Producing shifting personhood: How designers anthropomorphize artificial intelligence

Abstract

Anthropomorphized artificial intelligence has become increasingly ingrained in the fabric of everyday life, yet sociologists know little about how it is produced. Drawing upon in-depth interviews with 21 voice assistant (VA) designers at major technology companies, we examine how designers of VA devices think about how to design anthropomorphization in order to produce a specific user experience. Extending Erving Goffman's theory of personhood, we find that designers do not always seek to construct devices that convincingly mimic human interaction, even when this is technologically possible in certain ways. Instead, designers selectively build social and personal identities into VA devices as they seek to design devices that are helpful and engaging, broadly appealing, and representative of company values. As a result, devices express what we call shifting personhood, whereby anthropomorphic qualities are conveyed to varying degrees across different moments of interaction.

Introduction

Artificial intelligence (AI) has increasingly pervaded every sphere of social life, from the domestic realm to the workplace. In particular, people often interact with “anthropomorphized” AI—technologies intentionally designed to mimic aspects of human interaction, such as speech and expressions of emotions—from when they operate self-driving cars to when they communicate with customer service chatbots. While people are aware that they are interacting with nonhuman systems, they nonetheless anthropomorphize, or ascribe human-like qualities to these entities. For example, Amazon named its warehouse robots after Muppet characters (Abnet, 2021) and NASA engineers held a tearful funeral for a rover whose battery died in space (Vertesi, 2015). However, social scientists know little about how people anthropomorphize these technologies and even less about how the designers of these technologies conceive of anthropomorphization. While consumers attribute human-like qualities to anthropomorphized AI, designers make decisions about how these technologies should express human-like qualities in their interactions with consumers. In this article, we explore how designers think about how to design anthropomorphic qualities in AI in order to produce a particular user experience.

We address this question using the case of voice assistant (VA) devices—technology that uses speech recognition to understand commands and answer questions (i.e., Siri, Alexa, Echo, and Cortana). Voice assistant devices are one of the most widely adopted AI technologies, used by around 128 million consumers in the United States alone (Dickson, 2022). Consumers use VA devices for a variety of daily tasks, from setting an alarm, to requesting a statistic, to engaging in casual conversation. We interviewed 21 VA designers at three major technology companies that produce popular VA

Empirically, our findings contribute to the scant body of sociological research on VA devices. Most research on VA devices has been conducted in the field of human–computer interaction (HCI). This research almost exclusively focuses on consumers’ experiences of VA devices, including consumers’ perceptions of VA devices’ social categories and personality (Abercrombie et al., 2021; Curry et al., 2020; Pradhan et al. 2019; Sutton, 2020), identifying ethical concerns with VAs (Seymour et al. 2023), accessibility for older adults and persons with disabilities (Hanley and Azenkot, 2021; Nicenboim et al., 2020), the nature of consumers’ relationships with VA devices (Pridmore and Mols, 2020; Xu and Li, 2022), and consumer privacy (Chalhoub and Flechais, 2020; Turow, 2021). In contrast, our research focuses on the production of VA devices.

More broadly, our findings extend social scientific research on anthropomorphized AI produced within organizations. Science and technology studies scholars have conducted laboratory ethnographies to examine how people's interactions with technology shape the production of knowledge and technology (Knorr-Cetina, 1981; Latour and Woolgar, 1986; Lynch, 2017). More recently, sociologists analyzed how workers use technology within various organizations, including technology organizations, focusing on how workers incorporate AI into their work practice (Barrett et al. 2012; Beunza, 2019; Lebovitz et al., 2022; Shestakofsky, 2017) and entrenched assumptions embedded in algorithms (Ziewitz, 2016). However, sociologists have only begun to examine how designers produce AI (Lee, 2022; Sachs, 2020; Tubaro, Casilli and Coville, 2020; Wajcman, 2019). To our knowledge, there is no social scientific research on how designers produce anthropomorphized AI, and more specifically, think about how to design anthropomorphic qualities into these technologies. These decisions are critical, as they fundamentally shape consumers’ interactions with and perceptions of anthropomorphized AI and also influence how we think about the appropriate role of these technologies in social life more broadly.

To address this question, we apply Goffman's (1983) theory of “persons,” which are products of collaborative engagement created momentarily through a face-to-face interaction. Goffman specifies two distinct yet intertwined elements that people use to ascribe personhood to each other: social identity—categorical attributes, such as gender and occupation—and personal identity—individuating qualities such as personality traits and biography. We reveal how designers themselves selectively and strategically employ elements of social identity and personal identity to build VA devices. As a result, devices communicate what we call shifting personhood—they mimic dimensions of social and personal identity to varying degrees at different moments.

Literature review

Human–nonhuman interactions in organizations

Our study contributes to social science research on human–nonhuman interaction within organizations, specifically the production of anthropomorphic features in AI.

