Subjective Ratings Of Trust Research Articles

It's Not Only What You Say, But Also How You Say It: Machine Learning Approach to Estimate Trust from Conversation.

The objective of this study was to estimate trust from conversations using both lexical and acoustic data. As NASA moves to long-duration space exploration operations, the increasing need for cooperation between humans and virtual agents requires real-time trust estimation by virtual agents. Measuring trust through conversation is a novel and unintrusive approach. A 2 (reliability) × 2 (cycles) × 3 (events) within-subject study with habitat system maintenance was designed to elicit various levels of trust in a conversational agent. Participants had trust-related conversations with the conversational agent at the end of each decision-making task. To estimate trust, subjective trust ratings were predicted using machine learning models trained on three types of conversational features (i.e., lexical, acoustic, and combined). After training, model explanation was performed using variable importance and partial dependence plots. Results showed that a random forest algorithm, trained using the combined lexical and acoustic features, predicted trust in the conversational agent most accurately . The most important predictors were a combination of lexical and acoustic cues: average sentiment considering valence shifters, the mean of formants, and Mel-frequency cepstral coefficients (MFCC). These conversational features were identified as partial mediators predicting people's trust. Precise trust estimation from conversation requires lexical cues and acoustic cues. These results showed the possibility of using conversational data to measure trust, and potentially other dynamic mental states, unobtrusively and dynamically.

The frequency of positive and negative interactions influences relationship equity and trust in automation

The purpose of this study was to 1) examine whether frequency of positive and negative interactions (manipulated via reliability) with a computer agent had an impact on an individual's trust resilience after a major error occurs and 2) empirically test the notion of relationship equity, which encompasses the total accumulation of positive and negative interactions and experiences between two actors, on user trust on a separate transfer task. Participants were randomized into one of four groups, differing in agent positivity and frequency of interaction, and completed both a pattern recognition task and transfer task with the aid of the same computer agent. Subjective trust ratings, performance data, compliance, and agreement were collected and analyzed. Results demonstrated that frequency of positive and negative interactions did have an impact on user trust and trust resilience after a major error. Additionally, it was shown that relationship equity has an impact on user trust and trust resilience. This is the first empirical demonstration of relationship equity's impact on user trust in an automated teammate.

Human-Likeness of Feedback Gestures Affects Decision Processes and Subjective Trust

Trust is fundamental in building meaningful social interactions. With the advance of social robotics in collaborative settings, trust in Human–Robot Interaction (HRI) is gaining more and more scientific attention. Indeed, understanding how different factors may affect users’ trust toward robots is of utmost importance. In this study, we focused on two factors related to the robot’s behavior that could modulate trust. In a two-forced choice task where a virtual robot reacted to participants’ performance, we manipulated the human-likeness of the robot’s motion and the valence of the feedback it provided. To measure participant’s subjective level of trust, we used subjective ratings throughout the task as well as a post-task questionnaire, which distinguishes capacity and moral dimensions of trust. We expected the presence of feedback to improve trust toward the robot and human-likeness to strengthen this effect. Interestingly, we observed that humans equally trust the robot in most conditions but distrust it when it shows no social feedback nor human-like behavior. In addition, we only observed a positive correlation between subjective trust ratings and the moral and capacity dimensions of trust when robot was providing feedback during the task. These findings suggest that the presence and human-likeness of feedback behaviors positively modulate trust in HRI and thereby provide important insights for the development of non-verbal communicative behaviors in social robots.

Estimating Trust in Conversational Agent with Lexical and Acoustic Features

As NASA moves to long-duration space exploration operations, there is an increasing need for human-agent cooperation that requires real-time trust estimation by virtual agents. Our objective was to estimate trust using conversational data, including lexical and acoustic features, with machine learning. A 2 (reliability) × 2 (cycles) × 3 (events) within-subject study was designed to provoke various levels of trust. Participants had trust-related conversations with a conversational agent at the end of each event. To estimate trust, subjective trust ratings were predicted using machine learning models trained on three types of conversational features (i.e., lexical, acoustic, and combined). Results showed that a random forest model, trained on the combined lexical and acoustic features, best predicts trust in the conversational agent (R2adj = 0.67). Comparing models, we showed that trust is not only reflected in lexical cues but also acoustic cues. These results show the possibility of using conversational data to measure trust unobtrusively and dynamically.

