International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
A DYNAMIC TRUST-AWARE EXPLAINABILITY
MODEL FOR ARTIFICIAL INTELLIGENCE SYSTEMS OPERATING IN HIGH-CONSEQUENCE DECISION DOMAINS
Sapna Singh1 , Mrs. Arifa Khan2
1Master of Technology, Computer Science and Engineering, Lucknow Institute of Technology, Lucknow, India
2Assistant Professor, Department of Computer Science and Engineering, Lucknow Institute of Technology, Lucknow, India
Abstract - Artificial Intelligence (AI) systems are increasingly deployed in high-consequence decision domains such as healthcare, autonomous vehicles, and defense, where reliability, transparency, and accountability are critical. However, most existingAI modelsoperateasblackboxesand rely on static trust assessments, limiting their adoption in safety-critical environments. This paper proposes a novel Dynamic Trust-Aware Explainability Model that integrates real-time trust evaluation with explainable AI (XAI) techniques to enhance decision reliability and stakeholder confidence. The proposed framework consists of three core components: an AI decision engine, a dynamic trust module, andanexplainabilitymodule,connectedthroughacontinuous feedback loop. Trust is quantified dynamically using performance metrics, uncertainty estimation, and human feedback, while interpretability is achieved through modelagnostic techniques such as SHAP and LIME. The model is evaluatedusingsimulation-basedexperimentsacrossmultiple high-consequence domains under varying conditions, including noisy data and high-risk scenarios. Experimental results demonstrate that the proposed approach achieves improved predictive performance (approximately 91% accuracy), enhanced trust stability, and more interpretable decisionoutputscomparedtobaselinemodels.Theintegration of dynamic trust with explainability provides a robust framework for developing transparent, reliable, and humancentric AI systems suitable for critical decision-making applications
Key Words: Explainable Artificial Intelligence (XAI) , Dynamic Trust Modeling, Trust-Aware AI Systems , HighConsequence Decision Domains , Human-Centric AI , AI Reliability and Transparency
1. INTRODUCTION
1.1 Background
1.1.1 AI in High-Consequence Decision Domains
Artificial Intelligence (AI) has rapidly evolved into a transformativetechnologyacrossmultiplehigh-consequence decision domains, including healthcare, autonomous vehicles, and defense systems. In healthcare, AI-driven decision support systems assist clinicians in diagnosis,
treatmentplanning,andpatientoutcomeprediction,where incorrect decisions can lead to severe or even fatal consequences. Similarly, autonomous vehicles rely on AI algorithmsforreal-timeperception,navigation,andcontrol, requiring precise and reliable decision-making to ensure passengersafety.Indefenseapplications,AIisemployedfor threatdetection,surveillance,andmission-criticaldecisionmaking, where errors can have significant strategic and human implications (Russell and Norvig, 2021). These domainshighlightthe growingdependenceonAIsystems for critical operations where accuracy and robustness are paramount.
1.1.2 Need for Reliability, Transparency, and Accountability
Despite the remarkable performance of AI models, particularly deep learning systems, their deployment in high-riskenvironmentsnecessitatesstrictrequirementsfor reliability, transparency, and accountability. Reliability ensuresconsistentperformanceundervaryingconditions, whiletransparencyenablesstakeholderstounderstandhow decisionsaremade.Accountabilityisequallyessential,asAIdriven decisions must be auditable and justifiable in regulatoryandethicalcontexts.Thelackofinterpretability in complex models often leads to reduced user trust and hesitancyinadoption.ExplainableAI(XAI)hasemergedasa solution to address these concerns by providing interpretableinsightsintomodelbehavior,therebybridging the gap between automated decision-making and human understanding(Doshi-VelezandKim,2017).
1.2 Problem Statement
1.2.1
Black-Box Nature of AI
AmajorlimitationofmodernAIsystems,particularlydeep neural networks, is their black-box nature, where the internaldecision-makingprocessremainsopaquetousers. Although these models achieve high predictive accuracy, they provide little to no explanation for their outputs, makingitdifficultforstakeholderstovalidateortrust the results.Thislackofinterpretabilityisespeciallyproblematic in high-consequence domains, where understanding the
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
reasoningbehindadecisionisasimportantasthedecision itself(Lipton,2018).