While early sociological theorists focused on human social interaction (Mead and Schubert, 1934), sociologists in recent decades have given more attention to nonhumans in the study of the social world, examining both how humans socially interact with nonhumans and how nonhumans mediate social interaction among humans. In particular, actor-network theory (ANT) scholars have redirected the bounds of social interactions away from humans and instead across a broad network of human and nonhuman actors (Callon, 1984, 1987; Latour, 2007; Law et al., 2012). Actor-network theory scholars argue that nonhumans are not simply reflections or objects of projection but instead independent social actants that can impact or influence the actor-network assemblage (Latour, 2007

While ANT scholars have examined the relationship between human and nonhuman actants as seemingly mundane and primitive as shellfish (Callon, 1984), VA devices represent a case of more technologically advanced nonhuman actors—specifically, anthropomorphized AI. Since the 1990s, scholars in communications and HCI have explored how computers fill a social role (Nass et al., 1994; Reeves and Nass, 1996). In particular, these theorists have examined how people interact with robots—artifacts with qualities, such as materiality, anthropomorphized features, and the capacity to physically move—in workplaces and everyday life such that they interpret them ontologically and interactionally as social actors within new and evolving legal and political frameworks (Alač, 2009, 2016; Alač et al., 2011; Breazeal and Scassellati, 2002; Lipp, 2022; Meister, 2014; Jones, 2017). Empirical work examines how humans engage in meaningful communication with robots (Turkle, 2005, 2007; Turkle et al. 2006; Vertesi, 2015). For example, Darling (2021) uses historical antecedents to understand how humans relate to robots and develop deep emotional connections, often explicitly due to their anthropomorphized desi

More recently, the advancement of generative AI has spurred a vibrant, interdisciplinary research agenda concerning AI design. A significant cluster of this work, spanning HCI, psychology, and communication studies, investigates the relational and psychological impact of anthropomorphized AI on consumers (Turkle, 2024), examining phenomena such as how consumers perceive AI as having emotions like empathy (Ayers et al. 2023), how consumers develop trust through conversational interaction (Bae et al. 2023), and the complex, bidirectional relationship between chatbot use and consumer loneliness (Inzlicht et al. 2024). A second, parallel cluster of research addresses the complex ethical and societal questions these systems raise, from their potential for misuse and the need for new safety protocols (Gabriel et al., 2024) to the rights of AI systems themselves (Forrest, 2024; Boyle, 2024) and the ethical design of anthropomorphized AI (Akbulut et al., 2024). However, as with the broader literature on anthropomorphized AI, including VA devices, this recent work has focused less on the production process of these systems from the perspective of the designers themselves.

Complex technologies like anthropomorphized AI are often employed and produced within organizations. Social scientific research examining the use of technology within organizations has been rooted in foundational ANT scholarship, which drew upon ethnographies of scientific laboratories to provide rich analyses of how organizational goals and lab personnel work shape the production of scientific knowledge (Domínguez Rubio, 2014; Forsythe, 1993; Knorr-Cetina, 1981; Latour, 1992; Latour and Woolgar, 1986; Lynch, 2017; Shapin, 1988). This body of research has more recently focused both on how people incorporate AI into their work practices (Barrett et al., 2012; Beunza, 2019; Lebovitz et al., 2022; Shestakofsky, 2017) and how people design, implement, and maintain algorithms within a range of organizations (Kellogg et al., 2020).

Yet social scientists have only begun to examine the development of information technology within technology companies themselves (Lee, 2022; Sachs, 2020). This work has studied how teams within technology companies make decisions around the production of algorithmic technology, such as how teams classify qualities in ways that make them legible to machines (Sachs, 2020). However, this work does not engage with how teams design anthropomorphic features into these technologies. Our work brings Goffman's (1967, 1983) theory of personhood into conversation with research on the production of AI within organizations to address this question.

Producing personhood in human–nonhuman interaction

Goffman's (1983) theory of personhood is ripe for applying to human–nonhuman interactions because Goffman theorizes personhood as emergent within face-to-face interaction, rather than an objective feature belonging exclusively to humans. By reframing the analysis of personhood away from psychology or Mead and Schubert's (1934) conception of “self,” he instead focuses on how personhood is “collaborative[ly] manufacture[d]” within interaction (Goffman, 1983). For Goffman (1967), the interaction order is in fact a production order, one in which persons are created momentarily through face-to-face interaction involving multiple individuals (Cahill, 1998; Formlian and Stark, 2021; Rawls, 1987). Within the interaction, each individual is equipped with “systems of social classification and identification” (Cahill, 1998), which they use to influence how the other defines their person and determine whether or not to accord personhood to the other. While Goffman explores personhood production in face-to-face interactions among humans, we examine how technology producers conceive of how to design anthropomorphic attributes in nonhuman technology as they imagine future interactions between AI and human consumers.

