Voice Assistants for Mental Health Services: Designing Dialogues with Homebound Older Adults

Over the years, the HCI community has developed VIPAs to support older adults’ general wellbeing, studying approaches to incorporating VIPAs into older adults’ daily living routines [40, 77], and examining old adults’ current use of VIPAs to identify benefits and challenges [16, 19, 31, 45, 64]. Features commonly used by older adults were transactional and task-oriented (i.e., "getting things done"); in particular, the most used features include retrieving information, such as current time and weather, setting alarms and reminders, and launching entertainment services such as music streaming and listening to audiobooks [16, 43, 59, 63, 64, 84]. Some older adults also engaged in small talks [59] as they naturally would with humans [42, 43, 63, 64].

Compared to other smart devices such as computers and smartphones, VIPAs require lower adoption effort because voice-based interactions are similar to how people communicate and interact with one another [31, 42, 43, 45, 63]. Although VIPAs may present accessibility issues for older adults with auditory impairments [43, 61], relying on voice as the primary mode of interaction generally can mitigate other challenges older adults experience due to mobility and visual impairments [33, 42]. Older adults have also reported that they are more confident using VIPAs as opposed to other devices, and they enjoy the playfulness aspect of interacting with VIPAs [16, 86]. However, older adults may stop using VIPAs because they fear that using smart technology (e.g., turning on light switches at home) would diminish their ability to care for themselves or complete household tasks independently [45, 90], are unclear about VIPAs’ values and benefits for them [15, 42, 59, 90], have difficulties operating the device [15, 43], and are concerned with privacy [16, 37, 42, 90]. While most VIPAs are more affordable than alternative options (e.g., social robots and computers), the cost may be a barrier for a subset of the older adult population with lower socioeconomic status [61]. Furthermore, one of the most notable challenges arises from the VIPAs’ capabilities to process and understand natural language. Older adults’ speech tends to get slower, with more long pauses reflecting changes in their cognitive states [16, 24, 38, 58], causing VIPAs to dismiss or interrupt older adult users’ requests and responses [37, 99]. Other language-related issues may include older adults’ speech clarity and varying pronunciation [37, 42, 43, 59, 84]. In Harrington et al.’s study [34], Black older adults explained that they had to code-switch when interacting with VIPAs; they frequently adjusted their speech style from using African-American Vernacular English (AAVE) to a style (e.g., shorter, more formal) that would be understood by Google Home. Although several studies have investigated how older adults use VIPAs in the current technological state, little work has been done on conceptualizing how older adults might use VIPAs in daily lives for mental health support if they were capable of natural multi-turn dialogues and high level of socioemotional intelligence.

VIPAs can support individuals’ mental health in various ways: when experiencing stressful situations, users can request mindfulness activities such as breathing exercises and meditation, as well as soothing music to be played [64, 69, 81]. Studies have also explored using VIPAs for self-reported data collection around mood and activities of daily living [53, 54]. The anonymity and nonjudgmental characteristics of the technology prompted users to disclose more symptoms and personal information compared to a traditional assessment [52, 54, 55], leading to more implementation of mental health screenings through VIPAs [25, 52, 69].

Older adults in particular have expressed interest in using VIPAs for health and wellbeing purposes [57, 80]. They are open to receiving recommendations from VIPAs to engage in health and wellness activities [34, 91] and tracking positive behaviors and emotions [10]. They use VIPAs to retrieve information related to their physical health (e.g., their illnesses and symptoms, medication, nutrition, and vital signs [9, 34]) and to schedule medication intakes for managing chronic illnesses [63]. With regards to mental health, VIPAs can be used and perceived as companions that can provide emotional support for older adults in social isolation [20, 22, 43, 59, 63] by acting as “someone” who is nonjudgmental and always available to talk in the moment of need. Existing research focused on older adults’ mood-boosting activities via their everyday usage of VIPAs, or explored how conversational agents are used to deliver evidence-based mental health interventions, such as CBT, to the general population [29].

Additionally, family caregivers of homebound older adults assist with not only daily activities such as transportation but also physical health [60, 71], mental health management, and emotional support [60]. Because of the vast responsibilities they have, they experience extreme burnout [30, 60, 82] and lack peer support during the process [82]. As such, VIPAs could be a tool to complement their care through existing features such as playing music to improve older adults’ mood [20, 63], as well as monitoring and tracking older adults’ physical health and medication intake [20, 31, 57].