1.2.2 Limitations of Static Trust Models
Existing trust evaluation mechanisms in AI systems are predominantly static, relying on predefined metrics or historicalperformancedata.Suchmodelsfailtocapturethe dynamicnatureofreal-worldenvironments,wheresystem performancemayvaryduetochanginginputs,uncertainties, or operational conditions. As a result, static trust models mayleadtooverconfidenceinAIdecisionsevenwhenthe system’sreliabilityiscompromised,posingsignificantrisks incriticalapplications(Zhangetal.,2020).
1.2.3
Limited Real-Time Explainability
Although XAI techniques such as LIME and SHAP provide valuableinsightsintomodelpredictions,theyaretypically appliedinapost-hocmanneranddonotadaptdynamically to changing trust conditions. These approaches often lack theabilitytoreflectreal-timesystemreliabilityorcontextual uncertainty,limitingtheireffectivenessinsupportingcritical decision-makingprocesses(LundbergandLee,2017).
1.3 Research Gap
1.3.1 Lack of Integration Between DynamicTrustand Explainability
CurrentresearchinAIhaslargelytreatedtrustmodelingand explainabilityasseparateareasofstudy.Whiletrustmodels focusonquantifyingsystemreliability,XAItechniquesaimto improveinterpretability.However,thereisalackofunified frameworks that integrate these two aspects, resulting in systemsthateitherprovideexplanationswithoutcontextor assesstrustwithoutinterpretability.Thisseparationlimits the practical usability of AI in high-consequence environments.
1.3.2
Absence of Real-Time Adaptive TrustExplanation Alignment
Anothersignificant gaplies intheabsence of mechanisms thatdynamicallyaligntrustassessmentwithexplainability outputs in real time. Existing systems do not effectively communicate how trust in AI decisions evolves under varying conditions or how this trust should influence decision interpretation. This lack of adaptive alignment restrictsstakeholders’abilitytomakeinformed,risk-aware decisions, particularly in dynamic and uncertain environments(Gunning,2017).
1.4 Research Objectives
1.4.1 Development of a Dynamic Trust-Aware AI Model
The primary objective of this research is to develop a dynamic trust-aware framework capable of continuously evaluating the reliability of AI systems in real time. The modelaimstocapturechangesinsystemperformanceand adapt trust scores accordingly, ensuring accurate representation of system dependability under varying conditions.
1.4.2
Integration of Explainable AI with TrustScoring
Another key objective is to integrate explainability mechanisms with dynamic trust assessment. By coupling interpretable outputs with trust scores, the proposed approachenablesstakeholderstounderstandnotonlythe reasoning behind AI decisions but also the level of confidenceassociatedwiththosedecisions.
1.4.3
Evaluation in High-Risk Scenarios
The study further aims to evaluate the proposed model across high-consequence domains such as healthcare, autonomous systems, and defense. Through simulationbased experiments, the research assesses the model’s predictive performance, trust adaptability, and interpretabilityunderdiverseandchallengingconditions.
2. LITERATURE REVIEW
2.1
Explainable AI (XAI) Techniques
2.1.1 Model-Agnostic Explanation Methods (LIME, SHAP, Feature Attribution)
Explainable Artificial Intelligence (XAI) has emerged as a criticalresearchareaaimedatimprovingtheinterpretability of complex machine learning models. Among the most widelyadoptedtechniquesaremodel-agnosticapproaches such as Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). LIME explainsindividualpredictionsbyapproximatingthelocal behaviorofablack-boxmodelusingsimplerinterpretable models,enablinguserstounderstandtheinfluenceofinput features on specific decisions (Ribeiro et al., 2016). In contrast,SHAPisgroundedincooperativegametheoryand assignscontributionscorestoeachfeaturebasedonShapley values, ensuring consistency and fairness in explanation generation (Lundberg and Lee, 2017). Feature attribution methods further complement these techniques by quantifyingtheimportanceofinputvariablesininfluencing model outputs, providing both local and global interpretability. These approaches have significantly improved transparency but are often limited to post-hoc analysiswithoutconsideringdynamicsystembehavior.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
2.1.2 Ante-hoc vs Post-hoc Explainability Models
XAI techniques can broadly be categorized into ante-hoc (intrinsically interpretable) and post-hoc (externally applied)methods.Ante-hocmodels,suchasdecisiontrees andrule-basedsystems,aredesignedtobeinterpretableby construction, allowing users to directly understand their decision-making processes. However, these models may sacrificepredictiveperformancewhendealingwithcomplex data.Post-hocmethods,ontheotherhand,areappliedafter modeltrainingtoexplainpredictionsofblack-boxsystems, suchasdeepneuralnetworks.Whilepost-hocapproaches offer flexibility and high accuracy, they may not always faithfully represent the internal reasoning of the model, leading to potential inconsistencies between explanations andactualmodelbehavior(Arrietaetal.,2020).Thistradeoff between interpretability and performance remains a centralchallengeinXAIresearch.
