6 Thematic Analysis of Post-study Interviews
Each study concluded with a semi-structured interview comparing the experience of working with HIRO to working alone (see Section
4.1.4). In these interviews, 33 participants said they preferred to work with HIRO, 14 preferred to work alone, and 9 did not express a clear preference. We extracted a set of themes from our post-study interviews using a thematic analysis conducted via affinity diagramming. Precedents for this analysis method include, for example, Nunes et al. [
56] and Lucero [
55].
To perform this analysis, we first used a third-party service to transcribe the interview audio. Two authors then segmented the transcripts into notes bearing quotes representing the key ideas articulated by each participant. Three authors then collaboratively affinity diagrammed these user notes to identify the eight themes that are presented in Table
2 and described below. This affinity diagramming took place across 5 weeks with the three authors, who were not co-located, meeting in seven virtual collaborative sessions. The notes were divided randomly among the three authors, who each took turns adding them one at a time onto a shared board, where clusters represented thematically similar notes. Authors temporarily labeled clusters as the diagram developed, but would also reorganize if a more compelling grouping emerged as new notes were processed. Disagreements were handled as they arose by voting. At the end of this process, the affinity diagram contained labels at different levels of abstraction, which the authors organized into the eight themes discussed below. Themes are illustrated using representative interview quotes; the speaker for each quote is identified as
\(Px (yP, zR)\) , where
x is the participant identifier,
y is the number of HIRO card placements in their session, and
z is the number of reversals.
6.1 Second Perspectives Provided by the Robot
Participants identified several ways that having HIRO as a second perspective affected the experience of affinity diagramming. Some participants appreciated HIRO as an additional decision maker, especially when they were unsure about how to cluster certain notes. “I felt like I was just going along a straight one-way road when I’m working alone,” said P1 (20P, 2R), “but with the robot, I felt like I had some kind of company with the thinking process and some optional solutions to rely on when I’m stuck.” HIRO also led some participants to explore more alternatives than they did alone. P40 (14P, 2R): “I felt like I was changing a lot more when I was with the robot, […w]hereas, when I was working [alone], I didn’t switch a card once.” Even when HIRO’s perspective wasn’t convincing, evaluating it reinforced their confidence in their own decisions. P52 (16P, R8), despite disagreeing frequently with HIRO’s choices, said “I also feel more secure, because I actually thought more about it.”
HIRO’s second perspective helped participants resolve uncertainty. “Sometimes in my mind,” described P13 (13P, 4R), “I’m not really sure which group this would go to. And then, the robot would decide, and I’d be like, oh, that make sense.” P26 (14P, 4R): “I don’t feel very confident working with open-ended questions by myself. Another person’s decision would definitely help. And HIRO is pretty intelligent to me and I appreciated his help.” This was especially important at the beginning, as P11 (14P, 3R) told us: “I was indecisive at first. […] But with the robot […] it dictated a path […] it was much easier to categorize things in that way.” For others, simply seeing HIRO’s opinion could help clarify their decisions. As P28 (14P, 4R) explained, “When HIRO picked [the note] up, […] just the act of watching it make a decision, I was, like, okay, I actually know where I want this to go.”
At the same time, many participants found it more stressful or less efficient to deal with HIRO’s perspective on the data. Participants complained about having to double check HIRO’s work, inhibiting their ability to work intuitively. Instead of going off “the first thing that popped into my head,” said P45 (15P, 9R), “I had to think harder about it to figure out whether or not I wanted to change the way I initially approached it.” This was particularly challenging when participants frequently disagreed with the robot. P51 (10P, 7R) told us that working with HIRO was a burden, “because I was trying to fix things I didn’t agree with.” Some participants found it more stressful to have HIRO checking on their work. P56 (14P, 4R) remembered feeling judged in HIRO’s presence. As they described it, “HIRO is, like, a genius machine that has been programmed to know how to do this. And I’m just coming in as an undergraduate in college. And so, I felt like when you sit at the kitchen table and do math homework with your dad. And you know that they’re watching, knowing the correct answers, and thinking you’re stupid for making the wrong decisions.”
6.2 Mutual Understanding: Getting on the Same Page with HIRO
Participants perceived varying degrees of mutual understanding with HIRO about the structure of the diagrams they were constructing together. Some felt on the same page as HIRO early on in the process, while others struggled to understand HIRO or felt like HIRO didn’t understand their thought process.