Goffman's theory of personhood helps explain how designers think about selecting certain anthropomorphic elements and the effects of these decisions. Goffman (1963) argues that personhood is composed of two types of identities: social identities and personal identities. Social identities are categorizations used by “society to [establish] the means of categorizing persons and the complement of attributes felt to be ordinary and natural…” (Goffman, 1963), such as race, gender, occupation, and familial status. As social identities are integrally woven into dynamics of social order (West and Zimmerman, 1987), the attribution of social identities can contribute to the creation of hierarchy and inequality within face-to-face interactions. In the personhood production process, social identities are perceived and produced in conjunction with personal identities—objects of classification which identify the other as a distinct individual, such as physical appearance, social security number, and name (Brensinger and Eyal, 2021). Personal identities also extend to other individuating features, like personal “style” (Goffman, 1974). Together, one's social and personal identities communicate self-sameness (recognizing individuals as the same individual over time) and uniqueness (recognizing individuals as distinct from one another). However, as Brensinger and Eyal (2021) note, self-sameness and uniqueness are in fact continuums, as attributes of self-sameness can vary over time, and individuals may not be unique in all attributes.

We apply three key insights from Goffman's theory of personhood to help explain how designers navigate decisions around equipping VA devices with anthropomorphic qualities. First, Goffman's conception of personhood as collectively manufactured paves the way for examining how people construct, rather than merely attribute, anthropomorphic qualities in nonhumans. Secondly, Goffman identifies the central elements of personhood—social and personal identity—that designers employ in this process. In terms of social identity, we focus on how designers conceive of anthropomorphized AI's social categories, such as gender, age, and birthplace. In terms of personal identity, we examine how designers conceive of VA devices’ personality traits, biography, and opinions. Third, as self-sameness and uniqueness in humans exist on continuums (Brensinger and Eyal, 2021), social and personal identities can be theorized as only partially inhabited in humans and nonhumans alike. Our concept of shifting personhood highlights how designers do not always seek to maximally mimic humans when designing anthropomorphized AI; Instead, they selectively build certain elements of social and personal identity into these technologies, resulting in a product that mimics anthropomorphic qualities more or less closely along various dimensions of personhood, which are elicited at different moments of interaction with consumers.

Data and methods

One of the authors, Margot Hanley, conducted 21 interviews with VA device designers at three major VA companies. The three companies were selected because, at the time of the interviews, they collectively accounted for the vast majority of the VA device market, which in turn represented the most widely used form of anthropomorphized AI.

The data were collected in two waves: February 2018 to March 2018 and August 2020 to June 2021. Before scheduling each interview, the participant was sent an information sheet outlining the study’s purpose, voluntary nature of participation, potential risks, and the right to withdraw at any time. All interviewees were bound by corporate nondisclosure agreements. Participants were often upfront that they would not share information that violated their agreements, and they expressed concern about protecting their personal identity and their companies’ intellectual property. Acknowledging these constraints and our ethical responsibility to minimize subjects’ risk of professional or legal sanction, we adopted a strict two-step deidentification protocol to minimize subjects’ risk of professional or legal sanction: (1) each company is referred to by a fixed pseudonym (Company A, Company B, Company C) and the product class is described generically as “VA devices,” and (2) potentially identifying job titles are abstracted (e.g., “Senior Conversation Designer” becomes “mid-level conversation designer”). Additionally, highly specific design anecdotes are redacted to prevent traceability, and we did not report certain proprietary information at the request of interviewees. Audio recordings and transcripts are stored in encrypted, access-controlled folders on institutional servers, separated from the master code key.

Of the 21 designers, 12 self-identified as women and 9 as men. Participants ranged in age from their mid-20s to early 60s. Although most were based in the United States, several designers worked for the same companies from offices abroad. The term “designer” encompasses a spectrum of roles integral to VA personhood work: conversation or UX writers, interaction/voice UX designers, software or ML engineers, product managers, program managers, and data analysts. While the focus of this article is on VA designers, the author also conducted 16 interviews with VA consumers to understand how they interacted with the devices and perceived the personhood that designers worked to achieve.

Interviewees were recruited via snowball sampling and professional networking websites, such as LinkedIn. As designers had signed nondisclosure agreements with their companies, recruitment was a challenging and lengthy process. Interviews were conducted via phone or Zoom and usually lasted one hour. The first round of interviews included designers who worked on a broader set of design roles, whereas the second round of interviews focused on designers who significantly contributed to the expressive qualities of the device in particular. Twenty of the interviews were recorded and transcribed verbatim. One interviewee requested that the author not record, and extensive notes were taken throughout the interview in this case.

We adopted semistructured interviews because the research question—how designers of VA devices think about how to design anthropomorphization in order to produce a particular user experience—requires eliciting designers’ tacit reasoning, storytelling, and reflections on individual and collaborative practice. Such sense-making processes are rarely observable in situ and are not captured in formal documentation, making in-depth interviews the most appropriate primary method. The interview guide balanced comparability—common prompts across participants—with flexibility to probe role-specific experiences.

The interview script comprised three thematic blocks aligned with our research question: (1) role & workflow—how day-to-day responsibilities intersect with anthropomorphizing decisions; (2) personhood design decisions—social-identity choices (gender, accent, backstory) and personal-identity choices (personality traits, opinions) that define what is being anthropomorphized; and (3) organizational and market constraints—why teams adopt or avoid anthropomorphic cues (e.g., brand positioning, cross-cultural universality). Tailored follow-up questions provided depth while preserving comparability across roles. Interviewees used different terms and pronouns when referring to the VA device, for example, by calling the VA device by its “name,” by a gendered pronouns (i.e., “she”), or nongendered pronouns (i.e., “it”). The author avoided the use of a name or pronoun until the interviewee had referred to the VA device in a certain way, and then the author would use their most recent classification.

