Progress in automation will change the way we use and experience technology. The complex, unpredictable, and proactive nature of these technologies will transform them from tools controlled by humans to counterparts in dialogue with them. The existing models and frameworks on human-automation interaction (HAI), however, are insuficient to define this transition. They are mostly performance-oriented and neglect the experiential and meaning-oriented aspects of HAI, or highly rely on principles of direct manipulation limiting the afordances of automation. This creates a demand for alternative interaction paradigms based on quasi-social interaction. This paper elaborates on the problems of the existing frameworks and hints at new forms of interaction that consider experiential and performance goals.
Automation plays a crucial role in our lives. It subsumes a broad range of technologies able to autonomously carry out tasks formerly done by humans [1]. While current automation largely addresses pre-programmed routine tasks, advances in machine learning, natural language tive, “intelligent” automation, able to replace intellectual, creative, and non-routine work (i.e., knowledge work) [2]. This will change the way we use and experience technology. Automated systems difer from conventional computational artifacts, typically controlled by humans through forms of direct manipulation, such as Graphical User Interfaces (GUI), Tangible Computing, and in recent years embodied and ubiquitous computing. For example, automated vehicles, perceive the world through sensors, learn, and act in a proactive and autonomous manner. In fact, they are in dialogue with humans. People cooperate with them rather than use them. This shifts the perception of systems from a passive extension of self to active counterparts or collaborators [3].
Automation, however, does not happen overnight. It evolves continuously. The original idea of an automated tomation shape novel types of human-computer interaction in which humans and automated systems collaborate in diferent ways. This change raised concerns regarding the autonomy, trust, accountability, acceptance, and performance (efectiveness and eficiency in achieving task goals) of such ”teams” of people and machines in the field of human-computer interaction (HCI). A widespread shift collaborating with them will inevitably change the role of humans and consequently their experience of enjoyment and meaning in use. For example, feelings of autonomy as a source of well-being might sufer when delegating tasks rather than doing them. To this end, a big future challenge for HCI is to design meaningful interaction with automated systems while maintaining performance goals. While HCI has extensive knowledge about how to design for control, far less is known about appropriate interaction paradigms for automated systems perceived as counterparts.
system had been to relieve humans entirely from per- In the last decades, several models and frameworks are forming unwanted or dangerous tasks. This, however, is proposed that describe the interaction (or allocation of barely the case. More realistically, various forms of au- tasks) between human and automated systems. Most of (M. Hassenzahl) ∗Corresponding author. †These authors contributed equally. was later extended to 12 levels by Endsley [4]. These models of levels of automation (LOA), are originally develsystems, and their function allocation focuses on perfor© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License oped in the context of decision-making in safety-critical 3. Experiential Aspects mance metrics [5]. With increasing levels of automation (LOA), however, the relationship between human and automated systems changes [6]. At the higher levels, The LOA model was originally designed based on static humans are rather in dialogue with algorithmic, self- task allocation. This is done by either, allocating tasks learning, self-reliant, and proactive systems than acting to the better performer between human and automation through them. Personal assistants, chatbots, conversa- by comparison, automating everything possible and altional interfaces, and autonomous or even anthropomor- locating the rest to the human, or allocating based on phic robots are examples of such computational artifacts. cost and performance benefits [ 16]. Later, the concept of They tend to be perceived by their users as autonomous, ”Adaptive Automation” was introduced that allows more semi-intelligent agents, either incidentally because the lfexibility in task allocation to increase performance [ 17]. reasons for their actions are opaque (e.g., in the case of The HCAI framework also follows adaptive allocation. complex simulations or deep learning algorithms), or de- However, the performance-oriented perspective focuses liberately because they are designed that way (e.g., in the primarily on the functional capabilities of the technology case of anthropomorphic robots). This fundamentally when designing human-automation interaction (HAI). In impacts the relationship between humans and compu- its purest form, this leads to a so-called left-over allocatational artifacts and suggests an alterity relationship, tion of tasks to humans based on what the machine is where technology becomes “other”, i.e., a counterpart unable to do reliably (e.g., [18]). Even if the distribution technology [7, p. 99]. The change of relationship leads to of tasks is more considerate, it mostly follows the princiinteraction paradigms in which the automation can have ple of optimizing performance by, for example, designing roles with diferent goals from the human and could even automation in accordance with the cognitive abilities of supervise them (e.g., teaching robots). Thus, in these sys- humans (e.g., [1]). tems, using only LOAs for defining interaction between From the human perspective, this is a severely limhuman and automation is insuficient [ 5]. ited approach. In a meta-analysis, Onnasch et al. [19]
