Some participants worked with stereotypes of wisdom, maturity and experience associating them with older ages. Wisdom, particularly in terms of coping skills gained through experience, emerged in Lee and Riek’s study as a quality that older adults wanted researchers to focus on, as an alternative to the more deficit-oriented framings of health issues that they “needed help” with [64]. While this past work focused on implications for human-robot interaction, our findings indicate that wisdom as a design concept [64] is also a fruitful direction for researchers studying conversational interfaces and even informational retrieval. But, participants also brought up harmful associations with age such as “senility,” particularly around advanced age (assistants that were “too old”). In noting that older adults have internalized harmful associations with their demographic and those older than them, we arrive by another route to the same path as Rincón et al. in asking whether user representation alone can solve design problems [92].

5.4. Designing Anthropomorphic Technologies Posing Ontological Confusions

A phenomenon we found surprising was the discomfort that some participants experienced when asked to customize human-like traits for voice assistants. CSCW literature examining users’ ontological perception of voice assistants, i.e., do people think of voice assistants as human-like or as machines? [37, 82, 99], helps further contextualize these findings. Corroborating with this body of work, we find that users may continue to experience ontological confusions even when designing these technologies. Even participants who customized characteristics, at times, continued referring to the voice assistant as a piece of machine (e.g., P18’s “ideal dude box.”). The openness to ascribing human traits, while fluidly moving to technological descriptions, indicate a perception of a voice assistant that may not neatly fit either ontological category [82]. Another instance where a discomfort was notable was when some individuals resisted customizing human-like traits, either by assigning a trait that would be appropriate for a machine (e.g., region for a machine as region of the parent company) or by not customizing the trait altogether. These individuals perhaps deliberately put an effort to keep the categories of ‘human’ and ‘machine’ distinct, as a way to minimize the ontological confusion [99].

For researchers seeking to involve end users in designing anthropomorphic technologies posing ontological confusions, we discuss some strategies we found helpful. As noted in our work, humanlike traits associated with voice assistants can become invisible to end users through use over time. Therefore, careful scaffolding of the activity to first make these traits visible is necessary. Our approach of familiarizing participants with the concept of ascribing human-like characteristics by showing them examples of current voice assistants with these traits was helpful. Although one might anticipate that familiarizing participants could lead to idea fixation, most participants’ customizations differed significantly from the examples we had shown, suggesting that this approach may indeed facilitate meaningful engagement in brainstorming desired characteristics. We also found it useful to give participants the space to express the ontological confusion, by letting periods of silence during the sessions extend longer than usual.

6. Limitations and Future Work

Acknowledging the diversity of older adults as a user group [115], our sample largely reflects perspectives of older adults who would be considered highly educated, living in urban areas, from higher socio-economic groups, and technologically savvy. Further, participants were predominantly living in the the United States. As voice assistants begin to speak with several regional accents or dialects [53], it raises questions on what local and regional dominant norms could be encoded with the regional voice of the technology. There is a need for future work that brings in diverse voices from non-Western contexts. Our brainstorming sessions have limitations. Working with individuals who had used the technology for some time made it easier to brainstorm [83]. However, it may have led to idea fixation, where participants developed a preference for their current voice assistants’ characteristics (see 4.2.1). Second, our approach to ask participants about specific traits has trade-offs. Because the brainstorming was focused on an uncommon, and to an extent, an invisible topic— technology having human-like traits— providing some traits from prior work as examples was helpful to scaffold the activity. Moreover, this scaffolding allowed us get insights on individuals’ perceptions of how each of these traits relate to/are encoded in an anthropomorphic voice and their preferences for it. We may have missed these insights had we not specifically asked these traits, as individuals may not consciously realize these traits exist in an artificial voice or social desirability bias [43] could cause hesitance in bringing up/discussing them. However, a downside of presenting these specific traits was that it oriented most participants to think only about these traits. Although participants were encouraged to bring up new traits, apart from some who brought up race and ethnicity, most did not bring up new traits. Acknowledging how our findings are influenced by the specific traits that was presented to participants, and that the traits we presented were not exhaustive, and other qualities maybe associated with a human-like voice, such as, sexuality [109], there is an opportunity for future work to engage end-users in more open-ended activities to understand their perceptions and preferences for other human-like traits for voice assistants.

Finally, we reflect on our choice of working with a particular population, i.e., older adults. In doing so, our work aligns with prior work on voice assistants that have focused on specific user groups (e.g., children [40, 87]), or Black older adults [48]). While some of our findings and discussion specifically address how older adults’ preferences continue to be unaccounted for in the design and development of AI-enabled voice assistants, our findings should not be limited to one age demographic. Prior CSCW work has examined particular groups, such as, older adults, and yet this work is influencing the way that we think about ontologies and voice assistants more broadly [82]. We share a sentiment with past work [82] noting the harmful tendency of resigning older adults to a separate category. That Black feminist theory [19, 20] related to and shed light on our empirical data, reinforced our belief that knowledge derived by working with older adult may also be applicable for others groups as well.

7. Conclusion

By conducting ideation sessions with older adults, focused on customizing human-like characteristics of their voice assistants (e.g., age, gender, education, social class), we unpack the associations, assumptions and stereotypes individuals attribute to these characteristics, such as, associating older age with wisdom, woman as easygoing and accommodating, education with the type and quality of information, or style of speech with income level, class, and race. Drawing upon Collins’ matrix of domination, we illustrate how existing distribution of power between dominant and non-dominant groups plays a role encoding and perpetuating these associations and assumptions. We also examined older adults’ motivations for customizing or not-customizing voice assistant characteristics, and found that individuals attempted to make the technology relatable to them, broadly relatable for wide appeal, or made selections to align with a view of the voice assistant as not human-like. We discuss how AI-based voice assistant encode age-bias in content recommendation through algorithms and our design practices. Surfacing the invisible, yet systemic power structures encoded in anthropomorphic technologies, we call for an intersectional approach to design of these technologies.

CCS Concepts:

• Human-centered computing → Human computer interaction (HCI).