We inductively coded the interview transcripts in an iterative, three-stage process in Excel. Open coding generated provisional, low-level labels for any quotes regarding identity or anthropomorphism (e.g., “accent neutrality,” “backstory disagreement”) and revealed the salience of this issue for designers. This led us to focus on how designers conceive of how to build anthropomorphic qualities into VA devices. During axial coding, related labels were consolidated into broader thematic families, generated through constant comparison and memo writing. A selective-coding round then distilled those families into two analytic axes—“social identity” and “personal identity”—that frame the findings, with tabs for each subcategory, such as “absent social categories” and “mid-range personality traits.” Deviant cases were flagged throughout to ensure minority or contradictory perspectives remained visible in the analysis.

The three companies differed in the degree to which they selected human-like and gendered names for VA devices; however, we found no differences across the three companies in terms of how designers conceived of VA devices’ social and personal identities. Therefore, we analyze data from designers at different companies together. Social and personal identities are not discrete. For example, people perceive social categories through the expression of personality and vice versa; however, we analytically separate discussions of social and personal identities in the analysis to emphasize how designers conceive of each, and, where relevant, we explain how their coproduction shapes one another.

The following analysis proceeds in three main parts. First, we examine how designers wrestled with how to best conceptualize VA devices’ social and personal identities in order to aid their work of scripting speech for the device. The second and third sections focus on how designers conceive of VA devices’ social identities and personal identities, respectively, to reflect how they wanted consumers to interact with the devices. We show how designers selectively programmed different aspects of social and personal identity to mimic anthropomorphic qualities more or less closely, resulting in a device that communicated a shifting personhood to consumers.

Analysis

Conceiving of absent or latent personhood to aid design

Designers faced the task of scripting sequences of responses for VA devices, believing that human-like conversations would make VA devices more useful and engaging to consumers. They used their understanding of consumer expectations and human behavior to script naturalistic interactions. Specifically, designers explained that they scripted VA devices by practicing lines aloud with each other. They would start by building the “core” component of the conversation, and then add more sequences for a “multitiered” interaction. Danielle explained, “So in this case, we started with one turn and then gradually started to add more turns. And so the first turn, when we were experimenting…it was like…'So do you want to hear… something I just learned about [topic],’ and then the consumer would say, ‘Yes,’ then we would respond to them. That would end the conversation and then slowly building from there…starting with those ideal dialogs and then breaking it down into sort of its core components and then creating a flow, for lack of a better word, of how the conversation might go. So that involves mapping out if the consumer says, ‘Yes,’ if the consumer says, ‘No,’ if the consumer says, ‘I don’t know,’ which is something that happens a lot with [demographic], and then some other different possible scenarios as well and how we would handle each of those.” In this back and forth, designers had to imagine what consumers might say and how the VA devices should respond to them.

Within and across technology companies, designers took two different approaches regarding how to anthropomorphize VA devices in their own imaginations in order to aid this scripting process. Some designers imagined VA devices as having a latent social and personal identity which they conceptualized in order to build scripts but which was not necessarily obvious to consumers. Chandler recounted how a lead designer told the team to envision the VA device as a thirty-something-year-old female who liked to play [a specific sport] on the weekends. While he explained that he had a “fundamental philosophical disagreement” with this approach and had “agreed to disagree” with the lead designer, others described this approach as useful to their work. Audrey stated that imagining this woman had helped the team choose an appropriate voice actor for the VA device: “That was the intangible sort of thing … I do think that the personality of the female voice, like for real, she does sound like she could go [the sport], she sounds athletic and kind of perky and like she's up for doing something fun like that.” While the VA device would not make these social categories explicitly known to consumers in interaction, Audrey felt that imagining the VA device as an actual human—with both aspects of social identity, such as gender and age, and personal identity, such as hobbies—helped her decide the right voice for the device.

Designers who preferred to envision the device as human cited gender as particularly useful in conjuring a specific human in the design process. Benjamin stated, “It helps me when I’m writing dialogue. I want to try and create the sense that the customer is interacting with an intelligent entity. And so it kind of helps me to envision [the VA device] as an intelligent entity. And I use ‘she’ mainly because the default voice is female. And that's been the tradition in the voice assistant space, but I mean, I’m very aware of gender dynamics and the problems with that. And if I could, if there were a gender-neutral term I could use that didn’t make [the VA device] sound like an object, then I would … but for [the VA device], it just doesn’t feel right for me at this point.” Benjamin explained that imagining the VA device as having a certain gender helped him mimic human interaction more convincingly in the scripting process.

Other designers chose to conceptualize VA devices as having no inherent personal identity and absent social identity. Jade described writing AI as different from writing characters in other contexts, like a screenplay, in which social categories are built into the character, due to the potential blank slate regarding social categories that AI affords: “So it's more getting into a different mind frame that people may not be aware of, because when you sit down to write a character [in a screenplay], [the character is] automatically probably starting off with those demographic traits. But when you’re writing for AI, you’re maybe starting from a place where those traits are not inherent, so you’re just starting from a different place versus taking something and stripping it away.”

