Unmaking-with AI: Tactics for Decentering through Design

Over the past decade, unmaking and more-than-human design (MTHD) have emerged as valuable approaches for designers and researchers in Human-Computer Interaction (HCI), establishing themselves as promising avenues for generating critical responses to pressing environmental and social challenges [27, 39, 62]. While unmaking and MTHD have distinct corpora and approaches, they share common ground and motivations. At the core of both approaches is the proposition that to adequately engage with the current environmental and social challenges, designers have an ethical duty to “critically rethink the modern, colonialist, and anthropocentric inheritance that resonates in and through design cultures” [33, p. 2]. Rooted in feminist theory, both unmaking and MTHD approaches seem to belong to a broader paradigm shift in HCI known as the posthuman turn [27]. As such, they can be understood as approaches that converge two posthumanist critiques. On the one hand, they challenge human exceptionalism by expanding the focus of design to material processes and nonhuman agencies. On the other hand, they contest modernist conceptions of ‘humans’ built around the rhetoric of users and progress, centering white, male, and able bodies [7].

Exploring the resonances between unmaking and MTHD, the article is organized around two questions posed by the editors of this issue: “How can unmaking suggest new standpoints such as more-than-human thinking?” and “How can it support new ways of seeing and imagining technology?” To answer these questions, we build upon the reasoning articulated by Sable and colleagues [60], who provocatively asked, “What if we recognize unmaking as a design move for agonism?”. We pose a parallel inquiry: What if we recognize unmaking as a move for MTHD? More specifically, our exploration delves into the potentialities of aligning unmaking and decentering, an MTHD practice that shifts attention from traditional perspectives to marginalized voices [47].

To rehearse this alignment, we extend Lindström and Ståhl’s [40, 41] concept of the double movement in un/making to decentering. Through that framework, we analyze the outcomes of a series of workshops conducted in 2020, which invited participants to decenter the figure of the user in AI systems and decenter themselves when engaging with AI-enabled conversational agents (CAs) [45]. Using the double movement, we elucidate what was unmade/decentered and what was made/accounted for during the workshops. By exploring the resonances between these practices using a research-through-design (RTD) approach, we address two gaps in HCI scholarship: the imperative to delineate the epistemological scope of unmaking [62] and the urgency to comprehend how posthumanist thinking can be enacted through design [47]. Furthermore, our investigation introduces novel nuances to both practices: it articulates the alternative imaginaries that emerge from unmaking [39] and portrays decentering as processes involving material movements [47].

Following this introduction, the article unfolds as follows: Section 2 delineates the theoretical developments that frame our contributions, reviews pertinent literature from unmaking and MTHD, and explores potential synergies between them. Section 3 examines these synergies within our design case (focusing on AI-enabled CAs). Building on this, Section 4 introduces two key contributions: first, the concept of “unmaking-with” as an emergent notion to articulate posthuman unmaking practices, and second, a trio of decentering tactics to enable these practices: situating, materializing, and enacting. Section 5 revisits the research questions to discuss how this emergent notion of unmaking can offer new ways of envisioning technology and enriching MTH thinking and practice.

While the article seeks to reconcile two theoretical concepts, the emphasis on a practice-based approach is crucial. As designers and researchers active in and beyond the realm of HCI, we acknowledge that unmaking and decentering serve not only as theoretical constructs but also as practical methodologies for designers. Therefore, our aim is to delve deeper into understanding how HCI designers use and expand these concepts. Similarly, while the article can be seen as aiming to align two distinct approaches, the alignment we are concerned with is best seen through the lens of diffraction, a notion that the feminist theorist Karen Barad uses to describe an approach that attends to “relations of difference and how they matter” [3, p. 71]. That is, while we refer to “alignment,” we are not trying to bypass the differences but instead, working towards a space where the two practices are entangled to the extent that novel interferences can emerge. Akin to the way diffracted light creates a spectrum of possibilities from a single beam, we hope the article illuminates the nuances and potential synergies between unmaking and MTHD, revealing new perspectives and pathways for HCI design and research.