In his recent book Human-centered AI, Ben Shnei- showed that while the automation of routine tasks inderman, criticizes the LOA model: “[...]increases in au- creases the performance of a human-automation system, tomation must come at the cost of lowering human con- it has only a small efect on the experienced workload trol. But this zero-sum assumption limits thinking about of the humans involved. Increasing levels of automation ways to increase human control and the level of automa- reduce experiences of workload but lead to a decrease in tion. There is a better way” [8, p. 48]. To ensure human situational awareness and failure performance. Humans control while increasing automation, he suggests a two- start to lack insight into the tasks performed and become dimensional framework for human-centered AI (HCAI) less and less engaged. In the long-term, increasing levels with two axes for human control and automation. The of automation, thus, lead to deskilling and technological framework strongly emphasizes on supervisory control ”illiteracy”, as well as a sense of not contributing and, in interaction with automation (AI) in which the human consequently, to the difusion or even abandonment of is primarily assigned to the monitoring of automation, responsibility [20]. This reveals a fundamental dilemma and occasionally intervenes to overwrite decisions, ad- of automated systems: On their way to full automation, dress errors, or unexpected circumstances [9, 10]. While respective systems do not create more meaningful, rethis form of interaction is proposed to ensure reliability, laxing, and augmenting work experiences for the people previous research has shown that humans are bad at mon- involved, but disengagement and increased stress. In her itoring and vigilance due to reasons such as boredom, seminal work on the “Ironies of Automation”, Bainbridge fatigue, distraction, lack of situational awareness, and already discussed how the notion of replacing the human deskilling [11, 12]. A very well-known example of such through automation may actually lead to more and not a situation is the take-over situation in highly automated fewer problems [15]. These problems remain important cars. This situation requires the ”driver” of the automated when designing human-automation interaction (HAI) vehicle to take over control when the driving automa- (e.g., [21]), nevertheless, the actual design challenges are tion reaches its limits. To perform a safe maneuver, the more comprehensive than this: HAI is not only a cogni”driver” needs to shift his/her attention to the driving tive problem but an afective-motivational problem on context, gain situational awareness, have the required an individual and societal level. Despite this, research on skills to operate the car, and perform the right maneuver the experience of meaningful and fulfilling interaction in a few seconds[13, 14]. This example shows that human with autonomous systems and how to design for it is rare supervision over automation not only is insuficient to [22]. ensure safety or reliability but also leads to one of the One way to shape meaningful positive experiences in Ironies of Automation [15]. interaction with AI is through fulfilling human psychological needs. Previous research has widely discussed the role of universal human needs in shaping positive experiences. The Self-determination Theory specifies example, take the output of the autonomous typewriter the three needs for autonomy, relatedness, and compe- and further edit it to make it into her own; however, tence as essential substances for individual well-being editing is not writing. While both practices intersect, [23]. Later Sheldon et al. [24] extended this to a list of they are not integrated. In fact, through the design of au10 psychological needs. Drawn on these works, in the tonomous systems, humans are excluded from any new context of HCI, Hassenzahl et al. [25] showed that fulfill- practices performed by them. Consequently, humans ment of psychological needs leads to positive experiences might not feel responsible for the output of autonomous in interaction with technology. While the HCAI model systems and experience drastic changes in the meaning aims to put the human in the center, it only does so by of their work. By introducing an autonomous typewriter, addressing the need for autonomy, which comes at the the author’s work does not become easier but drastically cost of limiting automation. In this regard, a recent work changes its nature. on a collaboration with robots at work Smids et al. [26] Through their agency, opaqueness, and complexity, mentioned five characteristics of meaningful work: pur- autonomous systems can appear to humans as countersuing a purpose, social relationships, exercising skills parts rather than as tool-like extensions of their self ([3]). and self-development, self-esteem and recognition, and Despite being machines, people now tend to use social autonomy. These characteristics are in line with the metaphors in interaction, such as delegating, collaboratneeds for relatedness, competence, popularity, security, ing, being in a team, or trusting. Both aspects, that is, and autonomy, respectively [27]. Another approach is the entangled but separate