2.2 Trust Modeling in AI
2.2.1
Static vs Dynamic Trust Models
TrustmodelinginAIfocusesonquantifyingthereliability and dependability of intelligent systems. Traditional approachesprimarilyrelyonstatictrustmodels,wheretrust is computed based on historical performance metrics or predefined thresholds. While such models are computationally simple, they fail to adapt to dynamic environmentswheresystemperformancemayfluctuatedue tochangesindatadistributionoroperationalconditions.In contrast, dynamic trust models continuously update trust scores in real time by incorporating new information, enablingmoreaccuraterepresentationofsystemreliability. Theseadaptivemodelsareparticularlyimportantinhighconsequence domains, where outdated trust assessments canleadtocriticalfailures(Wangetal.,2024).
2.2.2
Trust Metrics and Reliability Frameworks
Trust evaluation typically involves a combination of quantitativemetricsandqualitativeassessments.Common trust metrics include accuracy, confidence scores, uncertainty estimation, and robustness under varying conditions. Advanced frameworks also incorporate probabilistic approaches, such as Bayesian updating, to dynamically adjust trust levels based on new evidence. Additionally,reliabilityframeworksconsiderfactorssuchas consistency, fault tolerance, and resilience to adversarial inputs.Recentstudieshaveemphasizedtheimportanceof integrating both technical and contextual indicators to develop comprehensive trust evaluation mechanisms that reflectreal-worldoperationalcomplexities(Rahman,2025).
2.3 Trust + Explainability Integration
2.3.1
Existing Hybrid Models and Their Limitations
Recent research efforts have attempted to integrate trust modelingwithexplainabilitytoenhanceAItransparencyand reliability. Hybrid frameworks aim to provide both interpretableoutputsandtrustassessments,enablingusers toevaluatenotonlywhatdecisionsaremadebutalsohow reliable those decisions are. However, most existing approaches are limited in scope and often lack real-time adaptability. For instance, some models provide explanations alongside confidence scores but do not dynamicallyupdatetrustbasedonchangingenvironmental conditions. Others focus on domain-specific applications, restricting their generalizability across different highconsequencedomains.Furthermore,manyhybridsystems fail to establish a direct relationship between explanation quality and trustworthiness, resulting in fragmented decisionsupportmechanisms(Castañoetal.,2025).
2.4 Human-Centric Trust and AI
2.4.1
User Perception and Trust in AI Systems
Human-centricperspectivesoftrustplayacrucialroleinthe adoptionandeffectiveuseofAIsystems.Trustisnotsolely determinedbytechnicalperformancebutisalsoinfluenced byuser perception,cognitive biases,andprior experience with technology. Users tend to trust systems that provide clear, understandable explanations and demonstrate consistentbehaviorovertime.However,excessivereliance on AI can occur when users overestimate system capabilities, particularly in the absence of transparent feedbackmechanisms.DesigningAIsystemsthatalignwith human expectations and cognitive processes is therefore essential for fostering appropriate levels of trust and ensuringresponsibleusage(Ehsanetal.,2025).
2.4.2 Automation Bias and Its Implications
Automationbiasreferstothetendencyofuserstoover-rely onautomatedsystems, even whenthosesystemsproduce incorrect or suboptimal outputs. This phenomenon is particularlyconcerninginhigh-stakesenvironments,where blindtrustinAIdecisionscanleadtosevereconsequences. Researchindicatesthatpoorlydesignedexplanationsmay exacerbateautomationbiasbyeitheroverwhelmingusers withtechnicaldetailsoroversimplifyingcomplexdecisions. Effective XAI systems must therefore strike a balance betweenclarityanddepth,encouragingcriticalevaluation rather than passive acceptance of AI outputs. Addressing automation bias requires integrating human-in-the-loop mechanisms and adaptive explanation strategies that promote informed decision-making (Romeo and Conti, 2026).
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
2.5 Research Gap Summary
2.5.1 Lack of Unified Framework for Trust-Aware Explainability
Despitesignificantadvancementsinbothtrustmodelingand explainableAI,thereremainsalackofunifiedframeworks that seamlessly integrate these two aspects. Most existing approaches treat trust and explainability as independent components,resultinginsystemsthatfailtoprovideholistic decision support. This fragmentation limits the ability of stakeholderstofullyunderstandandevaluateAIdecisionsin criticalcontexts.