Modeling and inferring human trust in automation based on real- time eye tracking data

Trust miscalibration remains a major challenge for human-machine interaction. It can lead to misuse or disuse of automated systems. To date, most trust research has relied on subjective ratings and behavioral or physiological data to assess trust. Those trust measurements are discrete, disruptive and quite difficult to implement. To better understand the process of trust calibration, we propose eye tracking as an unobtrusive method for inferring trust levels in real time. Using an Unmanned Aerial Vehicle simulation, participants were exposed to varying levels of reliability of an automated target detection system. Eye movement data were captured and labeled as high or low trust based on subjective trust ratings. Feature extraction and raw eye movement data were compared as input for multiple classification modeling methods. Accuracy rates of 92% and 80%, respectively, were achieved with individual-level and group-level modeling, suggesting that eye tracking is a promising technique for tracing trust levels.

Effects of an Unexpected and Expected Event on Older Adults' Autonomic Arousal and Eye Fixations During Autonomous Driving.

Driving cessation for some older adults can exacerbate physical, cognitive, and mental health challenges due to loss of independence and social isolation. Fully autonomous vehicles may offer an alternative transport solution, increasing social contact and encouraging independence. However, there are gaps in understanding the impact of older adults’ passive role on safe human–vehicle interaction, and on their well-being. 37 older adults (mean age ± SD = 68.35 ± 8.49 years) participated in an experiment where they experienced fully autonomous journeys consisting of a distinct stop (an unexpected event versus an expected event). The autonomous behavior of the vehicle was achieved using the Wizard of Oz approach. Subjective ratings of trust and reliability, and driver state monitoring including visual attention strategies (fixation duration and count) and physiological arousal (skin conductance and heart rate), were captured during the journeys. Results revealed that subjective trust and reliability ratings were high after journeys for both types of events. During an unexpected stop, overt visual attention was allocated toward the event, whereas during an expected stop, visual attention was directed toward the human–machine interface (HMI) and distributed across the central and peripheral driving environment. Elevated skin conductance level reflecting increased arousal persisted only after the unexpected event. These results suggest that safety-critical events occurring during passive fully automated driving may narrow visual attention and elevate arousal mechanisms. To improve in-vehicle user experience for older adults, a driver state monitoring system could examine such psychophysiological indices to evaluate functional state and well-being. This information could then be used to make informed decisions on vehicle behavior and offer reassurance during elevated arousal during unexpected events.

Eye Tracking: A Process-Oriented Method for Inferring Trust in Automation as a Function of Priming and System Reliability

Trust miscalibration, a mismatch between a person's trust in automation and the system's actual reliability, can lead to either misuse or disuse of automation. Existing techniques to measure trust (e.g., subjective ratings) tend to be discrete and/or disruptive. To better understand and support the process of trust calibration, a nonintrusive continuous measure of trust is needed. The present study investigated whether eye tracking can serve this purpose. In the context of an unmanned aerial vehicle simulation, participants monitored six video feeds to detect predefined targets with the assistance of onboard automation. Automation reliability (95% versus 50% reliable) and priming (reliability information provided or not) were manipulated. Eye movement data, subjective trust ratings, and performance data were collected. The eye tracking data show that people visit more frequently and spend more time on low reliable automation. Priming information could also affect the participants’ trust level and trigger different types of searching behavior, as reflected in eye tracking data such as mean saccade amplitude. In summary, these findings confirm that eye tracking is a very promising tool for inferring trust and supporting future research into trust calibration.

Feedback on system or operator performance: Which is more useful for the timely detection of changes in reliability, trust calibration and appropriate automation usage?

Creating safe human-machine systems requires that operators can quickly notice changes in system reliability in the interest of trust calibration and proper automation usage. Operators’ readiness to trust a system is determined not only by the performance of the automation but also by their confidence in their own abilities. This study therefore compared the usefulness of feedback on the performance of either agent. The experiment required two groups of ten participants each to perform an automation-assisted target identification task with “Automation Performance Feedback” (APF) or “Operator Performance Feedback” (OPF). Four different scenarios differed with respect to the degree and duration of changes in system reliability. Findings indicate that APF was more effective for supporting timely adjustments of perceived system reliability, especially with large and long reliability changes. Subjective trust ratings and performance were not affected, however, suggesting that these two factors are closely linked and more relevant for automation reliance.