The intersection of unmaking and decentering, as an MTHD practice, is particularly intriguing because both practices seem to involve a form of movement. In these practices, designers move away from an established position and gravitate toward a new one. Before exploring this resonance further, we will unpack the theoretical underpinnings of unmaking and MTHD and how they have been used in the field of HCI.

During the HCI conference Designing Interactive Systems in 2020, we hosted the workshop series titled “More-than-human Design and AI: In Conversation with Agents” [46]. The aim of the workshop was to bring together designers and researchers from various fields and regions worldwide to collaborate on designing with AI. The innovative aspect of these workshops was the experimental technique of conducting unstructured interviews to CAs, i.e., asking questions to them directly as well as enacting speculative alternative responses (for more details and outcomes, see the workshop documentation website [81]).

We chose to focus on CAs because it presented a compelling context for exposing anthropocentrism and a fertile ground for exploring more situated and inclusive designs. The field of conversational AI has been growing rapidly, driven by advancements in natural language processing and large language models. Despite the recent buzz around ChatGPT, the field of conversational user interfaces has been an area of research in HCI for many years [82]. The research in this area includes studies on how CAs are embedded in everyday life [35, 57] and how people interact with them [43, 63, 79]. These studies highlight some of CAs’ opportunities but also many risks [1, 8, 11, 36, 70–72, 77], such as CAs’ potential to reproduce gender and racial biases [16, 32, 68, 78]. In relation to these risks, HCI scholars have made significant efforts to explore how CAs can be designed otherwise [37, 65, 68, 78]. Strengers and Kennedy [68] have reviewed some of these efforts in their book “The Smart Wife.” The book examines the historical development of CAs and how they were conceived to take on domestic roles traditionally assigned to human wives. They describe how the gendered character of CAs today, which is designed to be friendly, flirty, docile, efficient, and occasionally glitchy but perpetually available, reinscribes these outdated and unfounded stereotypes. Our workshops expanded this critique by investigating how the discriminatory biases in the design of CAs might intersect with anthropocentric tendencies.

Understanding the anthropocentric tendencies of technologies is complex, primarily because, as scholars from different fields have noted, anthropocentrism is not just an ordinary human bias but a cultural agenda tied to dreams of progress through modernization [40, 74]. Another challenge is methodological. While human-centered techniques are valuable for understanding how technologies are used, they often fall short in addressing the increased agency of technologies like AI and the broad ecosystems they are a part of [21]. This is important in the case of CAs because with their tangible presence (and their often-female voice), they transcend mere devices and become things that live with us. In contrast to their domestic perception, these devices are developed within extensive networks of data, labor, and profit: “While consumers become accustomed to a small hardware device in their living rooms, or a phone app, or a semi-autonomous car, the real work is being done within machine learning systems that are generally remote from the user and utterly invisible” [13, p. 17].

To address these challenges, the workshops took a RTD [2, 67] approach, combining MTHD methods (Thing Ethnography [25] and Interview with Things [9, 58]) with a Speculative Design tradition [12]. This tradition has been valuable for pointing out the ethical considerations designers need in the context of CAs and for exploring alternatives [14, 53, 54, 59, 65]. More specifically, our workshops aligned methodologically with the approach of Lee and colleagues [37] and Parviainen and Søndergaard [53] in using CAs as probes for co-speculation.

Every workshop’ session had 9–15 participants. Due to COVID-19 restrictions, all the sessions were facilitated online through Zoom. We asked participants to bring one or more CAs if they had one, thus, there were a similar number of CAs as people in every session, including Amazon’s Alexa, Apple’s Siri, and Google’s Home. Participants had experience with CAs in various degrees. While some participants had no experience, others had been researching these devices [14, 54, 65]. Given that participants owned these devices, they were partly configured within homes. For example, some CAs were connected to participants’ accounts in applications for playing music and knew their local geographical position.

As shown in Figure 1, every session was divided into different activities. In the first three sessions, we did the same design activities but iterated on the outcomes from the previous ones. The last session included a broader discussion of the outcomes and a speculative design activity to understand how the insights gained could be used further. To facilitate introductions, we started every session with a simple exercise called “Some-Thing in Common,” where the next person to speak was anyone who identified a shared connection with the current speaker.

