How does HCI Understand Human Agency and Autonomy?

Agency and autonomy have long been focal to HCI scholarship. Schneiderman’s 1985 Eight Golden Rules of Interface Design urged designers to “support internal locus of control” [145, p. 75]. CHI workshops in 1996 [56] and 2014 [26] addressed issues of autonomy in current technology use, covering the risks of undermining human autonomy [56], and theory and design strategies related to user autonomy [26]. The first of these led to the influential Value-Sensitive Design approach [55]. Organisers of the latter developed the METUX (Motivation, Engagement, and Thriving in User Experience) framework [132], which provides guidelines for design for human well-being, and draws attention to the multi-level nature of autonomy. Autonomy and agency have figured in a range of work aiming to characterise good user experience (UX) [e.g., 68, 69, 159] and they remain central concepts in recent strands of HCI research: For example, in Human-Computer Integration, agency is suggested to be central to categorising and understanding situations where “user and technology together form a closely coupled system” [122, 34, p. 1].

Beyond theoretical work, much recent empirical and design work addresses the question of how and when autonomy and agency should be supported. This work, discussed in greater depth in our literature review below, varies considerably in the kinds of agency and autonomy it addresses – covering for example, the experience of causing discrete events during minimal interactions [e.g., 36]; the experience of autonomy in episodes of gaming [e.g., 44]; but also the lived experience and personhood of dementia sufferers [e.g., 31]; and the material fact of control or influence in tasks [e.g., 134]. At present it is unclear how such different approaches to autonomy and agency, at different scales of behaviour, relate to one another. Moreover the use of near-identical language across this diversity of cases can make it difficult for HCI researchers to identify work relevant to their particular concerns. At the same time, we argue that there is value in understanding the interrelation of these diverse approaches: The tight integration of technologies into our bodies, behaviours and lives [122] has implications for autonomy and agency across multiple levels. Recent work in philosophy of cognition, for example, has indicated risks for integrated technology whereby delegation to technology at sensorimotor timescales (milliseconds) might be expected to impact on autonomy and agency in decision-making and everyday life [115, 171]. To understand such scenarios it is imperative to grasp how different aspects of agency and autonomy relate to one-another.

While recent years have seen repeated calls to deal with autonomy in a nuanced and multi-faceted way [e.g., 26, 64, 75], it remains unclear what a multilevel understanding of agency and autonomy, adequate to the range of work in HCI, might look like. Some calls for nuance have focused only on particular domains, such as multiple sclerosis support [64]. The METUX framework [28, 132] offers a less domain specific approach and outlines several “spheres” of life into which interactions and their implications can be categorised — from interaction at the interface, to the user’s wider life and society. Specifically, the METUX calls attention to the presence of multiple levels of need satisfaction, and provides a heuristic for “organizing thinking and evaluation” [132, p. 7] in design work. However, METUX does not focus specifically on autonomy, and the authors do not relate their account to previous work on autonomy and agency in HCI, nor address the full range of issues presented here. Also, the basis for choosing its approach to distinguishing aspects of autonomy issue — “spheres” of life — is not made clear. As the results of our literature review will show, a number of other aspects seem operative in approaches to autonomy and agency, and offer clarity in distinguishing understandings and issues. To our knowledge, no work has attempted to survey the breadth of approaches to autonomy and agency in HCI, to understand how they relate, and how they might coordinate and build upon one another. This is the goal of our paper.

We searched the ACM Digital Library on May 21st, 2022, for relevant publications from the CHI conference, ACM Transactions of Computer-Human Interaction (ToCHI), and International Journal of Human-Computer Studies (IJHCS), Behaviour and Information Technology (BIT), and Human-Computer Interaction, as these have previously been considered high-impact venues in HCI [73]. We searched these venues for the keywords "agency" and "autonomy" in keywords and abstracts. This resulted in 207 full papers (see Figure 1 for breakdown per publication venue).