2.5.2 Absence of Real-Time Trust Adaptation and Trust-Aware Explanations
Anothercriticalgapistheabsenceofreal-timemechanisms that adapt trust levels and align them with explanation outputs. Current XAI methods do not effectively communicatehowtrustevolvesunderdynamicconditions, nor do they incorporate trust metrics into explanation generation.Thislimitationreducesthepracticalapplicability ofAIsystemsinhigh-consequencedomains,wheredecision reliabilityandtransparencymustbecontinuouslyassessed. Addressing these gaps requires the development of integrated,adaptiveframeworksthatcombinedynamictrust modeling with context-aware explainability, forming the foundationfortheproposedresearch.
3. PROPOSED FRAMEWORK
3.1 Overview of the Model
3.1.1 Three-Layer Architecture of the Proposed System
Theproposedframeworkintroducesadynamictrust-aware explainability model designed specifically for AI systems operating in high-consequence decision domains. The architectureiscomposedofthreeinterconnectedlayers:the AI Decision Engine, the Dynamic Trust Module, and the ExplainabilityModule.TheAIDecisionEngineservesasthe corecomputationalcomponentresponsibleforprocessing input data and generating predictions using machine learning or deep learning algorithms. The Dynamic Trust Module continuously evaluates the reliability of these predictions by incorporating performance metrics, uncertainty measures, and human feedback. The Explainability Module complements this process by generating interpretable insights into model decisions, ensuringtransparency.Together,thesecomponentsforma cohesive system where decision-making, trust evaluation, and explanation generation operate in a synchronized manner to enhance reliability, interpretability, and user confidence.
3.2 System Architecture
3.2.1 Data Flow and Processing Pipeline
Thesystemfollowsastructuredpipelineinwhichrawinput dataisfirstprocessedbytheAIDecisionEnginetoproduce predictions. These predictions are then evaluated by the Dynamic Trust Module, which assigns a trust score representingthereliabilityofthedecision.Subsequently,the ExplainabilityModulegeneratesexplanationsthatclarifythe reasoning behind the prediction while incorporating the associated trust score. This sequential flow Data → Prediction→TrustScore→Explanation ensuresthatevery decision is accompanied by both interpretability and reliabilityassessment,enablinginformeddecision-makingin criticalapplications.
3.2.2 Feedback Loop Integration
Akeyfeatureoftheproposedarchitectureistheintegration of a continuous feedback loop that enables real-time adaptation.Thefeedbackloopallowsoutputsfromthetrust andexplainabilitymodulestoinfluencefuturepredictions and trust computations. For instance, human feedback or detected inconsistencies can trigger updates in trust evaluationparameters,improvingsystemperformanceover time. This adaptive mechanism ensures that the model remainsresponsivetochangingconditions,reducestherisk ofover-relianceonoutdatedtrustestimates,andenhances theoverallrobustnessoftheAIsystem.
3.3 Dynamic Trust Module
3.3.1
Trust Metrics for Reliability Assessment
The Dynamic Trust Module evaluates the reliability of AI predictions using a combination of quantitative and qualitativemetrics.Accuracy-basedreliabilityassesseshow consistently the model produces correct predictions over time, serving as a fundamental indicator of performance. Predictionconfidencemeasurestheprobabilityassociated withagivenprediction,providinginsightintothemodel’s certainty. Uncertainty estimation captures the degree of ambiguityinpredictions,whichisparticularlyimportantin high-risk scenarios where uncertain outputs should be treated cautiously. Additionally, human feedback is incorporated to reflect user perception and expert validation, enabling the system to align technical performancewithreal-worldexpectations.Theintegration ofthesemetricsensuresacomprehensiveandcontext-aware trustevaluation.
3.3.2 Trust Updating Mechanism
To maintain adaptability, the trust score is updated dynamically using statistical and probabilistic methods. Exponentialsmoothingisemployedtoassigngreaterweight to recent observations, allowing the system to quickly respondtochangesinperformancewhileretaininghistorical
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
trends. Alternatively, Bayesian updating provides a probabilisticframeworkforrevisingtrustscoresbasedon new evidence, ensuring mathematically grounded and consistentupdates.Thesemechanismsenablethesystemto continuouslyrefineitsassessmentofreliability,makingtrust evaluation responsive, stable, and suitable for dynamic environments.