Get Ready for Being Chauffeured : Passenger's Preferences and Trust While Being Driven by Human and Automation.

We investigated passenger's trust and preferences using subjective, qualitative, and psychophysiological measures while being driven either by human or automation in a field study and a driving simulator experiment. The passenger's perspective has largely been neglected in autonomous driving research, although the change of roles from an active driver to a passive passenger is incontrovertible. Investigations of passenger's appraisals on self-driving vehicles often seem convoluted with active manual driving experiences instead of comparisons with being driven by humans. We conducted an exploratory field study using an autonomous research vehicle (N = 11) and a follow-up experimental driving simulation (N = 24). Participants were driven on the same course by a human and an autonomous agent sitting on a passenger seat. Skin conductance, trust, and qualitative characteristics of the perceived driving situation were assessed. In addition, the effect of driving style (defensive vs. sporty) was evaluated in the simulator. Both investigations revealed a close relation between subjective trust ratings and skin conductance, with increased trust and by trend reduced arousal for human compared with automation in control. Even though driving behavior was equivalent in the simulator when being driven by human and automation, passengers most preferred and trusted the human-defensive driver. Individual preferences for driving style and human or autonomous vehicle control influence trust and subjective driving characterizations. The findings are applicable in human-automation research, reminding to not neglect subjective attributions and psychophysiological reactions as a result of ascribed control duties in relation to specific execution characteristics.

Trust and Approachability Mediate Social Decision Making in Human-Robot Interaction

As humanoid robots become more advanced and commonplace, the average user may perceive their robotic companion as human-like entities that can make social decisions, such as the deliberate choice to act fairly or selfishly. It is important for scientists and designers to consider how this will affect our interactions with social robots. The current paper explores how social decision making with humanoid robots changes as the degree of their human-likeness changes. For that purpose, we created a spectrum of human-like agents via morphing that ranged from very robot-like to very human-like in physical appearance (i.e., in increments of 20%) and measured how this change in physical humanness affected decision-making in two economic games: the Ultimatum Game (Experiment 1) and Trust Game (Experiment 2). We expected increases in human-like appearance to lead to higher rates of punishment for unfair offers and higher ratings of trust in both games. While physical humanness did not have an impact on economic decisions in either of the ex-periments, follow-up analyses showed that both subjective ratings of trust and agent approachability medi-ated the effect of agent appearance on decision-making in both experiments. Possible consequences of these findings for human-robot interactions are discussed.

Transitioning to Future Air Traffic Management: Effects of Imperfect Automation on Controller Attention and Performance

This study examined whether benefits of conflict probe automation would occur in a future air traffic scenario in which air traffic service providers (ATSPs) are not directly responsible for freely maneuvering aircraft but are controlling other nonequipped aircraft (mixed-equipage environment). The objective was to examine how the type of automation imperfection (miss vs. false alarm) affects ATSP performance and attention allocation. Research has shown that the type of automation imperfection leads to differential human performance costs. Participating in four 30-min scenarios were 12 full-performance-level ATSPs. Dependent variables included conflict detection and resolution performance, eye movements, and subjective ratings of trust and self confidence. ATSPs detected conflicts faster and more accurately with reliable automation, as compared with manual performance. When the conflict probe automation was unreliable, conflict detection performance declined with both miss (25% conflicts detected) and false alarm automation (50% conflicts detected). When the primary task of conflict detection was automated, even highly reliable yet imperfect automation (miss or false alarm) resulted in serious negative effects on operator performance. The further in advance that conflict probe automation predicts a conflict, the greater the uncertainty of prediction; thus, designers should provide users with feedback on the state of the automation or other tools that allow for inspection and analysis of the data underlying the conflict probe algorithm.