We first reviewed abstracts and titles for all papers for whether they concerned human autonomy or agency. Where abstracts focused on non-human agency and autonomy (e.g., of robots, software agents), we reviewed the full text, excluding it if there was no discussion of human agency and autonomy. Example exclusions include one paper which dealt with the agency of parrots in animal-computer interaction [97], and another focusing on the calibration of autonomous braking systems in cars [57]. In total, 46 papers were removed in this way, leaving 161 papers for analysis.

The papers in our corpus covered a wide range of subject matters, from the design and evaluation of input and control methods [134], to issues of everyday living [7], family and parenthood [59]. The largest group dealt with issues of ageing and accessibility (n=33): Here autonomy and agency often concerned how individuals could be supported in retaining personhood [168], maintaining social and family ties [51, 83], and living independent lives [7], a focus also seen in work on parenting (n=6), children (n=6), and in work concerned with HCI4D (n=6). Across all these domains, issues of autonomy and agency commonly focused on material matters, addressing constraints imposed by health, and by social and material conditions. The next largest group focused on gaming, social media, and entertainment (n=32). Here, issues of autonomy and agency usually centred on people’s experiences: how free, independent, self-governing, and in control they felt during the interaction. The remaining papers focused on a range of other subjects; from work (n=17), to input methods (n=6).

Across our corpus it was evident that autonomy and agency are widely considered desirable qualities in HCI. However, in many papers this value was left implicit or unclear (n=57). For example, one paper reported that “users desire agency and control in all aspects of the system” [146, p. 1270] but did not elaborate on why this might be. Other papers sought to understand how technologies could support agency and autonomy, but did not address their value [e.g. 162, 78]. In the following, we examine in more detail the reasons why agency and autonomy are valued, as well as what beneficial outcomes HCI researchers associate them with.

Only one third of papers in our corpus (n=58) gave explicit definitions of autonomy or agency. Lukoff et al., for example, defined sense-of-agency as “an individual’s experience of being the initiator of their actions in the world [...] broken down into feelings of agency, that is, the in-the-moment perception of control, and judgments of agency, that is, the post hoc, explicit attribution of an action to the self or other” [110, p. 3]. However, even some explicit definitions were rather unspecific (“Human autonomy is the ability ‘to be one’s own person’ and impacts well-being” [5, p. 10]) or tautological (“autonomy ("a sense of freedom and autonomy to play the game as desired” [180, p. 6]).

In 47 papers, definitions were implied via associative statements. For example, the work by da Rocha et al. [39] contained a number of statements which together suggested that they used “autonomy” to refer to a player’s ability to act for themselves without support in gaming sessions (e.g., "The goal is to improve a player’s autonomy and game enjoyment by removing barriers that can hinder their ability to learn a game or engage in gameplay." [39, p. 10].) In another 55 papers, the meaning of autonomy or agency was less clear, or rather inconsistent. For instance, Gonçalves et al. described agency in terms of decision making, distinguished from mere execution: “decision making on the part of both players [...] Avoid the perception that one player is just an “executor" of the game or another player’s orders” [62, p. 5]. However, later statements in the same paper specifically associated task execution with agency, noting that the players who gave orders “would like to have more agency in the action (e.g. in shooting monsters, in rescuing people)” [62, p. 8, our emphasis]. A paper grounded in Actor Network Theory first stated that agency arises from network effects, and is not the property of individual artifacts [80, p. 828], but later discussed “the agency of the artifacts” [80, p. 829]. Finally, some papers [e.g., 2, 173, 100, 47, 52, 2] explicitly indicated in the abstract or introduction that agency or autonomy was a significant themes of the paper but then never used the terms directly again. Very few papers explicitly engaged with existing theoretical accounts of agency and autonomy. However, in some papers that did, a certain understanding of agency or autonomy was implicit via the concepts and theories cited. Self-Determination Theory (including METUX), for example, was most frequently mentioned (n=17), particularly in works focusing on the player experience [e.g., 96, 147, 44] or user experience [e.g., 129, 130]. Research focusing on the immediate sense of control over actions often resorted to concepts from cognitive science, such as the notion of intentional binding [e.g., 143, 36, 106]. Other conceptual frameworks mentioned include Actor Network Theory [e.g., 58, 4], value-sensitive design [e.g., 7], social anarchism [92], and Kantian philosophy [64].