practices as well as perceivthrough interactive (contrary to supervisory) allocation, ing autonomous systems as counterparts, lead to a focus and assignment of tasks between humans and automa- on the particular relationships established between the tion according to human psychological needs. Examples human and the system. These relationships, in turn, are of this approach are ”Hotzenplotz” by Klapperich and established through the roles assigned to the system. Is Hassenzahl [28], and its later adaptation by Frison et al. it an expert, apprentice, team member, subordinate, su[29] for automated vehicles, which propose interaction perior, or something completely diferent? In this view, concepts with automation that aim to fulfill the needs for interaction paradigms for autonomous systems and recompetence, autonomy, and stimulation while maintain- sulting experiences are inevitably social in nature and are ing automation advantages. heavily shaped by the relationships established through design. This has also been observed in previous studies (e.g., [34]) that demonstrate the pleasure and meaning 4. From a Tool to a Counterpart people derive from driving, the direct control over the car, and a feeling of being “one” with it. The very same Human-automation interaction (HAI) fundamentally study also shows that already simple automation, namely difers from traditional, cognitively oriented human- an adaptive cruise control, created feelings of the car computer interaction (HCI). In HCI, the human is the as other, which was experienced as supportive but also center of the design. Widespread interaction paradigms, diminished the meaning derived from driving. In this such as direct manipulation [30], gesture-based inter- case, driving even remained an activity, and was only action [31] or tangible computing [32] revolve around lightly supported by automation. Nevertheless, its meanhuman action, and agency and seek to extend the user’s ing changed considerably. It seems out of question, that control over the environment through computational more automation towards, for example, self-driving cars tools. Automation is diferent. It has its own, however, will inevitably lead to perceptions of the car as other. restricted, agency, which is often opaque and complex. Interestingly, while interaction paradigms based on Humans do not directly act through automation, but indi- the notion of technology as an extension of the self are rectly by instructing, supervising, or supporting it. Thus, highly developed and validated (e.g., direct manipulaHAI has the character of a co-performance of two more tion, [35]), the insights into the quasi-social interaction or less autonomous entities, the machine and the human paradigms with automated counterparts remain vague. [33]. This has fundamental consequences: Autonomous Current HCAI approaches tend to underplay the changes systems are not simply integrated elements of human in people’s relationships with technologies. For example, practices. In contrast, they have their own practices, Shneiderman [8, p. 55] mentions that computers are not which are entangled with the practices of humans, but teammates, collaborators, or co-active partners, as many not identical. To give an example: while a typewriter will shape the writing practice of a human author (for suggest “[…] Humans are responsible for actions of the technology assists that they use.” He is actually demandexample, compared to writing by hand), the author will ing to design HAI as empowering extension of the self, never have the feeling that the typewriter actually writes thereby perpetuating principles of direct manipulation her book. In contrast, an autonomous typewriter would and disregarding the challenges interaction with more have its own writing practice. The human author can, for and more autonomous systems will bring. This has been
also discussed widely in panels/duels he had with Pattie Maes [36, 37], where she mentions “some tasks I may just not want to do myself even if the interface was perfect. If my car had a perfect interface for fixing the engine, I still wouldn’t fix it. I just don’t want to bother with fixing cars.
I want someone else to do it” highlighting the importance of considering new paradigms for collaborating with, or delegating tasks to automation.
Good Human-Automation Interaction is an imminent, unsolved problem, despite the longstanding research into automation, its design, and efects. One problem is the inherent conflict between providing control to the human and the very fact that automation largely controls itself. Models of human-centered automation, provide no real solution to this. They simply demand to limit the autonomy of automation, and its power, in a way 5. Machine Autonomy ≠ that it remains controllable by a human. While this is an Anthropomorphism understandable demand, it does seem a very productive approach, and we doubt it will be successful merely by An often-used approach in exploring the interaction be- insisting on limiting automation. Another approach is tween human and automated systems and the distribu- to make automation more transparent. While applaudtion of their tasks is comparing these two interaction able, this approach seems doomed as well. True conpartners. Concepts such as “men are better at, machines trol requires a deep understanding of what is going on. are better at’’ (MABA-MABA) have been around since While a system may be able to better explain certain situthe 1950s [38]. Furthermore, the common notion of au- ations in hindsight, during the process itself, explaining tomation as a means to replace human