Leveraging human agency to improve confidence and acceptability in human-machine interactions

Abstract

Repeated interactions with automated systems are known to affect how agents experience their own actions and choices. The present study explores the possibility of partially restoring sense of agency in operators interacting with automated systems by providing additional information about the system's decision, i.e. its confidence. To do so, we implemented an obstacle avoidance task with different levels of automation and explicability. Levels of automation were varied by implementing conditions in which the participant was free or not free to choose which direction to take, whereas levels of explicability were varied by providing or not providing the participant with the system's confidence in the direction to take. We first assessed how automation and explicability interacted with participants' sense of agency, and then tested whether increased self-agency was systematically associated with greater confidence in the decision and improved system acceptability. The results showed an overall positive effect of system assistance. Providing additional information about the system's decision (explicability effect) and reducing the cognitive load associated with the decision itself (automation effect) was associated with stronger sense of agency, greater confidence in the decision, and better performance. In addition to the positive effects of system assistance, acceptability scores revealed that participants perceived “explicable” systems more favorably. These results highlight the potential value of studying self-agency in human-machine interaction as a guideline for making automation technologies more acceptable and, ultimately, improving the usefulness of these technologies.

Introduction

Human agents are used to interacting with sophisticated computer systems designed to help them in their activities and a significant number of our daily actions are technologically mediated. This is the case of the airplane pilot who controls their aircraft through increasingly sophisticated pilot assistance tools, but also that of the doctor assisted in their diagnoses by artificial intelligence algorithms. This paramount importance of technological assistance is further demonstrated by the growing role that virtual assistants – be it our phones, computers or specific planning devices – play in our daily lives.

Repeated interactions with automated systems can be expected to affect the way individuals experience their own actions and choices. Such experience is often referred to as “sense of agency” (SoA) and describes the subjective feeling associated with controlling one's own actions and, through these actions, events in the outside world (Haggard & Tsakiris, 2009). SoA plays a key role in guiding attributions of responsibility (Bigenwald & Chambon, 2019) and serves as a key motivational force for human behaviour (Di Costa, Théro, Chambon, & Haggard, 2018). A detrimental effect of human-machine interactions on SoA has been demonstrated in a number of experimental studies, one of the most typical of which implements an aircraft supervision task (e.g. Berberian, Sarrazin, Le Blaye, & Haggard, 2012). In this task, the participant is responsible for supervising the movement of an aircraft that may encounter unpredictable obstacles. When a conflict occurs due to the presence of another aircraft, the participant has to decide and implement the appropriate control to avoid the obstacle using a button-based interface. Following an established classification (Sheridan & Verplank, 1978), the levels of automation of the task were manipulated from the user having complete control (no automation) to the computer performing the entire task with the participant simply observing (full automation). The results showed a decrease in the participant's sense of agency concomitant with the increase in automation, and suggest that increasing the level of automation tends to distract operators from the results of the action and alter the emergence of a sense of control.

What makes automation particularly detrimental to the operator's sense of agency is not yet fully understood, but there is a relative consensus that the lack of transparency on how the system makes its decisions, or simply operates, is a key factor (Christoffersen & Woods, 2002; Klien, Woods, Bradshaw, Hoffman, & Feltovich, 2004). In most human-machine interactions, most of these decision processes are unknown, inaccessible, or even not explainable at all (Norman, 1990). Such opacity makes it difficult for the operator to link system intention to actual state and to predict the sequence of events that will occur. Predictive mechanisms are known to play a critical role in the development of individuals' sense of agency, by allowing the attribution of observed sensory events to prior intentions (Chambon, Sidarus, & Haggard, 2014; Chambon, Wenke, Fleming, Prinz, & Haggard, 2013; Haggard & Eitam, 2015). By affecting the predictability of actions, the inherent opacity of technological systems is therefore likely to alter perceived agency in human operators.