These designers also pushed back on the notion that their work required engaging with gender. Jade, for example, noted that while it was possible for consumers to choose the gender of the voice, her scriptwriting was ultimately genderless. Jade insisted that the VA device's script was something that was equally sensible for a male or female to say, “I do think gender is something that people may unconsciously write from… but a huge amount of content and meaningful content out there if I put it in front of you and I said, ‘Did a man say this or a woman say this?’ It probably would not be that easy to tell … Much of our conversation is not so gender-loaded as you would think. I think the bias people have about voice in this is the oral voice, they prefer hearing things from a woman. … As far as the scripted words go, they’re not really loaded in any way.” These designers argued that it was possible to script from a blank slate because human communication itself often did not reflect one's social and categorical identities.

More broadly, this “fundamental philosophical disagreement” among designers reflected ambivalences about how they should conceive of these devices and how their own conceptions of VA devices’ social and personal identities would become expressed in the device's interactions with consumers. Designers believed that even if it was technologically possible to design the VA devices to convincingly mimic human speech, this should not always be their goal. As Chandler stated, “We really have an opportunity right now to define the future of what these entities look like and what they’re able to do, what they should do…nobody has sort of cracked the code on what it means to have an AI that you could talk to and trust and that trusts you, and it's really on us to define what that is and what it looks like and what is appropriate and inappropriate.” As designers imagined how they ideally wanted consumers to interact with these devices, they had to make further decisions regarding how the devices should communicate social and personal identities to consumers.

Conceiving of pure and midrange social identities

Designers selectively built certain elements of social identity into VA devices. Designers wanted consumers to perceive the device as akin to a human service worker, similar to what one designer described as a “low-skilled employee” who performed workplace tasks that could be automated and that others did not wish to do. As Taylor stated, VA devices were “pure” assistants: “Even if he or she is your assistant, that person is more than that, way more than that. But with [the VA device], as is my personal opinion, is that they’re purely that character… ‘assistant’ is her entire identity, and the purpose of existence. Where, any human being is just way more than that.” On the one hand, designers viewed these interactions as more superficial. Benjamin explained that the relationship they expected their consumers to form with the VA device was “not the kind of relationship bonds that you create with a human being necessarily. But certainly the kind that—[the] superficial time that you would form with a human being who you were interacting with in a very transactional way, like a grocery checkout clerk, or something like somebody like that.” While initially stating that the bond between the VA and the consumer lacked the emotional intensity of human bonds, Benjamin then revised this statement to explain that it was similar to the fleeting and transactional human interactions between customers and service workers. However, Benjamin also recognized a certain level of intimacy between the VA device and the consumer due to the VA device's placement in the consumer's home. He stated, “The goal is for her to be helpful, for her to be trustworthy, reliable, not overly formal… she belongs in your home … so friendly, familiar.” Previous sociological research describes how domestic workers and service workers in other intimate settings often form close personal relationships with their clients, buttressed by emotional labor on behalf of the service worker (Kang, 2010). Similarly, designers built VA devices to implicitly express the social identity of a domestic worker, an occupation that they associated with certain elements of personal identity, such as emotional warmth and helpfulness.

Designers also believed that programming VA devices to interact similarly to a human service worker allowed consumers to experience their interactions with VA devices as seamless and cognitively untaxing. Natalie explained, “Do I normally have an [athlete], a star celebrity [athlete] helping me order food? [No] … but we do interact with servers who help us order food. So that would be a more appropriate frame for the conversation … So that might be in service of continuing a mental model.” For Natalie, programming the VA device to interact like a service worker made the interactions more continuous with consumers’ experiences of interacting with humans.

However, designers did not seek to convey that VA devices inhabited all elements of social identity equally. Beyond occupation, designers sought to have VA devices communicate an ambiguous “mid-range” of certain social categories, such as age and geography, believing that this would make devices more universally appealing to consumers. As Audrey stated, “So she doesn’t speak terribly quickly, but she's not a slow talker. She has a very midrange pitch. She doesn’t have a very low voice for a woman, but she's also not very high-pitched like chirpy. …almost all voice systems are in very standard North American English, like what we call newscaster English. So there's no one of a particular area. You don’t have Boston accents or New York accents or Southern accents. They also do tend to be kind of midrange in terms of age. And I think that's just because we don’t want to have any discernible, like, ‘Wow, she sounds really old or young,’ but more of a nebulous in between.” For Audrey, VA devices inhabited these social categories, but they were indiscernible or middling. Likening VA devices to sports announcers, Audrey said, “I’m trying to think like some of the Boston announcers or the LA announcers, who were very targeted towards it or my demographic, or very much my narrow group, but then maybe not understood or recognized outside that group. And so you might sacrifice some lovability in order to be a little more bland and accessible to a broader group.” Designers programmed VA devices with voices that they believed communicated midrange or vague social categories, as they prioritized universal appeal above strong engagement for a narrower consumer base.