This sense developed over the course of the task. “As the experiment went on,” P14 (13P, 3R) recalled, “HIRO started placing cards in clusters that I started agreeing with more. So, in that sense, I felt like he was learning and paying attention.” For P28 (14P, 4R), mutual understanding developed through collaboration: “We were learning a common language at that point. We were building off of things that can be used to inform understanding. But before that critical mass of information occurs, I did not understand what HIRO was doing.” Some participants never reached that tipping point with HIRO, leaving them frustrated and confused. For example, P6 (18P, 8R), who was excited to work with HIRO, at some point decided: “I’m very confused and I’m trying to generalize the idea behind this category but I have been struggling; so I’m just going to leave HIRO out of this, I’m going to start deciding on my own.”
Participants tended to categorize disagreements with HIRO as either reasonable or inexplicable. Reasonable disagreements, for example, could occur when participants were conflicted about where to place a card. “I moved it,” P3 (14P, 1R) explained, “But I could see why HIRO put it there because I was also kinda debating between those two categories.” When participants failed to parse HIRO’s reasoning, it could undermine mutual understanding, even if they mostly agreed with its placements. “I feel like I didn’t understand HIRO…we agreed on a good amount of items but there were some choices that I was very confused by,” said P2 (15P, 6R).
While repeated disagreements could erode trust, some participants worked to repair gaps in mutual understanding with HIRO. For example, P1 (20P, 2R) described reversing an earlier decision to override HIRO as P1 started to see its point of view, saying, “I thought it was an error, and then it made more sense after I knew what the other cards were and realized the robot was probably right.” When P23 (14, 4R) couldn’t make sense of HIRO’s choices, P23 tried to make clusters clearer, remarking that, “HIRO was pretty much telling me this category doesn’t make sense.”
6.3 How HIRO Affected the Cognitive Load Required for the Task
Despite the minimal differences measured by NASA TLX, participants described increased or decreased aspects of cognitive load when working with HIRO. As discussed in Section
6.1, it could be burdensome to check HIRO’s work or more stressful to think with HIRO watching. Participants also devoted effort to ensuring that HIRO could see and understand how they were organizing the cards. This manifested at a physical level, either in arranging cards, e.g., as P44 (12P, 6R) said, “myself, I could make the organization make sense in my head, but [with HIRO] I wanted to make sure it was clear so they could detect the different clusters,” or in sharing the workspace, e.g., as P23 (14P, 4R) said, “I had to make sure to keep my hands back because HIRO was trying to read.” At a conceptual level, P37 (16P, 5R) worked to make sure that HIRO could interpret P37’s clusters, reasoning that “I wanted the robot to implicitly recognize the themes…so this motivated me to have ones that are really similar together.”
On the other hand, some participants felt that simply having HIRO place some cards reduced their workload. P36 (9P, 2R) reasoned that HIRO “basically [made] the same decisions I would have made. I could have had the same outcome without him but maybe it was easier on me and less decision fatigue.” For some, working with HIRO reduced overthinking. “Alone…there are different aspects I would rethink over and over again,” explained P20 (9P, 2R), “compared to if the robot would just take the card and place it in a particular stack [and] I would try to think of the why and not.” For others, HIRO offered a form of automation, as per P1 (20P, 2R): “the robot just became more of a time-saving machine. It’s, like, a way to double my bandwidth.” In the end, P30 (13P, 5R) mused, “It comes down to, do you want to work on a group project by yourself? Or do you want to use it with somebody who is helpful to you and that you can try and figure it out together? I think just cognitively it helps you take a load off, right?”
6.4 How HIRO Affected the Pacing and Flow of the Task
In line with our quantitative findings, several participants noted that working with the robot slowed down the pace of the task. Some attributed this to the physical delay in waiting for the robot to find, pick, and place a card. As P11 (14P, 3R) described it, “You had to wait for the robot a lot. Sometimes, it would take a while to go and get the card and figure out where it wanted to put it.” Alternatively, P2 (15P, 6R) explained that thinking things through with HIRO took more time: “If HIRO put a card down…that I was unsure of, I would think about it more deeply. Whereas when you’re working alone, you’re, like, this is my thought. So, that’s why it’s faster.” In contrast, some participants actually found the robot to move too quickly, in a way that rushed them or disrupted their desired workflow. P15 (13P, 6R) told us, “When I work with the robot I feel a bit tense because I felt like I have to keep up with its pace.”
