the influence of cultural factors on trust in automation (fa9550-13-1-0129) pi: katia sycara...

38
The influence of Cultural Factors on Trust in Automation (FA9550-13-1-0129) PI: Katia Sycara (Carnegie Mellon University, USA) Co-PI: Michael Lewis (University of Pittsburgh, USA) Co-PI: Asiye Kumru (Ozyegin University, Turkey) Co-PI: Jyi-Shane Liu (National Chengchi University, Taiwan) AFOSR Program Review: Trust and Influence (May 11, 2015, Colorado Springs, CO)

Upload: emil-gallagher

Post on 29-Jan-2016

221 views

Category:

Documents


0 download

TRANSCRIPT

The influence of Cultural Factors on Trust in Automation (FA9550-13-1-0129)

The influence of Cultural Factors on Trust in Automation(FA9550-13-1-0129)PI: Katia Sycara (Carnegie Mellon University, USA)Co-PI: Michael Lewis (University of Pittsburgh, USA)Co-PI: Asiye Kumru (Ozyegin University, Turkey)Co-PI: Jyi-Shane Liu (National Chengchi University, Taiwan)

AFOSR Program Review: Trust and Influence (May 11, 2015, Colorado Springs, CO)Research MotivationAdvantages of automation (e.g. delivering more accurate information, lowering workload, allowing for more rapid decisions) may not be realized due to maladaptive use of the automation (Parasuraman and Riley 1997; Parasuraman and Manzey 2010; Lyons and Stokes 2012)Disuse/under-reliance: the human may fail to use the automation when it would be advantageous (e.g. ignoring or turning off alarms)Misuse/over-reliance/complacency: human accepts automations recommendations and/or actions when it is inappropriate (e.g. fail to monitor automation)Abuse: inappropriate application of automation by managers or designers.

Need for ways to ensure appropriate use of automationTrustMultiple studies have shown that trust mediates automation use and reliance, and in particular is involved in automation misuse and disuse (Lee and See 1994; Lee and Moray 1992; Muir 1988; Lewandowsky et al 2000; Merritt and Ilgen 2008)

Trust (Lee and See 2004) is an attitude that an agent (automation or another person) will help achieve an individuals goals in a situation characterized by uncertainty and vulnerability.Trust has been considered to consist of a set of attributional assumptions (trust dimensions) that range from trustees competence to its intentionsTrust is dynamicTrust establishmentTrust dissolution after occurrence of faultsTrust restoration

Factors that Affect TrustCharacteristics of the automationReliabilityNature of system faultsPredictabilityIntelligibility and TransparencyLevel of AutomationProperties of the OperatorPropensity to trustSelf ConfidenceCultureEnvironmental FactorsRisk

Current LimitationsDearth of studies on the influence of culture on trust in automation Current standardized measures for trust in automation are lacking (ad hoc measures, conceptualizations by analogy to interpersonal trust)Dearth of studies on the dynamics of trust in automation

Research GoalsDevelop a fundamental understanding of general principles and factors pertaining to dynamic trust in automation How trust mediates reliance on automation across cultures

Research StrategyReliable psychometric instrument that captures the nature and antecedents of trust in automation across culturesInitial pooling of items from 3 current measures Empirically Derived, Human-Computer Trust and SHAPE Automation Trust IndexTrust-Sensitive Task (TST) and associated computational and simulation infrastructure to enable replication studies of known effects so as to allow valid cross cultural comparisonsMulti-UAV control simulationTheoretically guided experimental studies to determine how cultural factors affect various aspects of trust and reliance on automationUse Hofstedes cultural dimensions plus cultural syndromes (Dignity, Honor, Face) (Triandis 1994;96, Cohen and Leung 2009; Leung and Cohen 2011)US-DignityTurkey-HonorTaiwan-Face