In a previous study, we directly investigated how system predictability impacts the development of agency experience during human-machine interaction (Le Goff, Rey, Haggard, Oullier, & Berberian, 2018). Particularly, we explored the benefit of prime messages regarding system intention while supervising an automated system. We tested whether providing information about what to do next mitigated the deleterious effect of reduced freedom of action on agency and, in doing so, increased the user's level of acceptability, along with increased control and performance. Our results suggest that displaying the system's intentions prior to an action is a good candidate for maximizing the experience of agency in supervisory task, and for increasing system acceptability as well. These preliminary results open interesting avenue as to how to modulate the emergence of the experience of agency during human-machine interaction.

The present study aims to go further in exploring the information required for making technological systems more intelligible, and to test whether improving intelligibility concomitantly increases the level of agency experienced by human operators. In two distinct experiments, we explored the role of communicating specific metacognitive information on improving the SoA of participants interacting with distinct automated systems. Specifically, both experiments implemented an avoidance task with different levels of automation and explicability.

Levels of automation were varied by implementing conditions in which the participant was free to choose which direction to take (free choice trials) or not (forced choice trials). Levels of explicability were varied by providing or not providing the participant with the system's confidence in the direction to take. Confidence can be seen as a measure of the uncertainty (or certainty) associated with one's choice or action (Fleming & Lau, 2014). Communicating confidence was intended to improve explicability of the system's decision, by increasing its transparency, that is, by providing the participant with additional information, such as the level of confidence associated with that decision (Tintarev & Masthoff, 2015). Indeed, the level of uncertainty (or confidence) associated with a decision is a key explanatory factor for why a decision is made or not, and whether or not that decision will be updated or revised in the future (Balsdon, Wyart, & Mamassian, 2020). The beneficial role of confidence on decision making has already been demonstrated in group settings, where sharing metacognitive representation increases joint performance (Bahrami et al., 2010; Fusaroli et al., 2012; Le Bars et al., 2020) and enhances team coordination (Lausic, 2009; Le Bars, Devaux, Nevidal, Chambon, & Pacherie, 2020; Poizat, Bourbousson, Saury, & Sève, 2009). Communicating confidence also makes performance more fluid and prospectively improves SoA (Chambon, Filevich, & Haggard, 2014; Sidarus, Vuorre, & Haggard, 2017), especially when sensorimotor information is not available (Pacherie, 2013) such as when interacting with an automated system. Finally, there is indirect evidence that improving the operator's SoA during interaction with automated systems concomitantly improves acceptability of the system's decision itself (Le Goff et al., 2018). In addition to exploring the relationships between explicability and SoA, we also tested whether an increase in the participant's SoA could be consistently associated with greater confidence in their decision and greater acceptability of the system.

In both experiments, the participant's choice and three additional measures were collected: (i) Temporal Binding (TB), a widely-reported temporal compression between a voluntary action and its consequence (hence originally referred to as “intentional binding”), as an implicit proxy of the participant's sense of agency (Caspar, 2017; Ebert & Wegner, 2010; Haggard, Clark, & Kalogeras, 2002; Vogel, Jording, Esser, Weiss, & Vogeley, 2021); (ii) a measure of the participant's confidence in either their decision (free trials) or the system's decision (forced trials); and finally, (iii) the perceived acceptability of the system by the participant (Van Der Laan, Heino, & De Waard, 1997).

In Experiment 1, we had three key predictions: (1) Participants would experience lower levels of agency when forced to follow the system's decision, compared to freely choosing (automation effect) (Barlas, Hockley, & Obhi, 2018; Berberian et al., 2012; Caspar, Cleeremans, & Haggard, 2018); (2) communicating to participants the system's confidence in the best decision would restore or even improve participants' SoA (explicability effect) (Sidarus et al., 2017); and finally (3) an increase in the participant's SoA would be associated with an increase in system acceptability (Le Goff et al., 2018). In Experiment 2, we further explored the relationships between our control measure, our decision confidence measure, and task demands. We leveraged a procedure developed in a previous study (Potts & Carlson, 2019) to clarify the contribution of task difficulty to the relationship between automation, explicability and sense of agency, using a modified version of the avoidance task from Experiment 1.