Beyond pacing, some found that working with HIRO changed the flow of the task, forcing them to process the cards incrementally. “I think working alone was a bit easier because first I got to read all the cards before deciding what clusters to make and while working with the robot I had to decide as it happened, as we placed the cards, which clusters to create,” explained P8 (13P, 3R).
While several participants preferred the faster pace when they worked alone, some participants found value working at a slower pace with the robot as it offered them more time to reflect on their decisions as well as to plan ahead as the robot was working. “Definitely, it takes more time to work with a robot, which is good because I think it allowed me to re-examine my decisions,” said P16 (26P, 8R). “I had more time while the robot was working to conceptualize the different sections that were going on,” said P40 (16P, 4R).
6.5 Enjoyment: The Experience of Working with HIRO
Participants found certain aspects of working with HIRO to be enjoyable in their own right. For some, working with a robot was simply more engaging. “I don’t really know why that is, but it was fun to work with the robot,” said P5 (15P, 4R). Others felt a sense of social connection to the robot. “It was friendlier,” said P3 (14P, 1R). “I don’t know how else to put it, but it was more enjoyable.” “I feel like we connected. I feel like we’re friends. I would be very sad if HIRO doesn’t like me,” said P52 (16P, 8R). Finally, for some, working with HIRO took on an element of mystery, and trying to understand HIRO added an interesting dimension to the task. “I would use it sort of like a game. I would see how it interacted. It would be more fun for me to interact in that way,” said P53 (18P, 2R).
6.6 Nonverbal Communication: Limits in Creative Collaboration
HIRO was designed to communicate with participants through placements in the shared affinity diagram and had no ability to communicate verbally. To some, the lack of verbal communication made the mechanics of collaboration more challenging, specifically wanting support for explanations and debate.
Participants desired to both explain their reasoning to HIRO and hear HIRO explain its reasoning to them. “I can’t talk to the robot and explain my thinking,” said P8 (13P, 3R). “And if HIRO was a human, I would ask them what they were thinking.” P7 (14P, 7R) put it a bit more bluntly, saying, “I don’t know its thoughts so I cannot brainstorm with it.” HIRO’s inability to communicate the rationale behind its choices led some to reject it as a collaborator. “I feel like because HIRO can’t express their opinion,” P21 (14P, 2R) told us, “I can’t really come to a compromise, so I decided to just go with what I thought. So it’s not very much a collaboration.”
Participants felt that verbal communication would allow for more argumentation between human and robot. “With a person,” reasoned P4 (12P, 1R), “you’d be constantly bouncing ideas back and forth and trying to come to some sort of middle ground that you both agree on,” whereas, as P37 (16P, 5R) put it, “The robot is just sort of saying, ‘Here,’ and you can take it or leave it.” This placed the burden of handling disagreements on the human. P17 (11P, 5R) explained, “[humans] can actually exchange information if there’s a disagreement or if any modifications need to be made…but with HIRO, it’s me handling the situation.” Overall, P18 (21P, 12R) argued, “a collaborator should demonstrate its argument.”
Participants, however, were split on whether they would rather work with an agent that could argue with them or not. Some participants appreciated the ability to overrule HIRO without having to argue with it or account for its feelings. As P9 (14P, 3R) told us, “From my experience working with other groups, there must be a compromise when there’s idea conflicts. And, for me, it’s very inefficient during the working process. But for the robot…I noticed that we do not have conflicts.” Some described HIRO’s muteness as a good fit for their personality and collaboration preferences. As P36 (9P, 2R) put it, “I’m a little introverted and he doesn’t talk back but he gets the job done.” That said, silence could be socially jarring, e.g., as P12 (15P, 6R) put it, “It feels like my thoughts, my worries, my insecurities are just put out there and I just don’t have anyone to reassure me or guide me in the right direction.”