A few papers in our corpus explicitly noted that agency and autonomy had “many different facets” [64, p. 12]. In line with this we identified a variety of facets of autonomy and agency in the reviewed works. By analysing the corpus as a whole we found that these facets, and author’ approaches to them, could be usefully organised around four broad aspects: (1) A focus on self-causality or identity. (2) A focus on the experience, or material aspects of agency and autonomy. (3) The time-scale over which autonomy and agency play out. (4) The degree to which either independence or interdependence are emphasised. These aspects are summarised and defined in Table 1 and discussed in detail in the subsections below.

Boundary objects are flexible yet robust concepts that can help coordinate the perspectives and activities of different communities of practice [126], without requiring strict consensus on precise definitions [103]. Previous work has emphasised that, to function well, boundary objects should be sufficiently flexible to meet the informational and communicative needs of the communities involved [103]. Our review suggests that agency and autonomy already fulfil this criterion, as illustrated by their continuing significance in a range of different issues and contexts, over three decades of HCI research. However, while such interpretive flexibility is a well known property of boundary objects [165], Star and others have emphasised that interpretive flexibility is not sufficient on its own [103]; boundary objects can fail to function well, if they are not also robust ¹ enough in usage to retain a broader identity across interpretations. This is required to support coherence across intersecting communities [150]. For example, in sustainability studies, it was found that the concept of “resilience” did not succeed as a coordinating boundary object [126], since different communities’ understandings did not overlap in important ways.

What do robustness and flexibility mean in practice? Star [103] emphasises that boundary objects rely on both 1) the refinement of the concepts to address local concerns, and 2) the coordination of these local understandings with the concept’s wider identity. In terms of autonomy and agency in HCI, we suggest this might mean 1) local attention by researchers on the aspects of autonomy and agency which are relevant to concerns in their area of work, and 2) integrative work to relate these aspects to one another, and to situate local findings and understandings in the context of wider understandings. Our analysis points to concrete ways in which the HCI community might approach these two activities, to develop autonomy and agency into better functioning boundary objects:

While many of the reviewed works associated autonomy and agency with a wide range of positive outcomes, we often found a lack of specificity in how these outcomes might be linked to particular expressions of autonomy and agency. This was least problematic in work which pointed to relatively modest conclusions, e.g., linking outcomes readily measurable at episode level [e.g., task performance, 149] to autonomy and agency during single-episodes of interaction. Here, even if it was not directly stated which aspects of autonomy and agency were considered relevant, this was often evident via the nature of experimental manipulations [e.g., 3], or via questionnaires which addressed well-defined operationalisations of autonomy [e.g., 19]. However, some such scenarios raised important questions. For example, in work which operationalised sense-of-agency via temporal binding [e.g., 15], the association between temporal binding and UX was generally asserted rather than directly tested. Only in more recent work was the relationship between temporal binding and user experience directly examined, and ultimately found to be somewhat unclear [e.g., 15].

Elsewhere the relationship between agency, autonomy and UX was unclear in other ways. In some instances, autonomy was treated as an antecedent of good UX [e.g., 3, 37]; in others as a component of good experience, and a positive experiential phenomenon in its own right [e.g., 96]. This suggests that the UX community lacks agreement on clear models of how these constructs relate to one another. In line with this, the links between autonomy or agency and UX were often articulated in quite general terms. Very few papers investigated how autonomy and agency related to specific components of user experience, such as sense-of-presence [81]. Future work might aim to move beyond the current understanding of agency and autonomy as components of generically positive experience, and instead seek to better understand how and why specific aspects of autonomy and agency foster specific aspects of positive experience.