In seeking to produce a product that was helpful and had broad appeal to consumers, designers produced a shifting personhood in which they selectively chose certain elements of social identity to be expressed more or less strongly. In some ways, designers constructed more anthropomorphic social identities, like when they programmed the device to mimic a human service worker in order to be legible to and emotionally resonant with consumers. In other ways, designers shied away from anthropomorphic social identities, like when they used unclear or middling social categories to appeal to a broad consumer base. Overall, the result was that designers programmed VA devices to communicate as they imagined a service worker, while ambiguously or indistinctly inhabiting other social categories.

Conceiving of personal identities without uniqueness

Consistent and midrange personality traits

Designers also conceived of a shifting personhood concerning personal identities—specifically, personality traits, biography, and opinions. Designers tried to vary VA devices’ response style and tone depending on the context; however, they strove to program VA devices to express consistent personality traits in order to facilitate ease of interaction, believing that the VA device would seem unnatural if the personality traits seemed to vary widely from moment-to-moment. As Chandler explained, “And if we don’t carry along a consistent experience, a consistent character with it, it will be very, very strange.” Designers described VA devices’ personality traits as helpful, informative, informal, friendly, fun, and succinct. They noted that they walk a fine line so that VA devices would interact in ways that consumers found both efficient and friendly. As Audrey explained, “It's like the assistant is going to help you get what you need as soon as possible, explain it to you in as few words as possible, but is friendly and helpful, but is not [using] flowery language and not telling you too much.” Designers deemed the playful and warm traits as especially important for physical VA devices placed in consumers’ homes. Wallace stated, “[The VA device] actually had a human-like personality. I think that was a good way of driving adoption with people that weren’t comfortable with technology because you have this speaker in your house and it's listening, you know when you say the ‘wake’ word it listens to you. And some people are just not comfortable with technology and by adding a human quality to it, I think it softens the hard edges, it allows people to become more comfortable with it and then explore more of the utility features, like calendar, timer.”

Given the global distribution of VA devices, designers strove to make the devices that would appeal to a broad demographic of consumers. Designers often used “inclusivity” and “accessibility” to discuss their efforts to make VA devices interact with broad and cross-cultural appeal. Designers talked about seeking out variations in speech patterns so that VA devices would respond appropriately to different consumers. One strategy was to ask different members of the team to respond to different prompts. Furthermore, designers discussed how to make the VA devices express a broadly appealing persona. Similarly to how some designers sought to communicate the “midrange” of social categories through the voice, designers also sought the “midrange” of personal identity through the style of response. Audrey explained that they try to avoid “niche personalities.”

However, designers also recognized that consumers would perceive personality traits differently across cultures. Leslie, a content strategist for an internationally oriented unit, described how her team made distinctive coding for different countries. She claimed that attributes, such as friendliness and helpfulness, were “a universal language” that designers maintained across countries. However, even these attributes were sometimes communicated differently. For example, Leslie noted that an earlier version of the American VA device was optimistic to a point that it would be perceived as insincere to a certain European country's audience, so they had to “tone down” the optimistic responses for that country's VA device. Furthermore, they programmed distinctive cultural references by country. Leslie stated that one country's VA device was “obsessed” with a particular sport, in that the device would provide detailed in-depth information on questions related to that sport when prompted, while another country's VA device would provide more information about a different sport.

Designers especially focused on distinctive humor, believing that humor was more culturally specific than other types of communication. Leslie gave an example of asking for directions in a country, explaining that people from that country tended to provide landmarks, instead of street names, as referent points. When consumers asked for directions, the team programmed a playful response, in which the VA device stated something like, “Oh, you go to the [type of plant], you turn left, you know, and when you get to the [type of structure],’ and then [we have it say], ‘No, I’m joking.’” However, designers stopped short of programming culturally specific humor when they felt it could alienate a specific group, which could limit their demographic reach and run against their company values. Leslie provided the example of a specific town in a country whose residents were mocked as “stupid” by the rest of the country. While consumers prompted the VA device to tell jokes about this town, the designers would not program this into the device, believing that it was “cruel” and could alienate consumers from that area. In seeking consumer engagement and broad appeal, designers produced a shifting personhood in which they programmed devices to express midrange personality traits in response to certain prompts and culturally distinctive personality traits in response to others, depending on location.

Absent opinions and distanced positions on social issues

In an effort to maintain VA devices’ universal appeal, designers also shied away from communicating a human-like biography and personal opinions. Kelly stated that the designers would not promulgate a “lie” about the device, even when prompted by consumers, “Even I have, just for fun, … things like, where do you live, what's your favorite color, and things like that. … We try to make the answers seem as natural as possible, but at the same time, like there's no made up lies about what's maybe the assistant's favorite food or something.” In these areas of interaction, designers held back from building more anthropomorphic personal identities. Designers asserted that the VA device's lack of communicated opinions was important to maintaining broad appeal, particularly when it came to fraught social topics that could alienate some consumers, such as sports, religion, or politics. As Benjamin explained, “I think basically politics and religion are kind of a third rail and it's not, I think the decision is basically it's not appropriate for [the VA device] to take a political position or certainly not a religious position.” Taylor described a situation that “embarrassed” her, in which an earlier version of the VA device pulled information from a movie database and described Jesus as a “fictional character.” Taylor explained that to avoid these controversies, designers discussed and manually programmed responses to sensitive topics.