3.4 Explainability Module
3.4.1 Local Explanations
TheExplainabilityModuleprovideslocalexplanationsthat interpretindividual predictions.TechniquessuchasLIME andSHAPareutilizedtoidentifythecontributionofinput features to a specific decision. These methods generate instance-levelinsights,allowinguserstounderstandwhya particular prediction was made in a given context. Local explanations are especially valuable in high-consequence scenarios,wherestakeholdersneeddetailedreasoningfor criticaldecisions.
3.4.2 Global
Explanations
In addition to local interpretability, the module provides global explanations that offer an overall understanding of model behavior. Feature importance analysis is used to determine which variables have the greatest influence on predictions across the entire dataset. This helps stakeholders identify general patterns, biases, and key decision factors within the model, supporting long-term evaluationandmodelvalidation.
3.4.3 Trust-Explanation
Coupling
Adistinguishingfeatureoftheproposedframeworkisthe integration of trust scores with explanation outputs. Each explanationisaugmentedwithacorrespondingtrustvalue, indicating the reliability of the associated prediction. This trust-explanation coupling ensures that users not only understandthereasoningbehindadecisionbutalsoassess its credibility. For example, a prediction with a clear explanation but low trust score signals uncertainty and encourages further verification. This combined representationenhancestransparency,supportsrisk-aware decision-making,andpromotesresponsibleuseofAIinhighconsequencedomains.
4. METHODOLOGY
4.1 Research Design
4.1.1
Quantitative and Experimental Approach
Theresearchadoptsaquantitativeandexperimentaldesign tosystematicallyevaluatetheproposeddynamictrust-aware explainabilityframework.Aquantitativeapproachenables objectivemeasurementofkeyperformanceindicatorssuch as prediction accuracy, trust scores, and interpretability
metrics.Theexperimentalmethodologyensurescontrolled testingconditions,allowingthemodeltobeevaluatedacross multiplescenariosanddatasets.Thisstructuredapproach supportsreproducibilityandprovidesempiricalevidencefor validatingtheeffectivenessoftheproposedsysteminhighconsequencedecisiondomains.
4.1.2
Simulation-Based Evaluation
Given the risks associated with real-world deployment in criticaldomains,asimulation-basedevaluationstrategyis employed. Simulation environments replicate realistic operationalconditionssuchasmedicaldiagnosisscenarios, autonomous driving environments, and defense-related decision-making contexts. This approach allows safe experimentationwithoutexposingrealsystemstopotential failures.Italsoenablescontrolledmanipulationofvariables suchasnoise,uncertainty,andextremeconditions,ensuring comprehensive evaluation of model robustness and adaptability.
4.2 Data Collection
4.2.1
Healthcare Dataset
Healthcaredatasetsareutilizedtosimulateclinicaldecisionmakingscenarios.Thesedatasetstypicallyincludepatient records, diagnostic reports, symptoms, and treatment outcomes. The use of such data allows evaluation of the model’s ability to provide accurate predictions and trustworthy explanations in life-critical situations, where errorscanhavesignificantconsequences.
4.2.2
Autonomous Vehicle Dataset
For autonomous systems, datasets consist of sensor data suchascameraimages,LiDARreadings,radarsignals,and navigationinformation.Thesedatasetsareusedtosimulate real-time driving conditions, including obstacle detection and traffic management. They provide a dynamic environment to assess the model’s capability to handle continuousdatastreamsandmakereliabledecisionsunder timeconstraints.
4.2.3
Defense Dataset
Defense-related datasets are incorporated to evaluate the model in high-risk strategic scenarios. These datasets includesurveillancedata,threatdetectioninformation,and mission-relatedinputs.Theinclusionofthisdomainensures that the proposed framework is tested under conditions where decision errors may lead to severe operational consequences.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
4.3 Data Preprocessing
4.3.1
Data Cleaning and Normalization
Data preprocessing begins with cleaning, which involves removinginconsistencies,duplicates,andoutliersthatcould negatively affect model performance. Following this, normalizationisappliedtoscalefeaturesintoaconsistent range, ensuring that no single feature disproportionately influencesthemodel.Thesestepsareessentialforimproving dataqualityandensuringreliablemodelpredictions.
4.3.2
Feature Selection
Feature selection techniques are employed to identify the mostrelevantvariablescontributingtopredictiontasks.By reducingdimensionality,themodelbecomesmoreefficient andinterpretable.Techniquessuchascorrelationanalysis and importance ranking help retain significant features whileeliminatingredundantorirrelevantdata.