Therefore, our work aims to explore how VIPAs can potentially increase access to evidence-based mental health interventions for homebound older adults without a clinician's presence, considering the specific design requirements and considerations made for this population and how VIPAs can support remote caregivers facilitate their older adults’ mental health care.

When designing conversation between older adults and VIPAs, it is crucial to consider the level of anthropomorphism, or human-likeness, of VIPAs. Although its definitions varied, anthropomorphism refers to attributing a non-human object "humanlike properties, characteristics, or mental states" [26]. The level of anthropomorphism influences users’ future interactions with the nonhuman agent [26]. Specifically for VIPAs, characteristics such as the agent's voice gender and tone can influence users’ perceptions and level of engagement [28] and expectations of VIPAs’ capabilities [15]. In the older adult population, those who experience social isolation tend to anthropomorphize VIPAs more, [33] often refer to the agent as a friend or “she/her” pronouns, and favor a polite and friendly conversational style from VIPAs [22, 31, 36]. As such, older adults generally prefer using polite expressions that follow typical social norms when interacting with VIPAs, such as thanking the devices for answering a question [42, 43]. Similarly, Taggart et al. [87] found that older adults sometimes project their emotions and relate to socially interactive agents. They found that older adults in a nursing home behaved as if they were building relationship with a zoomorphic social robot that express various emotional states, and the robot's animal-like behavior led to increased engagement. However, this could be influenced by the physical form of the robots that mimic pets or other lifeforms, which begs the question of authenticity in human-machine relationships [92, 93]. Unlike social robots, smart speakers do not typically have a life-like physical design, which makes studying conversational styles more crucial.

Although limited work has examined VIPAs’ conversational styles in delivering mental health services, studies on both health-related VIPAs and text-based chatbots have found conflicting results. Older adults expect conversational agents to exhibit competency in finding health-related information when conversing about health [46] and perceive the agent that uses a formal tone to be more appropriate and competent [21]. Yet, some also consider it as less human-like and lacking empathy [21, 22] and express preferences for a warmer and friendlier personality [9, 46]. Additionally, in a review of Amazon Alexa Skills, Shin and Huh-Yoo [81] found that people were more satisfied when the VIPA was more anthropomorphized. To assess anthropomorphism, past research often referenced pronouns that users labeled the agent (e.g., “she”, “my friend”). Some also attributed feelings and human-like behaviors to the agents, especially older adults who saw the agent as a companion. They may ask the agent questions they would normally ask acquaintances, such as whether the agent has a significant other [42]. Yet, these measures do not necessarily indicate preferences of VIPAs’ human-likeness [66]; Pradhan further highlighted that little is known if other characteristics of VIPAs also affect older adults’ preferences. For example, the use of expressive language could be another characteristic that should be explored.

Yet, little work has examined older adults’ preference for the VIPAs’ conversational styles when engaging in mental health intervention delivery and its role in such services. Unlike in most task-based interactions, such as retrieving weather forecast information or streaming music, treating mental health conditions requires engaging patients over long-term interactions with special considerations for their health and socio-emotional needs. As a result, more research is needed when anthropomorphizing VIPAs’ designs for mental health purposes [75, 85].

We recruited older adults and caregivers in the U.S. through research registries, such as ResearchMatch⁴, local senior and community health centers, and the university-affiliated health system for outpatient geriatrics and general internal medicine. Email advertisements and letters were distributed to provide details of the study and contact information. Once a potential participant expressed interest in participating in the study, a research assistant on the team performed an eligibility screening via phone calls. For homebound older adults, the eligibility criteria included the following: (1) 65+ of age, (2) having experienced symptoms of depression, (3) having received emotional and physical support from a family caregiver, and (4) showing no signs of dementia through a cognitive test [13]. Participants were not required to disclose any physical ailment such as motor or visual impairment. For family caregivers, participants must (1) be at least 18 years old and (2) have provided weekly unpaid support to an older adult with a history of depression. Caregivers were recruited to provide an additional perspective as part of the older adults’ immediate care team.