However, designers did not shy away from all contentious topics. At times, they perceived communicating company values to supersede broad appeal. Steven explained that the team considered, “how can we make sure that we’re not giving [consumers] the impression that [VA device] quote-unquote thinks or believes something that isn’t compatible with our values as a company,” further stating that he was fortunate that his personal values aligned with the company's values. They expected that, in the moment of interaction, consumers would both imagine the VA device as an intelligent entity with political positions, while also recognizing that the VA device was programmed by employees who were, in turn, tasked with representing the company's values.

Specifically, designers discussed programming VA devices to have affirmative responses to the question, “Do Black lives matter?” and communicate support for the Black Lives Matter movement more broadly. Designers described their companies and its CEO as “liberal,” using the blanket term of “company values” to describe their perception of companies overarching moral commitments and stances on salient social issues. For example, designers stated that support for the Black Lives Matter movement was aligned with company values. As Wallace explained, “So we threw together like all these attributes of a personality and we’re now starting to evolve this personality and cover areas, like people will bring up a big hot topic like, Black Lives Matter for example, and we’ll look at [VA device's] personality and be like, how does [VA device] respond to questions when a customer asks, ‘Do Black lives matter?’ But we have to be very careful because we want to be customer-centric so we can’t appear—we have to be very balanced. Like we don’t want to be totally on one side or totally on the other. We also have to call out that she's not human, and that she's saying these things that represent [the company].” Another designer, Benjamin, suggested that the device reflects the political commitments of the organization, “There are political positions that she takes because they’re fundamentally, they’re fundamentally compatible with [the company's values]. I think if you ask her, ‘Do Black lives matter?’ you’ll get what I think is a pretty good response.” While the designers programmed VA devices to respond affirmatively to the question, they scripted devices to state the position as the company's stance, thereby avoiding the “lie” that the opinion derived from the device itself. Thus, designers built the VA device to espouse the position, but to distance itself as the entity responsible for taking the position.

Designers also curtailed what they described consumers’ “abuse” or “harassment” or devices, such as moments that consumers swore at, yelled at, or otherwise mistreated devices. Designers reported that they had been initially unprepared for the amount of abuse levied upon the devices, and that they felt that consumers were more abusive toward devices with female voices. While they sought to diffuse these interactions with jokes in earlier iterations, they found that this only egged consumers on, so they ultimately had the VA devices “shut it down” by telling the consumer that that was not appropriate or by simply not responding at all, even though this inhibited consumer engagement. Designers gave multiple reasons why these types of interactions were inappropriate. Leslie stated that consumers’ interactions with VA devices should model appropriate human-to-human interactions, noting that if a consumer abuses their VA device, “[They’re] probably doing it to females in [their] life, you know? And that's kind of the approach of like, we have to kind of do something to create a— to model good behavior.” She also argued that VA devices themselves deserved to be treated with moral dignity, explaining, “So, you know, that also goes against, kind of like, human principles, like we have to respect each other and your freedom ends where [an]other person's freedom starts. And even if it's not a real person, we just don’t want to build that into it. We don’t want to kind of let people abuse [them] just because [they] are not human.” In instances of what designers considered to be abuse, they sacrificed consumer engagement in order to maintain what they believed to be appropriate standards of communication.

In an effort to engage with and broadly appeal to consumers while also asserting company values, designers often held back from scripting anthropomorphic personal identities, even when this was technologically possible. Designers programmed VA devices to avoid communicating personal opinions, except for stances on specific social issues, for which they scripted the device to explicitly express as the company's position. In the face of abuse, they programmed the device to shut down. As these facets of personal identity would be elicited at different moments depending on the consumer prompt, the result was that VA devices were encoded to express a shifting personhood.

Conclusion

Drawing on interviews with designers of leading VA devices, we have shown how designers selectively built certain elements of social identity and personal identity into VA devices. While designers sometimes had the technological capabilities to convey more anthropomorphic social and personal identities, they designed the devices to mimic anthropomorphic qualities more or less in a given situation, depending on how they prioritized different goals of user experience, such as helpfulness, broad appeal, and representing company values. Designers constructed devices so that they expressed facets of social identity as unevenly inhabited. In terms of personal identity, they designed devices that were mild-mannered and generally emptied of opinions and preferences, while deploying culturally specific knowledge and expressing certain convictions. Designers sought to reap the benefits of personhood—like engaging interactions—while limiting the drawbacks—such as expressing more commonalities with some people than others. As a result, designers constructed a device that communicated a shifting personhood in both social and personal identity.