Effects of in-vehicle navigation assistance and performance on driver trust and vehicle control

Driver trust in in-vehicle navigation systems may be an important factor influencing use of such automation. This research investigated the trust effect of in-vehicle navigation aiding performance that degraded over time and assessed differences in trust among automated versus human aids in a simulated driving task. Twenty participants drove through a suburb area following traffic signs and directions from one of the navigation aids, the human aid via a cell phone or the in-vehicle automation aid presented on a separate simulator display, with different levels of aid information accuracy in each trial (100%, 80% and 60%). A control condition was also investigated to assess any benefit of aiding in which participants were presented with a telemarketing survey instead of navigation information. A 2×4 mixed experimental design was used with the two levels of aid type (manipulated between-subjects) and the four levels of information accuracy (manipulated within-subjects). Driver trust was measured using a subjective survey of initial trust expectations, as well as subjective ratings of trust at the close of each trial. Driving speed control deviations and the number of errors in driver conformance with posted signs and navigation instructions were recorded. In agreement with prior work in other domains, results demonstrated drivers had a higher initial trust expectation for the automated aid than the human aid. However, once the automated aid was used, trust declined dramatically across degrading aid performance conditions. In general, there was a trend for trust to decrease more sharply with the automated aid than the human aid, but there was no statistical difference among the aids. Trust declined as both aids degraded in performance. Driving errors also increased as the navigation aid performance decreased, and the control condition produced the highest number of errors. This study demonstrated the role of driver trust in in-vehicle navigation aid use and has implications for designing navigation systems that support trust and performance.

Measurement of trust over time in hybrid inspection systems

A study was conducted to measure the effects of human trust and to determine how it develops over time in a hybrid inspection system given different types of errors (i.e., false alarms and misses). The study also looked at which of the four dimensions of trust (competence, predictability, reliability, and faith) were the best predictors of overall trust. Results from the study showed that trust is sensitive to the type of errors made by a system. There was a significant change in overall trust between the stages for the conservative and risky systems, but no significant change in the neutral system. In regards to the best predictors of trust, faith appeared as one of the predictors in all three trial blocks for the conservative and risky systems. As time progressed, predictability emerged in the second and third trial blocks for the conservative system. Competence played an important role in the development of trust for the risky system, whereas reliability played an important role for the neutral system. These results suggest that subjective ratings of trust and the properties of the system can be used to predict the allocation of functions in hybrid inspection systems. © 2005 Wiley Periodicals, Inc. Hum Factors Man 15: 177–196, 2005.

Measurement of human trust in a hybrid inspection system based on signal detection theory measures

Human trust plays an important role in influencing operator's strategies toward the use of automated systems. Therefore, a study was conducted to measure the effect of human trust in a hybrid inspection system given different types of errors (i.e., false alarms and misses). The study also looked at which of the four dimensions of trust (competence, predictability, reliability and faith) were the best predictors of overall trust. Results from the study reveal that trust is sensitive to the type of errors made by a system and suggest that subjective ratings of trust and the properties of the system can be used to predict the allocation of functions in hybrid inspection systems. Relevance to industry The study conducted here is applicable to inspection tasks in manufacturing and service industries. The results obtained indicate that subjective ratings of operators’ trust can be used as a basis for predicting and optimizing operator's allocation behavior and system performance. Furthermore, designers can use these results to help decide which functions to allocate to the human or to the system based on previous experiences and interaction with the system.

Trust in automation. Part II. Experimental studies of trust and human intervention in a process control simulation.

Two experiments are reported which examined operators' trust in and use of the automation in a simulated supervisory process control task. Tests of the integrated model of human trust in machines proposed by Muir (1994) showed that models of interpersonal trust capture some important aspects of the nature and dynamics of human-machine trust. Results showed that operators' subjective ratings of trust in the automation were based mainly upon their perception of its competence. Trust was significantly reduced by any sign of incompetence in the automation, even one which had no effect on overall system performance. Operators' trust changed very little with experience, with a few notable exceptions. Distrust in one function of an automatic component spread to reduce trust in another function of the same component, but did not generalize to another independent automatic component in the same system, or to other systems. There was high positive correlation between operators' trust in and use of the automation; operators used automation they trusted and rejected automation they distrusted, preferring to do the control task manually. There was an inverse relationship between trust and monitoring of the automation. These results suggest that operators' subjective ratings of trust and the properties of the automation which determine their trust, can be used to predict and optimize the dynamic allocation of functions in automated systems.