Cultural Dimensions and SyndromesDignity CulturesFound in Western Europe and North AmericaIndependent individuals focusing on personal goalsEgalitarian system backed up be effective system of lawTend to make swift trust assumption, i.e. trust until proven otherwiseHonor CulturesMiddle East, Latin America, Southern USExternally driven self-worthEmphasis on protecting self and family, not allowing to be taken advantage ofUnstable social hierarchiesLow interpersonal and institutional trustFace CulturesFar EastSelf worth extrinsically derived (what is important is view others have of you)Power and status is relatively hierarchical and stableSociety governed by norms, monitored and sanctions are imposedThus little need for interpersonal trust in general, high in-group trust

Comparison in Hofstede Cultural Dimensions

Power distance & Uncertainty avoidanceTurkey scores the highest PD (66) and UA (85) dimensionsUnited States scores PD (40) and UA (46)Taiwan falling in between, PD (58) and UA (69)Example Hypotheses Process TransparencyHypothesesFace culture operators will trust and more likely accept recommendations, even if the underlying process is unclearDignity and Honor culture operators will be relatively less likely to trust or accept recommendations if the process or purpose is not well understood Honor cultures will require greater support of trust from knowledge of process and/or purpose than Dignity and will be prone to disuseResearch ProductsReliable psychometric instrument that can be used across cultures to measure trust and its antecedents. This will be made widely available.Trust-Sensitive Task (TST) and associated computational and simulation infrastructure will become available to the scientific communityRecommendations and implications for interventions and training to enable proper trust calibration of human operators across culturesBasic Framework for Studying Trust in Automation Automation/environmentTrustDecisionOBSERVEMEASUREMANIPULATEFramework for Measuring Cross-cultural Effects of TrustDisposition to TrustIndividual/Cultural DifferencesTrustingBeliefsTrustingIntentionsTrust relatedBehaviorsTask variablesReliabilityRiskLOATask contextWorkloadSubjective measuresExperimentalmanipulationsObservedBehaviorsCross cultural samples, Hofstedes CVS, Big 5Distinguishing Predisposition to Trust from Trust of a Particular SystemItem pool contained items characterizing:General attitudes towards automationItems involving predisposition to trustI am confident in automationAttitudes across cultural-technological contexts E.g., uncertainty avoidance or subjective normsI feel okay using automation because it is backed by vendor protectionsAttitudes invoked after cueing to think about particular instances of automationThe advice GPS provided is as good as that which a highly competent person could produce

Cross-cultures validationDemographicsJune 16, 2014AFOSR Trust and Influence Program Review15SourcesUnited States(Army War College)Taiwan (Chengchi Univ.)Turkey(Ozyegin Univ.)Num of participants10012091GenderMale892946Female119145EducationHigh school0125Undergraduate57751Graduate954215Age< 200201221~300677631~40212141~506812051~602770> 6032215The Mturk info are shown on slide 26 15Cross-culture scale validationReliability Test (Cronbach's alpha)June 16, 2014AFOSR Trust and Influence Program Review16General AutoUnited States(War College)Taiwan (Chengchi Univ.)Turkey(Ozyegin Univ.)Performance /Ability.888.862.878Process/Integrity.869.856.855Purpose /Benevolence.844.777.850Task contexts .704.743.800Specific AutoUnited States(War College)Taiwan (Chengchi Univ.)Turkey(Ozyegin Univ.)Performance /Ability.847.859.903Process/Integrity.813.824.886Purpose /Benevolence.809.840.887All the Cronbach's alpha values exceed the threshold value of 0.7The Mturk reliability test are shown on slide 28

16Specific AutoMturkUnited StatesTaiwan TurkeyGermanCronbachs AVEAVEAVEAVEAVEPerformance 0.9870.7380.8470.5870.8590.5940.9030.6750.8540.585Process 0.9620.7180.8130.5310.8240.5390.8860.6390.8830.637Purpose 0.9790.6640.8090.5160.8400.5600.8870.6420.8710.614Cross Culture Specific Scale ValidationSuggestive threshold value- Reliability: Cronbachs > 0.7, Validity: AVE > 0.5Cultural Differences in Fullscale Means from SurveyAFOSR Trust and Influence Program Review18Mean ValuePerformProcessPurposeMturk3.824.003.83Army3.893.753.78TW3.373.463.40Turkey3.553.493.55Performance Process Purpose MANOVA shows cultural differences