Although participatory sessions are traditionally done in-person, we chose to provide both the option to participate in-person (i.e., at the University or at their home) or virtual. The rationale of this decision was both in consideration of the homebound nature of our participants and the geographical reach of recruitment. When asked, all participants chose to participate in the study virtually. Consent was collected verbally during the screening process. A total of 6 older adults and 7 family caregivers (n=13) participated in the study. Table 1 details participant demographics, including their experience with voice assistants (i.e., smart speakers, mobile voice assistants). Family caregiver participants were not related to the older adults who participated in the study, but they all were either spouses, adult children, or siblings of the older adults they supported. Participants received $50 for their participation in this study.

We conducted 13 co-design sessions via Zoom, where each participant participated in a 90-minute session in a one-on-one setting with either the first or the second author. All sessions were video recorded, and individuals from a university-approved third-party transcription service transcribed the audio data for analysis purposes. For answering the research questions listed, co-design is a useful method because it allows participants to openly create their ideal version of a technology, directly communicate their preferences based on their personal experience. Co-design aims to "produce new knowledge as people develop and experiment with (new) ideas around a matter of concern" [103]. As such, the range and specificity of co-design activities vary depending on various factors, including the participant population (e.g., older adults) and their health conditions (e.g., those with visual impairments).

Although limited activities have been used to study human-voice assistant interactions, several prior works have used dialogue trees and dialogue designs to explore how children and adults would interact with voice-based devices [11, 44, 96]. Similarly, in our sessions, participants completed a dialogue design activity where they drafted conversations between a voice assistant and themselves for six to seven different scenarios (Table 2) centered around behavioral activation as a sample evidence-based mental health intervention. Behavioral activation intervention is commonly used by mental health professionals to help older adults plan and encourage them to participate in pleasurable and rewarding activities to manage their depression [65]. The scenarios were chosen based on existing sample behavioral activation protocols provided by mental health providers on the research team, in which they guide the interactions providers had with their patients. Because participants might not be familiar with participatory design as a research method or writing dialogues, we used the first scenario, Greeting, as a practice scenario for participants to familiarize themselves with the activity.

In each session, the researcher shared their screen to show different virtual dialogue trees for each scenario. The materials for the activity were designed and stored on an online whiteboarding tool, Miro ⁵ . For each scenario, the research team prepared two example dialogues the voice assistant could say to the participants, with one being more formal (e.g., “Now, let's rate the activities you listed. First, on a scale of one to ten, how rewarding do you find (a)? Ten meaning very rewarding, and one being not at all rewarding.”) and another being more casual through more expressive language (e.g., “I'm curious how doing these activities make you feel. On a scale of one to ten, what would you say doing [a] make you feel? Ten for very rewarding and one for not at all rewarding.”). Participants were also asked to create their own or rewrite the dialogues as they desired. Figure 1 shows an example of the initial dialogues provided for a caregiver in scenario 3, where caregiver participants were asked to help curate pleasant and rewarding activities for the older adults they supported using VIPAs. By default, the researchers read aloud the example dialogues for both older adult and caregiver participants. It is worth noting that one participant requested to read through the text on their own instead.

Because the aim of the design sessions was to involve older adults and caregivers as designers when creating technology to support their mental health and routines, the dialogue variants were designed to facilitate ideation and acted as probing materials [103] to encourage participants to think aloud about their ideal interaction(s), rather than being experimental. In fact, participatory design and co-design studies with older adults commonly relied on probing to generate feedback on prototypes [6, 46, 76] because older adults have shown to be more comfortable with discussing their life experiences as opposed to generating abstract new ideas [51]. Therefore, throughout the session, the facilitator regularly reiterated the purpose of the co-design activity. When some older adult participants began to provide answers to the assessment-based questions in the scenarios instead of designing dialogues, the researchers took notes of their responses then refocused them on the dialogue design.

Furthermore, the researchers intentionally probed the participants to re-write the dialogue options when participants showed hesitation and uncertainty about the example dialogues. Participants were then asked to provide a rationale for the dialogues they chose or wrote for each scenario. We also recorded other data generated in the design sessions on Miro, including the responses participants gave to answer questions asked by the VIPA in each scenario and the ordering of dialogues. The word choices and language participants drafted as part of the activity were used to understand how older adults and caregivers would prefer to interact with a VIPA for mental health intervention use. Figure 2 shows an example of the rewritten version of the example dialogues seen in Figure 1. The participant first added an introductory dialogue for the VIPA to initiate the conversation. They then wrote new dialogues using word choices and styles (i.e., the right most yellow card labeled "V:" in each row) they preferred instead of choosing the example dialogues.