6.7 Perceived Intelligence: Can a Robot Understand Human Issues?
Participants’ mental models of HIRO could color their interactions with it. Participants rated HIRO as relatively intelligent (see Figure
13). However, assumptions about how HIRO was processing information shaped expectations about how it could assist on this task. For example, some differentiated the levels of insight at which an algorithm and a human could characterize human needs. “I think it went about as I expected it to go, a robot sorting human concerns,” said P5 (15P, 4R), “people obviously innately understand human struggles more than many robots are capable of.” Occasionally, HIRO’s decisions challenged such assumptions. P12 (15P, 6R), for example, believed that, “this context as a human that you have surrounding each of these comments…HIRO might not have that sort of context.” However, they told us, as the study went on, “a lot of the choices [HIRO] made, I could reason myself and get to the same place he did. So, there was more thought I guess, and less algorithm.”
HIRO’s movements also influenced participants’ perception of its intelligence and agency, for example, the way that HIRO scanned the diagram before placing a new card. “It was almost like watching someone and sort of following their thought process,” said P22 (7P, 2R). P12 (15P, 6R) suggested, “It’s very mechanical, but also I could see a human quality kind of like pondering as he went.” “Sometimes it feels as though it’s fast. Other times it feels like it’s slow,” said P21 (14P, 2R), “sometimes it feels enthusiastic. Other times, it’s like taking time to process and be careful.” P36 (9P, 2R) interpreted HIRO’s movements in a collaborative sense, “like he was taking a second to think about how I was thinking before placing his card.” Not everybody felt the same way. P1 (20P, 2R) and P31 (7P, 2R) expressed confusion as to why HIRO was moving over the diagram or moving in between turns, respectively, whereas P18 (21P, 12R) found it distracting if HIRO moved when P18 was trying to think.
6.8 Roles and Power Dynamics: Making Creative Decisions with HIRO
Participants attributed different collaborative roles to HIRO. Some viewed HIRO as a reference tool to check their thinking or to offer ideas when needed. For example, P1 (20P, 2R) said that HIRO, “gave me more possible solutions or options around which way I should go with organizing.” Others allowed HIRO to assume more control, including delegating clustering decisions. P9 (14P, 3R) explained that, “when I make sure that it’s in the same logic with me, there’s no need for me to carefully read every card. And it’s automatic.” At an extreme, P43 (13P, 0R) told us that if given the chance to work with HIRO again, P43 would let HIRO “place everything and just play on my phone, and then after HIRO does it, I can just check its work.”
Oftentimes, roles split across defining categories and sorting cards. P23 (14P, 4R) described HIRO as being “constrained to what I had set forth,” and “I was the sole person making categories, then [HIRO is] somebody who’s just helping me sort things.” Others simply preferred the role of creating categories. As P27 (12P, 3R) said, “I don’t want the robot to start making new categories for me, because I kind of wanted to make the categories.” Some participants, however, used HIRO’s input to define categories, e.g., P5 (15P, 4R): “[HIRO] would put things down and instead of having to generate categories myself I could either agree or disagree with the robot.”
Sometimes, roles reflected power dynamics with HIRO. P16 (26P, 8R) described deferring to HIRO: “when I first started doing the card sorting, I definitely saw our power dynamic; the robot had more power and knowledge in this area.” For P15 (13P, 6R), power imbalance inhibited collaboration: “I think it’s smart. So I don’t want to challenge it. Sometimes I think it places the card in the wrong place, but I’m not sure if I should move it. So I prefer working alone.” Others felt more comfortable overriding HIRO as they saw fit. “I think my reasoning trumps his,” P12 (15P, 6R) explained, “but not in a disrespectful way.” In the end, some participants observed that HIRO’s limitations, physical or otherwise, gave more power to the human. “I think ultimately the human will always have at least just a little bit more control,” P37 (16P, 5R) told us, “because the robot can move it so many times but I can also just move it…a lot quicker than the robot.” Sometimes, this came with emotional consequences. P52 (16P, 8R) worried about hurting HIRO by rejecting its suggestions: “It felt like I was offending him, because it can’t defend itself, right? It can’t be like, ‘No, actually, I’m going to put it back. So I’m kind of the decision power here.”
7 Discussion
We have described a system and study that we developed to research human–robot collaboration on a sensemaking design task, specifically, need-finding through an affinity diagram. We were interested in three questions: how working with HIRO might influence how humans construct affinity diagrams, how affinity diagrams perform as a human–robot collaborative medium, and what to consider when designing robots to support activities such as affinity diagramming. We start by discussing the first two of these questions, followed by a consideration of the unique context of our participants’ limited experience with affinity diagramming. We then present a set of implications for the future design of human–robot collaborative design systems.