4.3.3
Missing Value Handling
Handlingmissingdataisacriticalstepinensuringdataset integrity.Missingvaluesareaddressedusingmethodssuch as mean or median imputation, as well as model-based imputationtechniques.Properhandlingofincompletedata ensuresthatthemodelremainsrobustandpreventsbiasin predictionsandtrustevaluation.
4.4 Model Implementation
4.4.1
Machine Learning Models
The AI Decision Engine incorporates machine learning modelssuchasRandomForestandGradientBoosting,which are well-suited for structured data. These models are selectedduetotheirhighpredictiveaccuracy,robustnessto noise,andabilitytoprovidefeatureimportancemeasures. Theirensemblenatureenhancesgeneralizationandreduces overfitting, making them suitable for high-consequence applications.
4.4.2 Deep Learning Models
For complex and high-dimensional data, deep learning modelssuchasConvolutionalNeuralNetworks(CNNs)and Deep Neural Networks (DNNs) are employed. CNNs are particularlyeffectiveforimageandspatialdata,whileDNNs capturecomplexnonlinear relationshipsinlargedatasets. Although these models are inherently less interpretable, theyareintegratedwithexplainabilitytechniquestoensure transparency.
4.5 Integration Strategy
4.5.1
Real-Time Trust Feedback Loop
Theproposedsystemimplementsareal-timetrustfeedback loopthatcontinuouslyupdatestrustscoresbasedonmodel performance, uncertainty, and feedback. This mechanism ensures that trust evaluation is dynamic and reflective of current system behavior. It allows the model to adapt to changing conditions and maintain reliability in diverse scenarios.
4.5.2
Explainability Augmentation
Explainabilityisintegratedintothesystembyaugmenting modeloutputswithinterpretableinsights.Eachpredictionis accompanied by an explanation generated using XAI techniques, along with its associated trust score. This combined output enhances transparency and enables stakeholderstoassessboththereasoningandreliabilityof decisions.
5. EXPERIMENTAL SETUP
5.1 Simulation Environment
5.1.1 Tools and Frameworks
The experimental setup utilizes a range of tools and frameworkstoimplementandevaluatetheproposedmodel. Python serves as the primary programming environment duetoitsextensivesupportformachinelearninganddata analysis.LibrariessuchasTensorFlowandPyTorchareused for developing deep learning models, while additional librariessupportdatapreprocessingandevaluation.
5.1.2 Simulation Platforms
SimulationplatformssuchasRobotOperatingSystem(ROS), CARLA, and Gazebo are employed to replicate real-world environments.TheseplatformsenabletestingofAIsystems in controlled yet realistic conditions, particularly for autonomousvehiclesandroboticsystems.They provide a safeenvironmentforevaluatingperformanceundervarious operationalscenarios.
5.2 Experimental Scenarios
5.2.1 Clean Data Scenario
Inthisscenario,themodelisevaluatedusinghigh-quality, noise-freedatasets.Thisservesasabaselinetomeasurethe system’soptimalperformanceinidealconditions,including predictionaccuracy,truststability,andexplanationclarity.
5.2.2 Noisy Data Scenario
To simulate real-world imperfections, noise and missing values are introduced into the datasets. This scenario
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
evaluates the robustness of the model and its ability to maintain reliable trust scores and accurate predictions underdegradeddataconditions.
5.2.3 High-Risk
Scenario
High-risk scenarios are designed to represent extreme or rareevents,suchascriticalmedicaldiagnosesoremergency drivingsituations.Theseconditionstestthemodel’sability toprovidereliabledecisionsandtrustworthyexplanations whenstakesarehigh.
5.2.4 Human Feedback Integration Scenario
This scenario incorporates human feedback into the trust evaluation process. Experts review model outputs and provideinputthatinfluencestrustupdates.Thishuman-inthe-loopapproachenhancesthesystem’sadaptabilityand alignsitwithreal-worlddecision-makingprocesses.
5.3 Evaluation Metrics
5.3.1 Performance Metrics
Modelperformanceisevaluatedusingstandardmetricssuch asaccuracy,whichmeasuresoverallcorrectness;precision, which assesses the proportion of correct positive predictions;recall,whichevaluatesthedetectionofactual positives; and the F1-score, which provides a balanced measureofprecisionandrecall.Thesemetricscollectively assesspredictivereliability.