Outside UX and player experience, papers often connected support for agency and autonomy to “high value” outcomes, such as sense-of-self and well-being [e.g., 110]. However, this connection was always drawn on the basis of prior work in other domains, and not explicitly examined in the case of technology use. Again, reviewed works often left it unclear which aspects of autonomy and agency were expected to lead to these positive outcomes. Exceptions to this were found in some qualitative field work, addressing for example, how low level “proprioceptive” agency in execution could be expected to lead to higher level agency outcomes and well-being [102, p. 2157]. Or in work that reported on the complex trade-offs in well-being support between what we have characterised as execution, decision, and self-congruence [64]. However, as we have noted, autonomy and agency were generally studied in single sessions of interaction, without addressing the user’s wider life context, where high-value outcomes like well-being, motivation, and sense-of-self play out. In fact, our corpus contained no works that compared such outcomes pre- and post-session, or observed it over long term studies. Of course, it is challenging to address long-term outcomes in HCI studies and interventions [95]. However, without such work, it is hard to see how we can meaningfully investigate the impact of agency and autonomy on outcomes such as well-being and sense of self. In addition to the qualitative work discussed above, we suggest that the wider resources of Self-Determination Theory, beyond Basic Needs Satisfaction might be drawn upon here, and that work by Tyack and Wyeth provides an exemplary and achievable model for such future work [161].

Another complex question raised by our analysis concerns how we can disentangle experiential and material aspects of autonomy and agency and their impact on outcomes. As reported in section 4.4.2, we found suggestions that some experiential and material aspects could be decoupled. For example, that the experience of autonomy and agency can be altered without affecting the user’s causal involvement or material control over outcomes [e.g., 44]. Similarly, it was found that sense-of-agency over sensorimotor actions could predict training outcomes even where users had no causal role in the outcome [90]. These examples indicate that there are clear opportunities for exploitation of such gaps – sometimes to the detriment of the user: for instance, using the feeling of autonomy or agency as an alibi for materially reduced efficacy, choice, or scope of control. This is illustrated in our corpus in work by Bakewell et al., where companies granted workers self-management in certain aspects of their working life in a way which created wider, and sometimes undignified, restrictions on behaviour and efficacy [8]. Based on our findings, one straightforward recommendation for future work is to be (more) attentive to this distinction between experiential and material aspects of autonomy and agency. It is crucial that HCI researchers clearly report the relevant aspect(s) of autonomy and agency, and focus analysis appropriately, particularly where technology is likely to impact outcomes meaningful to the user. Future work might also seek to understand whether — as predicted by SDT [135, 139] — the congruence of outcomes will impact on autonomy independently of changes in, e.g., available choice, executional involvement, or the user’s independence of action. For example, the degree of control afforded in game and VR environments may make them a suitable vessel for such work, though this will require careful study design.

Finally, our analysis raises ethical challenges. First, there is the question of who benefits from the user’s agency or autonomy. Our corpus contained examples where these benefits primarily accrued to others — employing organisations, or sellers of products and services. This might not always be a problem in itself. One paper, for example, reported that delegation of worker decision-making to AI reduced attentiveness and engagement, leading to poorer task outcomes [104]. In this case while the organisation is the primary beneficiary of the worker’s agency, it is not immediately clear that this is problematic for the worker. However, other papers gave examples where benefits and burdens were more problematically distributed: users gained only limited and localised control or choice, or only the experience of agency and autonomy, while in return accepting increased burdens and material restrictions on aspects of behaviour [e.g., 8].