7.1 How Did Working With a Robot Influence How Humans Constructed Affinity Diagrams?
Our findings suggest several ways that HIRO’s presence affected participants’ affinity diagramming experiences. Participants tended to work more slowly with HIRO. While this could be attributed to the speed of the system or the overhead of collaboration, it is worth noting that some participants spent
less time on the task when working with HIRO (see Figure
8). Further, despite the turn-taking dynamic, the division of time between human and robot effort was not a zero-sum game: several participants described thinking about or working on the task while HIRO was moving. While they did not report overall differences in cognitive load, several participants described ways that HIRO affected their cognitive load, stress, enjoyment, or social engagement. Working with HIRO could inspire participants to consider alternative interpretations of the data. Sometimes, this increased perceived effort and the difficulty interpreting choices could be frustrating. HIRO could also mitigate uncertainty about individual cards or the overall direction of the diagram.
These findings reflect prior studies on how interaction with other humans affects cognition and creativity. Tversky and Hand [
67], for example, found that the mere presence of a human actor in a photograph encouraged people to adopt the actor’s perspective when describing objects in the photo. Paulus cites findings that suggest that exposure to others’ perspectives can increase individual creativity [
57], affording a wider base to explore ideas, commingling cognitive styles, and applying heterogeneous knowledge sets to a problem. At the same time, group social dynamics such as productivity blocking, evaluation apprehension, and free-riding can inhibit creativity [
16].
That said, while participants saw HIRO as intelligent and likeable, they tended not to perceive it as anthropomorphic. Some participants believed that a robot could never understand human concerns, and participants were frustrated that HIRO was unable to explain and defend its ideas like a human. This mirrors the argument by Guckelsberger et al. that creative agency in machines requires not just creative acts but also explanations that reflect and maintain a creative identity [
26]. This remains a significant challenge for human–machine collaborative design.
7.2 How Might Affinity Diagrams Support Creative Collaboration Between a Human and a Robot?
The context we investigated here, affinity diagramming for need-finding, is unexplored in the human–robot creative collaboration realm. How did the medium of shared note sorting perform in the design process?
As in human collaboration in affinity diagramming, participants usually used spatial position to communicate with the robot about the relationships between notes. Even when they disagreed with HIRO, participants attributed conceptual meaning to where the robot placed the cards. In many cases, they constructed mutual understanding with HIRO about how to interpret the data over the course of several card placements.
That said, we observed clear communication limitations to collaborating through the diagram alone. In a few cases, participants were confused by ambiguous card placements. Participants also disliked the unequivocal nature of HIRO wordlessly placing cards, wanting to hear its reasoning or to negotiate with it. This, combined with uncertainty around how HIRO was thinking, could restrict the degree to which participants felt like they were collaborating well with the robot. In short, the affinity diagram was sometimes effective at communicating opinions, but not reasoning, and HIRO’s movements over the diagram expressed that it was thinking but not what it was thinking.
Ultimately, while the shared affinity diagram did not consistently fulfill participants’ collaboration needs with HIRO, we also saw glimpses of how such a diagram could serve as the engine of mutual understanding in a highly unstructured task, such as need-finding.
7.3 Interpreting Participant Experiences Through a Lens of Task Expertise
On reflection, several elements of the themes we observed reflect novice design behaviors that may have emanated from our participants’ relatively low self-reported experience with affinity diagramming.
Novices and experts are known to exhibit different behaviors, and much work has gone into characterizing these differences in the context of unstructured problems. Cross’s comprehensive survey of expertise in design [
12] finds several important tendencies that characterize experts in early-stage design work: experts are solution focused and use conjectures to scope problems early on; experts tend to adopt to and stick to early design concepts rather than exploring many alternatives; and, finally, experts are opportunistic in their methods and frequently switch between parallel cognitive activities.
The divide between the tendencies that Cross attributes to expert designers versus novice designers surfaced in our participants’ stories and the themes we extracted from our interviews. For instance, several participants remarked that HIRO gave them a direction to explore or injected objectivity into the task, suggesting a lack of willingness or ability to conjecture from experience. Participants also noted that working with HIRO required a collaborative overhead compared with working alone. This could be distracting, preventing participants from following “the first thing that popped into [their] head” or applying a more top-down structured process. In contrast, others noticed themselves exploring different options or getting un-stuck by watching HIRO make a decision, evoking the kind of cognitive switching that an expert might already use effectively in unstructured problem-solving.