Example dialogues from scenario 3, curating activities. A description of the scenario is displayed for the participants. Below are dialogue options presented in card formats, starting with examples of how the voice assistant would begin the interaction. Then, participants can enter how they would respond to the voice assistant.

Example of dialogues rewritten by a caregiver participant. The participant added arrows to indicate their preferred order of dialogues, linking one card to another. Additional cards have been generated on the virtual whiteboard because the participant chose to rewrite some dialogues and add new ones to the scenario.

We analyzed the transcripts using inductive thematic analysis [7]. This provided further insights into the rationale behind the participants’ design decisions and the specific word choices used in rewriting the dialogues, as well as their perceptions of the dialogues used as probing materials. Three members of the research team developed two code books (one for the older adult data and another for the caregiver data) using a bottom-up approach, identifying codes that emerged from the study data. After two iterations of the codebooks and an additional round of re-coding the initial transcripts, the three team members reached a consensus on the codebooks. Subsequently, the remaining transcripts were distributed among the same three members where they finished the coding process using the established codebooks. They also held weekly discussions to identify conflicts by reviewing each other's codes and write analytic memos. Example codes included “Voice assistant is too human-like”, “Greeting voice assistant”, and “Data sharing between older adult and caregiver”. The researchers then generated themes based on the codes, including “autonomy and control” and “preferences for voice assistant conversational style”. Themes were then used to organize the Findings in this paper.

This study was reviewed and approved by the university's ethical review board (Institutional Review Board). Participant information was collected during the phone screening and each participant was assigned a participant ID to anonymize their identity throughout the study. All recordings, demographic and contact information are stored on a university-encrypted server with restricted access to only the research team. Study data on Miro is anonymized using participant IDs. The research team interacting with the participants also received training on risk assessment.

Two caregiver participants experienced the loss of the person they were supporting within the past year. The research team reached out to the participants prior to the session to inquire about the participants’ comfort level in continuing participation. During the session, we asked if it would be appropriate to bring up past memories with their loved ones, and provided a space for sensitive conversations where they could share memories and experiences of loss as needed, as recommended by Sakaguchi-Tang et al. [74]. Participants also had the option to discontinue the session at any time.

One of the challenges in conversation design for VIPAs is determining the level of anthropomorphism of the agent that is contextually appropriate for the task it is engaging in with users. In our study, we found that all participants used anthropomorphic language in the conversation design task but showed varied levels of preferences in the use of colloquial languages. In most scenarios, participants chose to use direct and short language for the agent's utterances, especially when the task was focused on delivering the behavioral activation intervention through the VIPA. They commonly chose dialogues or wrote their own that were straightforward and simple instead of ones with pleasantries. For instance, during the Goals and Reminders scenario, participants were given two examples of possible VIPA utterance, (A: “Would you like to set weekly goals for these activities?” and B: “Now that I know what you/they like doing, I want to know how often you/they would want to do them. Would it be ok if we talk about that?”). Six participants preferred option A and two rewrote the dialogues because the pleasantries in option B felt inauthentic and off-putting.

Although participants shared different reasons for their preferences against pleasantries, the most common reason was the mental model of VIPAs being an artificial intelligence instead of a person, an object rather than a sentient technology. One caregiver said, “‘Thanks for sharing that.’ I don't think it needs to go into too detailed of being empathetic because [...]‘Whoa! You're not a person. What are you trying to say to me?’” (CG7). Because a VIPA is not perceived as a person, participants felt that they did not need to uphold the social responsibility to “have dialogue or a conversation with it” (OA5). Furthermore, participants perceived certain word choices to be violating their mental models of insentient technology. Common phrases that led to negative reactions from participants include VIPAs expressing an emotion such as being excited (e.g., “I'm excited to work with you for the next 16 weeks.”) and curious (e.g., “I'm curious how doing these activities make you feel.”). One older adult thought that it was “kind of silly for a digital assistant to say, ‘I'm excited,’ because it's a machine. How excited can it be?” (OA4). Similarly, a caregiver commented that the VIPA could not be curious because it was an AI, and it was merely “gathering this data ‘cause [it's] programmed to” (CG2). As such, emotional expression that would anthropomorphize the VIPA was consistently perceived as pretentious. Particularly, participants strongly disliked VIPA's comment of “that's a great name” when participants initially shared their name with the VIPA. Because participants understood that the VIPA “is AI, so that's programmed in there” (CG5), which seemed “kind of patronizing” (CG5) when VIPAs made such comments. Instead, participants chose the more direct response (i.e., “Thank you.”) that proceeded to the next interaction. In the same scenario, some participants chose to anthropomorphize the agent, attributing feelings and human-like colloquial language use to the VIPA. However, differences existed; while some participants felt that the VIPA was “a little too perky, like it's trying too hard” (OA4) and “trying to get on [their] good side” (OA3), which felt inauthentic, one participant felt that emotional expression represents “a more casual type of a statement” (CG4).