All differences Significant exceptThose circled

Additional analysescontinuingMTurk ArmyTurkey TwFramework for Measuring Cross-cultural Effects of TrustDisposition to TrustIndividual/Cultural DifferencesTrustingBeliefsTrustingIntentionsTrust relatedBehaviorsTask variablesReliabilityRiskLOATask contextWorkloadSubjective measuresExperimentalmanipulationsObservedBehaviorsCross cultural samples, Hofstedes CVS, Big 5Trust Sensitive TaskStandard task(s) for evaluating effects of trust in automation across culturesNeeds to provide multiple tasks to produce effects such as complacencyNeeds to accommodate manipulations known to affect trust/trust-mediated behaviorsNeeds to be inclusive of types of tasks used in prior research

LOW 1 The computer offers no assistance, human must take all decisions and actions 2 The computer offers a complete set of decision/ action alternatives, or 3 Narrows the selection down to a few, or 4 Suggests one alternative, and 5 Executes that suggestion if the human approves, or 6 Allows the human a restricted veto time before automatic execution 7 Executes automatically, then necessarily informs the human, and 8 Informs the human only if asked, or Informs the human only if it, the computer, decides toThe computer decides everything, acts autonomously, ignores the humanHIGH Types of AutomationWho does it?What gets automated?Sheridan & Verplanck 1978 10 levels of automation (LOA)Parasuraman, Sheridan, & Wickens 2000Four stages of processing

LOW 1 The computer offers no assistance, human must take all decisions and actions 2 The computer offers a complete set of decision/ action alternatives, or 3 Narrows the selection down to a few, or 4 Suggests one alternative, and 5 Executes that suggestion if the human approves, or 6 Allows the human a restricted veto time before automatic execution 7 Executes automatically, then necessarily informs the human, and 8 Informs the human only if asked, or Informs the human only if it, the computer, decides toThe computer decides everything, acts autonomously, ignores the humanHIGH Levels of automation of Decision and Action Selection (Sheridan & Verplanck, 1978) Types and Levels of Automation (Parasuraman, Sheridan and Wickens, 2000) ClassifiersWho does it?What gets automated?Classifier?53% reviewed papersTypical Experiment:Parasuraman, Molloy, & Singh (1993)Monitoring gauges with reliable/unreliablealarm for out of range readings

LOW 1 The computer offers no assistance, human must take all decisions and actions 2 The computer offers a complete set of decision/ action alternatives, or 3 Narrows the selection down to a few, or 4 Suggests one alternative, and 5 Executes that suggestion if the human approves, or 6 Allows the human a restricted veto time before automatic execution 7 Executes automatically, then necessarily informs the human, and 8 Informs the human only if asked, or Informs the human only if it, the computer, decides toThe computer decides everything, acts autonomously, ignores the humanHIGH Levels of automation of Decision and Action Selection (Sheridan & Verplanck, 1978) Types and Levels of Automation (Parasuraman, Sheridan and Wickens, 2000) AdvisorWho does it?What gets automated?Expert system/advisor?26%, some classify &advise

Typical Experiment:Mosier, Skitka, Heers, & Burdick (1998)Altitude, heading, & radio frequencyChanges suggested by decision aid