Experts who were more familiar with this particular design activity may have had very different experiences working with HIRO than our participants. For example, switching between different ideas or ways of thinking could be more distracting for an expert who is already thinking along parallel tracks. Likewise, using HIRO as a proxy for conjecture and problem-scoping would likely not have appealed to an expert with the accumulated experience to make those decisions confidently and intuitively. While HIRO making choices could offer a safe sense of direction to someone who feels uncertain, the same behaviors could be burdensome for an expert user who would prefer control and flexibility to opportunistically define their own path. On the other hand, an expert’s solution-oriented focus might provide more useful context to the robot compared with a novice more fixated on the present.
In sum, we might expect experts at tasks such as affinity diagramming to look for different kinds of help than novices would in ways that are more tailored to their own processes rather than broadly suited to the general design activity at hand.
7.4 How Should We Design Robots to Support Conceptual Aspects of Designing?
Based on what we learned about affinity diagramming user needs with a robot, we propose the following guidelines for designing robots to support similar conceptual design activities, particularly with novice users.
(1)
Account for the robot’s speed. Consider how fast a robot moves when determining the roles it plays to support design activity. For various reasons, participants in our study tended to spend more time on their diagrams when working with HIRO than when they were working alone. One factor that contributed to this was the perceived need to wait for HIRO while it was making a move. For at least one participant, the need to take turns stifled the individual’s preferred flow for running through all the cards before choosing placements. Others, in contrast, suggested it moved too quickly, disrupting them or making them feel tense.
Our findings suggest that the perceived effects of a collaborative robot’s speed may influence the flow of a creative activity in nuanced ways. By conventional human–robot fluency metrics such as idle time [
33], a slow robot should be judicious about undertaking tasks with long-running motions. That said, idle time can also offer a human partner time to stop and think while sharing the floor. For some of our participants, the difference in pacing from working alone provided opportunities to re-examine their choices or conceptualize the meaning of the current clusters. Overall, for different participants and purposes, the speed at which a robot moves could have both benefits and challenges in terms of cognitive collaboration.
This suggests that the speed at which a robot is designed to move should align with the needs of a specific human collaborator and task. Of course, technical or safety limits may constrain the degree to which a robot’s speed can be adjusted. With this in mind, an alternative approach is to consider the kinds of tasks the speed at which a robot can move are suited for.
This guideline connects to the different roles participants gave the robot. For instance, a slower robot might be better suited to define clusters or handle difficult cases, whereas a faster robot might be more appropriate for automation or tasks that can be completed asynchronously without the human’s attention.
(2)
Pursue mutual understanding in creative collaborations.
Second, consider the system design goal of mutual understanding rather than more straightforward goals such as agreeing with the user or exposing the user to alternatives. Getting on the same page with HIRO was a process that informed many participants’ experiences. One participant described the experience as building up to speaking a common language with HIRO, something only possible once some critical mass of shared experience had been achieved. Participants’ willingness to engage in this process suggests several opportunities, most notably for novices at an unstructured task.
For example, establishing mutual understanding with a robotic partner could provide a gentle entry point for exploration. Our interviews revealed that some participants credited HIRO with providing early direction or a sense of objectivity. That said, simply providing initial directions could inhibit a novice’s willingness to explore more deeply [
9]. The process of developing mutual understanding with HIRO suggests a better alternative. Many participants described testing, rationalizing, and rebutting HIRO’s choices, with a desire to engage them more discoursively. If the system prioritizes this process of developing mutual understanding with a tool such as HIRO, it could offer low-stakes incentives to nontrivially engage in speculative directions for those without the confidence that comes with expert intuition.
What might this look like? Bratman describes shared cooperative activity as rooted in mutual responsiveness, demonstrated commitment to the joint activity, and commitment to mutual support [
10]. Through utterances or actions, a robot might might explicitly signal its commitment to developing mutual understanding with a user beyond simply finishing the task. For a robot such as HIRO, actions such as frequently inspecting the human’s clusters before making a move or identifying and pointing out clusters that it finds less cohesive might encourage the human to reciprocate and, ultimately, support both more collaboration and exploration.