Despite critiques toward the VIPA prototypes in some scenarios, participants expressed a desire for the VIPA to use friendly and colloquial languages in other scenarios. For example, when presented with two options introducing behavioral activation (A: “My goal is to help you to learn about behavioral activation technique which focuses on using behaviors to activate pleasant emotions. It has been shown to improve depressive symptoms.” and B: “I'm really interested in learning about how to help people feel more positive and came across this technique called behavioral activation. Want to hear more about it?”), one caregiver commented that she option B was “more colloquial, more conversational, as opposed to a little more formal” (CG5), which was more inviting to use with older adults. An older adult compared the more formal languages to VIPAs “assess[ing them] like a psychiatrist or something” (OA2). Although a more formal language may be more clinically appropriate for delivering evidence-based mental health service, this older adult felt that receiving an evidence-based mental health service through VIPAs should not be the same as speaking with a mental health professional or taking a mental health questionnaire. The desire for informality may be related to participants wanting VIPAs to maintain a natural flow of conversation while keeping the dialogues brief and straight to the point. Most of these interactions stemmed from the participants’ habitual interactions with other people. As a result, when VIPAs greet the users, older adults expect a friendly greeting because that mimics how their typical greeting interaction is like: “when I greet someone, I like to say ‘Good Morning.’ It seems more natural to have that. A sound effect isn't quite there” (OA3).

Some participants displayed contradictions in their conversational preferences. For example, OA33 previously stated that they “don't want to have a conversation with artificial intelligence” (OA1), but later said they “want it to be flowing all the time – real easy conversation kind of thing” (OA1). To resolve this tension, one participant designed their dialogues in a scaffolding approach, starting with questions that were “a little more basic, and not so intimidating when you're just starting to do that” (CG7). Others provided concrete examples of how this could be achieved. When presented with the Tracking and Assessment scenario where the VIPA conducted a progress check with the participant (A: “After you did not complete the activity, have family time, how do you feel?” and B: “I want to check in with you to see how you're feeling after not having family time. Can you tell me how you're feeling?”), an older adult rewrote the dialogue to be “Are you okay with not having family time tonight” (OA5). This aligns with other participants’ preferences for shorter, more direct, and less clinical language that does not suggest VIPAs having emotional expression while maintaining the interpersonal characteristics of a conversation. In sum, while participants anthropomorphized the agent at times, they also recognized the agent was non-human and found the use of overly-human-like language use to be unpleasant.

VIPAs can offer social and emotional support for older adults not only by being a companion but also a listener and a personal entourage, thus lowering the burden and reducing the responsibilities caregivers may experience when caring for their family members. Older adults in our study saw the VIPA as a useful tool for emotionally expressing themselves without worrying about burdening their caregivers. They valued VIPAs’ inability to be hurt and the ability to provide nonjudgmental responses during moments of depression and hardship: “I will snap back at it. It's easier to snap back at it than to snap back to my husband or somebody else” (OA1). Interestingly, OA1 followed up by sharing that they did not actually expect a response from the VIPA, stating that “She's grouchy. Leave the lady alone” (OA1). Other older adults also appreciated the mere presence of VIPAs and did not expect responses from the agent in similar situations. This is because older adults mainly wanted an outlet to express themselves and were aware of VIPA's limitations as a machine. OA5 provided a specific example in which they could not “get in touch with their friends and doing something with them” (OA5). Although frustrated, they explained that they would tell the voice agent about their frustration and what happened, but they “don't expect Alexa to really do anything about it” (OA5). The mere act of telling VIPAs about their feelings act as an emotional outlet for older adults in the study. Because of that, participants chose to have the VIPA simply acknowledge their feelings with generic language (e.g., "Thank you") or no response at all.