After completing the introductions, we began with our first activity, “Interview with Thing-Like Humans.” In this activity, participants (in groups of 3–5) were asked to role-play (impersonate) their own CA—or to speculate on one. Based on that experience, participants identified and reflected on emergent themes. These themes were used as starting points to question the CAs in the second activity, “Interview with human-like things,” in which participants interviewed real CAs (Alexa, Google Home, and Siri). From this second activity, the participants selected the three most provocative or surprising questions and added these to a co-created questionnaire for CAs. All these selected questions were presented in the next session, with different participants, to be used as starting points for their inquiry. In that way, the questions in the final questionnaire, featured in Figure 2, resulted from several rounds of iterations.

Beyond the questionnaire, which was the envisioned outcome, other outcomes from the workshop had similar, if not more, richness. We produced a video from the recordings of the first activity and three videos from the third one (see Figure 3). The latter probed the questions with real agents—Amazon’s Alexa, Google Home, and Apple’s Siri. Moreover, in the fourth session, we invited all participants to prototype a speculative agent with objects in their immediate surroundings and have a conversation with them (see Figure 4). In this conversation between the speculative agents, participants pointed their cameras to the agents and impersonated them. To initiate the conversation, we began with the same prompt as in “Some-Thing in Common.” We requested one agent to introduce itself, then invited any other agent with something in common to respond. The conversation evolved organically from this point, as it is shown in the snippet below.

Analyzing the double movements of unmaking and decentering during the workshops revealed several insights. On the one hand, the activities unmade anthropocentrism by exposing that CAs’ responses lacked situatedness. Beyond their inability to understand their context, CAs were limited in revealing their positionality and accounting for their failures, i.e., how the knowledge they produced was linked to broader power structures, materials, and resources and how their workings and failures had implications for different humans and nonhumans. Beyond revealing the anthropocentric biases of CA’s, the activities challenged the designers’ standpoints, decentering their traditional human-centered perspective. In this process, participants were encouraged to reflect on their own subjectivity, positionality, and biases.

On the other hand, the activities made more-than-human imaginaries, interactions, and bodies of CAs. The reimagined CAs differed significantly from traditional ones. Instead of being friendly, flirty, docile, efficient, occasionally glitchy but perpetually available [68], speculative CAs acknowledged their limitations, interrupted, had their own points of view, asked a lot of questions, refused to serve humans, and were only available in certain conditions. Instead of presenting neutral facts, which should be understood as the ultimate truth, speculative CAs were uncertain, quiet, volatile, and made mistakes. Furthermore, they openly disclosed their interconnected relationships within the intimate domestic settings where they operated and in the broader ecosystems and human/nonhuman systems that supported them. The new imaginaries of CAs were far from neutral; they embodied diverse forms and incorporated distinct materials, values, and voices.

The main difference between the reimagined CAs and the existing commercial ones was not that they were more anthropomorphic nor that they were more technical. The main difference was that the reimagined CAs were more situated: they reflected their positionality and acknowledged the position of their users. Moreover, the knowledge CAs (co)produced with humans was also situated, as it was connected to the context in which the agents were embedded, what they did or did not know, and how they were made—including the relations of humans and nonhumans involved in sustaining their infrastructures.

Thus far, we have utilized the dual movements of unmaking and decentering to analyze and annotate the workshop series. Building upon this analysis, we have proposed unmaking-with as a practice within MTHD and presented a series of decentering tactics to support it. These were further explored through two design experiments conducted subsequent to the workshops. Reflecting on these insights, this section revisits the article’s research questions, discussing the opportunities and potential challenges associated with practicing unmaking-with.

One question the article initially aimed to address was “How can unmaking suggest new ways of imagining the role of technologies in HCI?” The workshop has provided one example of how unmaking, coupled with decentering, could help HCI researchers and designers to imagine CAs otherwise. By intersecting discriminatory biases with anthropocentric tendencies, the workshop activities suggested that the design of CAs did not only perpetuate gender stereotypes but also lacked situatedness: CAs often failed to address critical questions about where they were made, who owned them, and what data they used and collected. Conversely, they did not account for how the knowledge they reproduced was entangled in broad infrastructures of power, materials, and resources, including humans and nonhumans.