(3)
Identify opportunities for constructive disagreements.
Beyond the dynamics of developing mutual understanding between a human and a robot, our findings suggest that there may be opportunities for a robot to challenge humans in a collaborative sensemaking task. Several participants described needing to think more about their choices or changing their mind when working with HIRO. One of our participants described needing more
surprise from HIRO to see it as a compelling creative partner. This resonates with prior work in computational creativity support, including several projects that suggest ways to encourage creative shifts (e.g., [
42,
54]) or explore formalizing surprising ideas [
24], pointing to the usefulness of constructive disagreement.
That said, not all disagreements are equal: our participants distinguished between disagreements they could rationalize and ones that didn’t make sense to them. In some cases, the content of the disagreement could be immaterial: one participant told us that simply seeing a decision from HIRO helped the individual to reach one’s own, regardless of agreeing or not. In short, characterizing the role of disagreement in a creative collaboration with a robot is multifaceted.
Unsurprisingly, this mirrors the complexity of conflict in human creative collaborations. Badke-Shaub et al. found that design teams that were relatively more confrontational and less collaborative tended to generate more functional and innovative ideas, although they still exhibited mostly collaborative behavior [
5]. However, conceptual conflict can escalate to damaging affective conflict in teams [
3,
64], a relevant design consideration for any collaborative robot with social behaviors. We saw elements of this in participants who perceived a social connection or power dynamics with HIRO that affected their willingness to override it when they disagreed with HIRO. This kind of dynamic could present both a constraint and a degree of opportunity to push creative boundaries within reason, using the social connection as a kind of buffer to affective conflict. Others appreciated that HIRO was non-confrontational, offering a second perspective without apparent social consequences. Finding the right balance of creative agreement and confrontation between a human and robot in an unstructured task demands attention to what constitutes constructive or destructive disagreements in each context.
The context of physical collaboration with a robot adds a particularly salient dimension to this discussion. Klemmer argues that embodiment carries risk, because choices are more visible and physical actions express commitment to those choices [
45]. This sense of risk is an interesting lens through which to view our findings around second perspectives and power dynamics, whether participants felt hesitant to override HIRO’s choices or simply anxious feeling their own actions were being scrutinized. Klemmer et al. point out that this riskiness can work both ways: increasing focus and attention on the task but also constraining willingness to think divergently. Beyond how participants perceived risk in their own actions, the visibility of HIRO’s actions also carried risk in terms of how participants perceived its intentions, intelligence, or helpfulness. In literally changing a participant’s arrangement of cards, HIRO expressed a commitment to a particular interpretation of the data that demanded a response. In a sense, the purely physical and uncompromising nature of HIRO’s behaviors amplified this risk—it had no way of, for example, lowering the level of commitment by verbally expressing uncertainty about a placement. This leads into our final design guideline, which addresses the nature and limits of communicating solely through shared physical materials.
(4)
Use other modalities of communication in conjunction with physical materials.
Participants’ desire to converse with HIRO reflected limitations on what HIRO was able to communicate nonverbally. While the diagram communicated relationships between notes, participants wanted to discuss the motivations behind choices that they or the robot made. To support this, we suggest using physical representations such as affinity diagrams in conjunction with other modalities that can add depth when needed.
While verbal dialog that includes explanations and debate is a straightforward way to add nuance, this is not the only solution. For example, one participant suggested highlighting key text on cards. HIRO could also use gestures to broaden communication; Heiser et al. describe teammates gesturing over maps to support collaboration and cognition [
32]. Some of our participants gestured with or over note cards, and many interpreted meaning in HIRO’s movements over the diagram. Beyond pointing gestures, the functions of metaphoric and iconic gestures in collaboration and creativity have been studied in human teams [
53,
71]. While less explored in HRI ( e.g., [
36]), metaphoric and iconic gestures may be particularly useful to ground creative exploration in human–robot collaborative design.
A physically shared diagram may afford forms of communication beyond the intended conceptual organization of the diagram itself. For example, while coding placements, we observed participants covering a cluster to hide it from HIRO or placing cards between clusters to indicate indecision. Overall, there is a rich interaction space to explore at the boundaries of what material representations explictly afford, a space that might yield more intricate human–robot dialogues of the sort that arise in unstructured problem-solving.