actions leads the everything will most likely lead to information overload existing interaction paradigms for automated systems to on behalf of the human. The “transparent” automation mainly copy natural ways of interacting with humans. may rather foster a pseudo-accountability, which we find This is why some computational artifacts, such as social deeply disturbing and unethical. We need more refined robots, deliberately prompt social metaphors through interaction paradigms here that will certainly be hybrid anthropomorphism. Those “anthropomorphic [interac- in using essentially social forms of interaction but dotion] paradigms […] augment the functionality and behav- ing so in a reinterpreted, machine-like way. While this ioral characteristics of a robot […] that we can relate to is far from settled, it seems important that we do not and rationalize its actions with greater ease” [39]. This neglect the experiential view of human-computer interhas clear advantages since it transfers already existing action. Whether direct manipulating computational tools knowledge and skills from human-human interaction to or quasi-socially working with autonomous systems, the human-technology interaction [40]. However, the dele- provided arrangements should be meaningful to the hugation of tasks to another “human” might lead to feelings mans involved. This is certainly a challenge. From the of losing own autonomy and can be detrimental when perspective of psychological needs, the experience of expectations about the communication and interaction autonomy will change. While paramount and out of capabilities of the automated system cannot be fulfilled. question in direct manipulation paradigms, interaction This implies that, just because technology feels like a with otherware will entail negotiating autonomies rather. counterpart, we should not blindly anthropomorphize Direct manipulation is also bound to foster competence its interaction with us. Contrarily, we need to recognize experiences, which may become more dificult if it starts its non-human characters and shape and define it as an to get more unclear whether a work result can be at“Otherware”–a counterpart with quasi-social relation- tributed to the human or the machine. In contrast, the ships with us [3]. This is not limited to accepting that quasi-social nature of otherware may lead to new forms automation may be best framed as an otherware, with its of relatedness, the experience of mutual support, shared own goals and practices, but needs to be reflected in the values, and goals. Nowadays, all this remains largely interaction paradigms we provide. Therefore, the stan- unexplored. And we fear that this might remain so if the dard of direct manipulation simply makes no sense here, HAI and HCAI community does not acknowledge the since others are not to be direct-manipulated (e.g., you fundamental diference between interacting with compudo not direct-manipulate your colleague). Thus, we need tational tools as we know them and highly autonomous to accept the pseudo-social nature of interacting with systems. otherware, and the altered relationship to technology it implies. ration with non-human coworkers, in: 27th Inter- [40] L. Damiano, P. Dumouchel, Anthropomorphism in national Conference on Intelligent User Interfaces, human–robot co-evolution, Frontiers in psychology IUI ’22, Association for Computing Machinery, New 9 (2018) 468.
York, NY, USA, 2022, p. 27–35. URL: https://doi. org/10.1145/3490099.3511128. doi:1 0 . 1 1 4 5 / 3 4 9 0 0 9 9 .
3 5 1 1 1 2 8 . [28] H. Klapperich, M. Hassenzahl, Hotzenplotz: reconciling automation with experience, in: Proceedings of the 9th Nordic conference on human-computer interaction, 2016, pp. 1–10. [29] A.-K. Frison, P. Wintersberger, A. Riener, C. Schartmüller, Driving hotzenplotz: A hybrid interface for vehicle control aiming to maximize pleasure in highway driving, in: Proceedings of the 9th international conference on automotive user interfaces and interactive vehicular applications, 2017, pp. 236–244. [30] E. L. Hutchins, J. D. Hollan, D. A. Norman, Direct manipulation interfaces, Human–computer interaction 1 (1985) 311–338. [31] L. Gupta, S. Ma, Gesture-based interaction and communication: automated classification of hand gesture contours, IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and
Reviews) 31 (2001) 114–120. [32] E. Hornecker, J. Buur, Getting a grip on tangible interaction: a framework on physical space and social interaction, in: Proceedings of the SIGCHI conference on Human Factors in computing systems, 2006, pp. 437–446. [33] L. Kuijer, Automated artefacts as co-performers of social practices: washing machines, laundering and design, in: Social Practices and Dynamic Non
Humans, Springer, 2019, pp. 193–214. [34] K. Eckoldt, M. Knobel, M. Hassenzahl, J. Schumann,
An experiential perspective on advanced driver assistance systems (2012). [35] B. Shneiderman, The future of interactive systems and the emergence of direct manipulation, Behaviour & Information Technology 1 (1982) 237–256. [36] B. Shneiderman, P. Maes, Direct manipulation vs.
interface agents, Interactions 4 (1997) 42–61. URL: https://doi.org/10.1145/267505.267514. doi:1 0 . 1 1 4 5 / 2 6 7 5 0 5 . 2 6 7 5 1 4 . [37] D. Wang, P. Maes, X. Ren, B. Shneiderman, Y. Shi,
Q. Wang, Designing ai to work with or for people?, in: Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, 2021, pp.
1–5. [38] P. M. Fitts, Human engineering for an efective
air-navigation and trafic-control system. (1951). [39] B. R. Dufy, Anthropomorphism and the social robot, Robotics and autonomous systems 42 (2003) 177–190.