Generalizing the experiences from the workshop to AI, the insights suggest that AI agent’s bodies and co-produced knowledge must be situated for designing more responsible interactions. Situating AI interactions emerged here as a practice of accounting for the positionality of users, agents, and designers. A lack of situatedness can pose two significant risks. First, not situating AI interactions can compromise Explainability, i.e., keeping the failures and infrastructure of AI systems hidden or in the background can prevent people from developing their own sense of trust in AI applications. Secondly, since the seemingly objective design of CAs relies on a humanist definition of humans, it can compromise inclusivity, as it may disregard the perspectives (and voices) of humans and nonhumans that are inadvertently categorized as “others.” Unmaking-with can support responsible AI development by revealing the social and political structures and biases that shape the design of AI, as well as problematizing the limitations and biases of designers and researchers in envisioning AI differently. Moreover, it can assist designers and researchers in overcoming these limitations by aiding them in reimagining AI—not necessarily emulating human-like or machine-like interactions or forms but adopting a relational approach that lets MTH bodies and relations to emerge. Ultimately, by giving a voice to speculative imaginaries, unmaking-with can do more than reimagine AI differently; it can help rehearse MTH relations and thus support care and response-ability.

The second question the article aimed to address was “How can unmaking support MTH thinking and practice?” Through the article, we have shown one way in which unmaking can actively support MTHD. Through design and diffractively, we have tried to align those two communities of practice. Unmaking’s double movement gave a particular shape to our workshop outcomes and added more nuances to the practice of decentering through design. For example, the emphasis on the process that unmaking provided, further supported the idea that decentering should not be understood as an ultimate goal but a constant practice [24, 42, 47] and that decentering through design is not just a change in perspective but, ultimately, a tangible and material move [47].

While our initial proposition was that both approaches might involve a double movement, in light of the analysis, we can consider how this movement is often set in motion through an active engagement with materials, i.e., how materiality and movement are intrinsically related in both unmaking and decentering. By engaging in creative activities like prototyping and speculating, designers dismantle human-centered conceptions and ways of doing and expand their scope to consider new perspectives. The materiality of the double movement seems to be one way in which designers reorient their efforts toward posthuman outcomes. Acknowledging this process may aid designers in recognizing that what emerges after unmaking embodies a new perspective. This requires designers to be mindful of the spaces, possibilities, potentials, hesitations, and tensions inherent in ontological and epistemological shifts.

Ultimately, as a process of carefully and deliberately dismantling an established center and giving space for the agencies that are in the boundaries to take the central stage, unmaking-with has the potential not only to critique, but to actually instantiate more-than-human worlds and relations.

The article has explored the connections between unmaking and MTHD. Taking a practice-based (RTD) approach and focusing on workshops and design experiments, we have shown some potential of unmaking to suggest situated ways of knowing and doing in HCI that support MTHD thinking and practice. We have developed that argument in several steps. We have extended unmaking’s double movement to the MTHD practice of decentering through design, and used de/centering and un/making to make sense of the outcomes of a series of workshops focused on AI and CAs. These notions helped us to unpack in which particular ways the workshops unmade the anthropocentric biases in the design of CAs and instantiated more-than-human relations. That process was facilitated by situating the interactions of CAs and the positionality of the designers, the users, and the agents; by materializing new imaginaries; and by enacting alternative relations. Thus, we demonstrated how these moves enabled designers and researchers to go beyond critiquing AI or exposing its pitfalls and helped them to conceptualize, enact, and materialize more-than-human relations.

From the diffractive entanglement between unmaking and MTHD, we have proposed unmaking-with as a posthumanist practice and drafted three decentering tactics that can bring nuance to the unmaking processes—situating, materializing, and enacting. We hope that our explorations may serve as an illustration of existing and potentially new synergies between the two research HCI communities and illustrate how entangling them could support designers and researchers to better address the current global challenges.

Many people contributed to the ideas discussed in this article. We are indebted to all of our workshop participants (whose names are listed on the website www.more-than-human.com). We also thank James Pierce, who co-organized the workshop that was central to the development of ideas presented in the article.