LOW 1 The computer offers no assistance, human must take all decisions and actions 2 The computer offers a complete set of decision/ action alternatives, or 3 Narrows the selection down to a few, or 4 Suggests one alternative, and 5 Executes that suggestion if the human approves, or 6 Allows the human a restricted veto time before automatic execution 7 Executes automatically, then necessarily informs the human, and 8 Informs the human only if asked, or Informs the human only if it, the computer, decides toThe computer decides everything, acts autonomously, ignores the humanHIGH Levels of automation of Decision and Action Selection (Sheridan & Verplanck, 1978) Types and Levels of Automation (Parasuraman, Sheridan and Wickens, 2000) ControllerWho does it?What gets automated?Closed loop controller?17%, 5 of 7 involved single failureTypical Experiment:Chen & Barnes 2012Path planner- detect & replanTrust Sensitive Task for exploring culture x manipulation interactionModifying Cummings/Nehmeslight-weight RESCHU simulator to provide manipulations for task variables reported to affect trust, reliance, misuse, & disuse behaviorsTypes of interaction:Monitoring- ComplacencyClassifying- likelihood alarm systemNon-alerts (Miss-prone) ComplianceAlarms (FA-prone)Reliance Warnings (FA-prone)Uncertainty avoidanceDecision Aid- path planningautomation bias (FA-prone)Construct-focused manipulationsPlanning transparencyTask characteristics:Workload & Auto reliabilitySystem Feedback

25Trust Sensitive TaskPayload TaskclassifierNavigation Taskclosed-loop oradvisorTrust SensitiveTaskReliability

workload forNavigation TaskReliabilityLOATransparency

workload forPayload TaskExperimental DesignBetween VariablesControlIndication OnlyRe-routing OnlyIndication & Re-routingHigh ReliabilityLow ReliabilityNavigation TaskPayload Tasks - ClassificationCHECK: after clicking CHECK receiving a picture with better resolution for further identifying the existence of the target HIT: after adding a box on the suspected target, the operator will be able to attack the targetSAFE: if the operator believes the assigned hostile target does not exist in the image, click SAFE to proceed with other tasks

28Payload Tasks Decision AidTarget Finder is used but the reliability is less than perfect Based on the automated diagnosis, the payload window will be highlighted in different colors and a square will be added on the suspected target Red: alarm condition, high likelihood to truly indicate the assigned targetYellow: warning condition, higher level of uncertainty, informing the operator that there might be the assigned targetGreen: non-alert condition, low possibility the assigned target is included in the image

29Payload Tasks

Indication without Re-routing

Conflict Detection is used to detect the UAV conflicts but with imperfect reliabilityOnce conflicts are detected the area on the map in which the conflict is predicted to occur is highlighted and alerting messages are displayed in the status panel.Operator must manually add waypoints to direct UAV(s) away from the conflict

Highlight the conflict spots without re-planning paths, the operator must manually add waypoints to avoid the conflicts 31Indication Only

Re-routing without Indication

Conflict Detection is used to detect UAV conflicts and generate new paths to resolve the conflicts. Paths follow Dubins curves which do not match expectations for straight trajectories. Reliability in conflict detection is imperfect.

Once a conflict is detected, new paths are automatically generated for the UAVs and the AUTO button in the status panel becomes available

Re-generate paths with no further auto information 33Re-routing Only

Re-routing with Indication

Conflict detection is used to detect the UAV conflicts and generate new paths for UAVs to avoid the conflictsTwo ways to apply the re-plan paths:Click the Red square shown on the map and then select Yes to apply the new pathsClick the AUTO button in the status panel to apply the new paths

Highlight the conflict spots and re-generate paths for UAVs to avoid the conflict 35Re-routing with Indication

Plans for Next YearStudies in US, Turkey and Taiwan, using the TST to determine differences in disposition of trust, trust formation, effects of trust on behavior

Publications Chien, S., Semnani-Azad, Z., Lewis, M.,& Sycara, K. (2014). An Empirical Model of Cultural Factors on Trust in Automation. Proceedings of the 58th Annual Meeting of the Human Factors and Ergonomics Society, Chicago, IL., October 27-31, 2014.

Chien, S., Semnani-Azad, Z., Lewis, M.,& Sycara, K. (2014). Towards the Development of an Inter-Cultural Scale to Measure Trust in Automation. Proceedings of the 2014 HCI International, Crete, Greece, June 22-27, 2014 (Best paper award)

Chien, S., Hergeth, S., Semnani-Azad, Z., Lewis, M.,& Sycara, K. (accepted). Cross-Country Validation of a Cultural Scale in Measuring Trust in Automation, 59th Annual Meeting of the Human Factors and Ergonomics Society

38