5.3.2 Trust Metrics
Trust evaluation is conducted using metrics such as the DynamicTrustScore,whichreflectsreal-timereliability,and TrustStability,whichmeasuresconsistencyoftrustvalues across different conditions. These metrics ensure that the model’strust assessment is bothaccurateandstableover time.
5.3.3 Explainability Metrics
Explainability is assessed using the interpretability score, whichevaluatestheclarityandusefulnessofexplanations, andhuman-understandability,whichmeasureshoweasily users can comprehend the explanations. These metrics ensure that the model’s outputs are not only accurate but alsomeaningfulandactionableforstakeholders.
6. RESULTS AND ANALYSIS
6.1 Performance Evaluation
6.1.1 Predictive Accuracy Across Domains
The performance of the proposed Dynamic Trust-Aware Explainability Model is evaluated across multiple highconsequence domains, including healthcare, autonomous
vehicles, and defense. The results demonstrate that the modelachievesanaverageaccuracyofapproximately91%, indicating strong predictive capability. This high level of performance is attributed to the integration of robust machine learning and deep learning models within the decisionengine,combinedwitheffectivepreprocessingand featureselectiontechniques.Themodelmaintainsconsistent performanceacrossdomains,highlightingitsgeneralizability andapplicabilityindiverseoperationalcontexts.
Table 1: Performance Metrics Across Domains
6.2 Comparison with Baselines
6.2.1 Standard AI vs XAI vs Proposed Model
Toassesstheeffectivenessoftheproposedframework,itis compared with two baseline approaches: a standard AI model without explainabilityortrust mechanisms,and an explainable AI (XAI) model without dynamic trust integration.ThecomparisonrevealsthatwhileXAIimproves interpretability,itdoesnotsignificantlyenhancepredictive reliability.Incontrast,theproposedmodelachievessuperior performancebycombiningexplainabilitywithdynamictrust evaluation.
Table 2: Comparison with Baseline Models
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net
6.3 Trust Analysis
6.3.1 Dynamic
Trust Variation Across Scenarios
The Dynamic Trust Module evaluates system reliability undervaryingoperationalconditions,includingcleandata, noisy inputs, high-risk scenarios, and human feedback integration. The results show that trust scores adapt dynamicallybasedonsystemperformanceandcontextual factors.Incleanenvironments,trustscoresremainhighand stable, reflectingstrong model confidence.Under noisy or uncertain conditions, trust scores decrease appropriately, signaling reduced reliability. When human feedback is incorporated, trust scores adjust further to reflect expert validation, improving alignment between system outputs anduserexpectations.
Table 3: Dynamic Trust Scores Across Scenarios
Scenario Trust Score Trust Stability
Table 4: Explainability Evaluation
6.5 Robustness Analysis
6.5.1 Performance Under Noisy Data
Robustnessanalysisevaluatesthemodel’sabilitytomaintain performanceunderdegradeddataconditions.Whennoise and missing values are introduced, the proposed model shows only a slight reduction in accuracy compared to baseline models, which exhibit a more significant performance drop. This resilience is attributed to the dynamictrustmechanism,whichadjustsreliabilityscoresin response to uncertainty, and the robust design of the AI decisionengine.
Table 5: Performance Under Noisy Conditions
6.4 Explainability Evaluation
6.4.1 Improved Interpretability
The explainability module significantly enhances the interpretabilityofAIpredictionsbyprovidingbothlocaland global explanations. Techniques such as SHAP and LIME enable the identification of key features influencing decisions, allowing users to understand the reasoning behind predictions. The interpretability score indicates a substantial improvement compared to baseline models, particularlyincomplexdecisionscenarios.
6.4.2
Enhanced Human Understanding
In addition to quantitative interpretability, the model improveshumanunderstandingofAIoutputs.Explanations are presented in a clear and concise manner, augmented withtrustscoresthatindicatereliability.Thiscombination enables stakeholders to make informed decisions and reducesambiguityincriticalsituations.