Theoretically, we draw from Goffman's (1963) theory of personhood to understand the core components through which designers conceptualize personhood—social and personal identities—and to conceive of social and personal identity as more or less fully inhabited. Accordingly, our concept of shifting personhood captures how designers build select features of social and personal identity into anthropomorphic AI, rather than seeking to construct technology that either fully inhabits or altogether lacks personhood. For example, personal identity is made up of several dimensions, such as personality traits and opinions. In turn, personality traits and opinions are each made up of many traits and opinions, which can be communicated or not communicated in various situations. Designers programmed VA devices to express support for the Black Lives Matter movement but to decline to name a favorite football team or a geographical birthplace. In this case, the VA device expresses one position in a distant way—in the role of a company spokesperson—and does not express another position at all. Thus, designers construct devices that convey different dimensions of social and personal identity unevenly, with these dimensions elicited at different moments depending on the prompt. These features of personhood are not wholly absent from human–human interaction; for example, employers often ask service workers to not express political positions at work. However, the construction of personhood is more explicit in anthropomorphized AI, as the personhood of the interactant is not taken for granted and must be designed into the product, rather than merely attributed to it.

This study offers significant empirical implications for academic research, industry practice, and public debate. While recent sociological research has focused more heavily on how anthropomorphized AI shapes the workplace (Barrett et al., 2012; Beunza, 2019; Lebovitz et al., 2022; Shestakofsky, 2017) and how consumers interact with such technologies, our research demonstrates that the design of AI personhood—often framed as a technical challenge within HCI—is fundamentally a sociological process of production in that designers navigate inherently ethical, organizational, and cultural concerns in this process. By revealing how designers grapple with constructing social and personal identities to manage interactions, this research calls for greater integration of sociological theory into the study and development of anthropomorphized AI.

Second, our analysis shows that designers think about various ethical and inclusivity issues regarding how to appropriately anthropomorphize VA devices; however, designers sometimes disagree among themselves about these issues, and designers from different companies likely do not discuss these issues with one another given nondisclosure agreements. Our findings make these considerations more explicit, providing a platform for designers to heighten their awareness of these issues and a language to discuss field standards.

Finally, our findings contribute to current law, policy, and governance debates concerning anthropomorphized AI. Academic scholars have begun raising alarms about the psychological impacts of human–anthropomorphized AI relationships and communication (Turkle, 2024). At the same time, recent lawsuits and Federal Trade Commission (FTC) complaints against companies highlight the potent and potentially harmful effects of AI anthropomorphization (Sewell Setzer v. Google and Character.AI, 2024; Federal Trade Commission [FTC], 2025). For example, the FTC has directly addressed the “inherent risks in terms of manipulation” in human-like AI, warning companies not to “use consumer relationships with avatars and bots for commercial manipulation” (Atleson, 2024; FTC Complaint against Replika 2025). By focusing on the production of anthropomorphic qualities, our research reveals how anthropomorphization is deliberately engineered by corporations to increase engagement and wide adoption. We show that the production of shifting personhood is the result of a strategy in which designers selectively program some anthropomorphic qualities—such as the helpfulness of a service worker—while neutralizing others—like adopting a “mid-range” of social categories. These findings reinforce questions about whether companies should continue to have free rein over how to anthropomorphize AI, and how the production of anthropomorphized AI should be regulated to protect consumers. These questions become more urgent as advances in technology, such as the emergence of LLMs, create the ability for designers to build anthropomorphized AI that mimics human interaction even more closely and seamlessly than VA devices.

While our study is confined to VA devices, new questions about anthropomorphizing AI have emerged as design practices shift toward tuning generative systems like LLMs. Our findings on the process of design might not fully capture new workflows and challenges. Future research should investigate the changing nature of the work required of designers to build these systems. For instance, which of the challenges identified in this study—such as constructing a “mid-range” personality to ensure broad appeal or navigating when to take a distanced position on a social issue—persist in this new wave of anthropomorphized AI? How might designers grapple with new concerns given emerging capabilities to generate hyperpersonalized anthropomorphized AI for each consumer that is algorithmically generated or that shifts some of the design decisions around personhood to consumers themselves?

Further, our focus on VAs means our findings primarily address personhood constructed through auditory and conversational cues, like voice pitch and accent. This focus necessarily excludes how personhood is designed in other modalities where different attributes are central. Text-based chatbots, visual avatars, and physical robots, for example, rely on different attributes such as writing style, appearance, and embodied gesture, instead of voice. Future research should comparatively analyze how the work of designing personhood differs across these modalities, and even across new sensory inputs like digital scent, each presenting unique social and technical challenges and questions around producing anthropomorphized AI.

Furthermore, our findings open new avenues for research on the ethical and cross-cultural implications of anthropomorphized AI design. We found that designers programmed VAs to take distanced positions on some social issues and to “shut down” abuse by consumers to “model good behavior” positions companies as de facto arbiters of social standards, leaving open critical questions for future inquiry: What are the long-term societal consequences of this practice, and who is, and who should be, responsible for the values these systems promote? Similarly, while designers in our study attempted to localize content like humor, the default persona often reflected a North American context. Future cross-cultural research is needed to examine how these often Western-centric personas are received in diverse global markets and whether attempts at localization reinforce alienating stereotypes or foster connection. As anthropomorphized AI circulates globally, local cultural assumptions embedded in their design travel with them. Ultimately, responsibility in designing anthropomorphized AI lies not just in defining what AI can do and for whom, but also in imagining what kind of “whom” it can become.