One key problem here is that autonomy and agency are resonant words, associated with a wide range of meanings. Sometimes, autonomy and agency are “human rights” [e.g., 162, 7] or are meaningfully associated with well-being [e.g., 99, 133]. In other cases, agency and autonomy have more limited ramifications, making it more acceptable to treat them as manipulable parameters which can improve work outcomes [e.g., 8, 104] or the appeal of our products [e.g., 3, 153]. This ambiguity can allow vague and under-specified claims of autonomy support to serve as cover for exploitative or manipulative practices. Clarity in addressing agency and autonomy therefore has not only practical implications for HCI research, but also ethical importance. A simple step towards addressing this is to spell out (expected) outcomes; not every study addressing autonomy and agency can plausibly attach its outcomes to the most lofty ramifications and values of these concepts. For example, while agency has been hypothesised as the ground of identity and the self-other distinction [98], the relevance of this to agency during button pushing seems limited. Instead, authors should focus on plausible, direct, outcomes which are relevant to the time-scale and scope of the work carried out, and relevant to the specific aspects of autonomy and agency involved. Publication pressures can easily drive us to inflate the value of our work, but in the case of autonomy and agency there seems a concrete risk in this value-inflation: it can provide a vague, peer-reviewed alibi for those who might exploit humans under the banner of apparent autonomy support.

Another ethical question raised by our analysis concerns autonomy in self-construction, and how we should handle the delicate issue of becoming involved in the user’s processes of self-change, and self-control. This seems particularly relevant for behaviour change technologies and technologies which support reflection. Across our corpus, three papers addressed different contexts of self-change: helping children internalise behavioural norms [70], supporting trafficked women to understand and address their exploitation by others [156], and helping abusive men take responsibility for their behaviour [14]. These particular examples stand out as being carefully conducted, and in each case working towards outcomes likely congruent with users’ long-term goals and values. However, these cases also illustrate the delicate balance involved in such work. The third example [14], in particular, involved a degree of short-term thwarting of autonomy, intentionally restricting meaningful choices to encourage the men to reappraise their identities and responsibility. Such work requires careful reasoning about what aspects of autonomy and agency will be restricted and supported, on what time-scales, and with what consequences. It requires reasoning about whether the long-term outcomes can be expected to be congruent with the user’s values and goals, and if not, then what warrants the intervention. Not all cases of persuasive technology are so dramatic and significant as the above examples, but the issues raised here can be informative elsewhere. One special case, for example, concerns apps for self-control, where user’s goals, values and intentions may vary over time, and care will be required in deciding how, and how much, to hold users accountable to their past intentions [111]. Again, we suggest that one route to the required care in such cases is to understand the different scales and aspects of autonomy and agency involved, and clarity about what is supported and thwarted, when and where. Prior work that deals with such trade-offs, and with the warranting of agency delegation [e.g., 166, 64, in our corpus], can be useful in guiding reasoning in such situations. Thinyane and Bhat’s work on the support of trafficked women [156] offers a sophisticated discussion of relevant issues, articulated with reference to Sen’s Capabilities framework [142]. Likewise, we see much promise in future work developing guidelines specific to user autonomy and agency in persuasive and reflective technologies, and more broadly in technologies for self-change.

Addressing concepts of such complexity inevitably means that certain issues must be left out of scope. First, this review focuses on agency and autonomy, and does not consider any of the agency-adjacent terms which have proliferated over the past few decades, such as “empowerment” [141], “efficacy” [10], or “competence” [13]. Agency and autonomy seemed to us to be the primary concepts here, whereas the other concepts can often be considered special cases, derivatives, or complexes which involve aspects of autonomy and agency alongside other factors. Moreover, “empowerment” is already the subject of a recent review at CHI [141]. That said, fitting these adjacent constructs into the overall landscape of agency and autonomy (e.g., by drawing on the four aspects outlined in this review) could provide useful insight. Second, the scope of the review — spanning over 30 years of research — means that we have addressed agency and autonomy at a relatively high level. This is in line with the goals of this paper: to provide a view of how these concepts and their accompanying issues are understood and can be better coordinated across HCI. However, future reviews which address understandings of these themes in specific domains and areas of research would help clarify understandings in particular communities and contexts, as well as furthering the development of agency and autonomy as boundary objects for HCI.