Additionally, anhedonia and overall low motivation and energy are common depression symptoms. As a result, depressed older adults may struggle to come up with pleasant activities, which is a challenge to implementing behavioral activation treatment. One of the scenarios (i.e., Curating Activities) in the study included probes that asked the participants to list pleasant activities they would like to engage in, during which one older adult stated that “list[ing] three activities that give [them] pleasure or make [them] feel accomplished” is a “difficult question to answer” (OA6). Instead of placing this additional responsibility on caregivers who may already feel taxed, caregivers particularly believed that VIPAs can support older adults by providing recommendations based on their preferences, physical conditions, and mood data. Namely, caregivers recommended that VIPAs could suggest and provide guided mindfulness activities or “a deep breathing exercise” (CG6) when older adults are “having a very anxious day” (CG40) or simple activities to get the person started” (CG1). Older adults and caregivers felt that this approach could provide social support without suggesting that the VIPA has emotional agency.

Furthermore, in the Tracking and Assessment scenario where the VIPA checked on the user's progress on activity goal-completion, participants often rewrote the dialogues to support emotion processing. Particularly, if older adults reported that they were unable to complete an activity as planned due to depression, responses from VIPAs should be “empathetic” (CG2) rather than being “a homework checker” (CG7). Participants suggested that the VIPA can offer words of encouragement that do not compromise their mental model of the technology. For example, VIPAs could acknowledge the hardship and encourage participants to try again: “Okay. You can try again tomorrow” (OA2). VIPAs could also ask if the older adults would like to hear a joke (“Would you like to hear a joke to cheer up?” (OA3)) or affirmations, such as “Great job” (CG1) and “I'm glad that you did [activity]” (CG7). Caregivers believed that VIPAs could provide social support to motivate older adults to complete activities for mental health intervention. This is crucial to behavioral activation, which uses behavioral goal setting to activate elevated mood among patients who struggle to find intrinsic motivation to engage in pleasurable activities.

Autonomy and privacy are important topics to consider when creating technologies, so special considerations should be placed on preserving these two components in human-VIPA dialogue exchanges. In addition to the existing research on autonomy and privacy in VIPAs, our findings provide additional insight into the similarities and differences in priorities of user autonomy and privacy between older adults and family caregivers, especially when using VIPAs for mental health data collection and data sharing.

One of the main challenges in developing conversational agents is to design engaging conversational interactions that are appropriate for the interaction's purposes and contexts. Researchers across different domains have generally found that anthropomorphism in VIPAs leads to increased engagement. Yet, users’ preferences also depend on the context [66] of the interactions and the perceived benefits of incorporating human-likeness characteristics [41]. Epley also theorized factors such as the need for social connection and perceived similarity can further affect individuals’ anthropomorphism of nonhuman agents [26]. Our findings provide additional evidence and explanation for such variations, such as the unresolved tension of perceiving the VIPA as both object-like and human-like throughout their interactions.

In behavior change contexts, VIPAs that use polite and suggestive languages [104] and high-emotion have been rated more likable [29, 105]. Other studies further suggest that users often view computers as social actors (i.e., CASA paradigm [62]), and emotional connection can lead to increased trust in the computer and its likability [49, 62]. In contrast, our older adult and caregiver participants found emotional and affective expressions to be unnatural of a nonhuman agent, which conflicted with their mental model of VIPAs as objects (i.e., an information retrieval and data collection tool). As such, they described themselves as lacking interest in anthropomorphizing the VIPA. However, because a voice-based interaction closely mimics a natural human-human conversation, specific conversational characteristics for mental health conversations were still perceived as valuable as the situation calls for such traits (e.g., empathy). These two seemingly conflicting perspectives therefore created a dissonance internally among our participants.

Continuity in care is crucial in mental healthcare, which includes not only informational continuity (e.g., access to patient's information), but also management (e.g., flexible goals and services based on patient's preferences) and relational continuity (i.e., ongoing relationship between the patient and their care team) [32]. VIPAs have the potential to provide continuity of care through (1) data collection and data sharing, and (2) social and emotional support. The unique nature of VIPAs being always available and in between being perceived as more “human-like” despite being an “object” [66, 95] allows them to fluidly switch between these two labels, potentially leading to developing a different relationship with VIPAs than with their human care teams.