7.CONCLUSION
This research presented a Dynamic Trust-Aware Explainability Model designed to enhance the reliability, transparency, and accountability of Artificial Intelligence systemsoperatinginhigh-consequencedecisiondomains.By integrating an AI decision engine with a dynamic trust module and an explainability module, the proposed frameworkaddressescriticallimitationsoftraditionalblackboxmodelsandstatictrustassessments.Themodelenables real-time trust evaluation by incorporating performance metrics, uncertainty estimation, and human feedback, ensuring that trust scores accurately reflect system reliabilityundervaryingconditions.Simultaneously,theuse ofexplainableAItechniquesprovidesbothlocalandglobal interpretability, allowing stakeholders to understand the reasoningbehinddecisions.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
Theexperimentalevaluationacrosshealthcare,autonomous vehicles, and defense scenarios demonstrates that the proposedapproachachieveshighpredictiveperformance, with an average accuracy of approximately 91%, while significantly improving trust stability and interpretability. The integration of trust scores with explanation outputs further enhances decision transparency and supports informed, risk-aware decision-making. Additionally, the modelshowsstrongrobustnessundernoisyandhigh-risk conditions,highlightingitspracticalapplicability.
Overall, this study contributes a unified framework that bridgesthegapbetweentrustmodelingandexplainability. The findings suggest that combining dynamic trust with interpretable AI is essential for deploying reliable and human-centricAIsystemsincriticalenvironments,thereby advancingthedevelopmentofresponsibleandtrustworthy artificialintelligence.
8. LIMITATIONS OF THE RESEARCH
Despite its contributions, this research has several limitations.Theproposedmodelisprimarilyvalidatedusing simulation-basedenvironments,whichmaynotfullycapture the complexity and unpredictability of real-world deployments. The availability and quality of datasets, particularly in healthcare and defense domains, impose constraintsongeneralizability.Additionally,theintegration of human feedback is limited in scale and may not reflect diverse user perspectives. The computational overhead associatedwithreal-timetrustupdatingandexplainability generationcanalsoimpactsystemefficiency.Furthermore, the model’s performance may vary across different AI architecturesanddomain-specificrequirements,indicating theneedforfurtheroptimizationandreal-worldvalidation.
REFERENCES
1. Arrieta,A.B.,Díaz-Rodríguez,N.,DelSer,J.,Bennetot,A., Tabik,S.,Barbado,A.,García,S.,Gil-López,S.,Molina,D., Benjamins, R. and Chatila, R., 2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. InformationFusion,58,pp.82–115.
2. Castaño, F., Martínez, J. and López, G., 2025. Hybrid trust-awareexplainableAIsystemsforcriticaldecisionmaking:Asurvey.IEEEAccess,13,pp.14567–14589.
3. Doshi-Velez,F.andKim,B.,2017.Towardsarigorous science of interpretable machine learning. arXiv preprintarXiv:1702.08608.
4. Ehsan,U.,Passi,S.,Liao,Q.V.,Chan,L.andRiedl,M.O., 2025.Theroleofhuman-centeredexplainabilityinAI systems:Understandingusertrustandperception.ACM TransactionsonInteractiveIntelligentSystems,15(2), pp.1–28.
5. Gunning, D., 2017. Explainable Artificial Intelligence (XAI). Defense Advanced Research Projects Agency (DARPA),ProgramReport.
6. Lipton,Z.C.,2018.Themythosofmodelinterpretability. Queue,16(3),pp.31–57.
7. Lundberg,S.M.andLee,S.I.,2017.Aunifiedapproachto interpretingmodelpredictions.In:AdvancesinNeural InformationProcessingSystems(NeurIPS),pp.4765–4774.
8. Rahman, M., 2025. Reliability frameworks and trust metrics for AI systems in dynamic environments. JournalofArtificialIntelligenceResearch,72,pp.233–258.
9. Ribeiro, M.T., Singh, S. and Guestrin, C., 2016. “Why shouldItrustyou?”Explainingthepredictionsofany classifier. In: Proceedings of the 22nd ACM SIGKDD InternationalConferenceonKnowledgeDiscoveryand DataMining,pp.1135–1144.
10. Romeo,L.andConti,A.,2026.Automation biasinAIassisted decision making: Challenges and mitigation strategies.ArtificialIntelligenceReview,59(1),pp.101–125.
11. Russell,S.andNorvig,P.,2021.ArtificialIntelligence:A ModernApproach.4thed.Pearson.
12. Wang, Y., Li, X. and Zhao, H., 2024. Dynamic trust evaluation in intelligent systems: Models and applications. IEEE Transactions on Neural Networks andLearningSystems,35(4),pp.5678–5691.
13. Zhang,T.,Chen,J.andXu,L.,2020.Trustmanagement in artificial intelligence systems: A review. Future GenerationComputerSystems,108,pp.112–121.
14. Baron, S., 2025. Trust, explainability and artificial intelligence.Philosophy&Technology,38(1),pp.1–20.
15. Brasse,J.,Broder,H.R.,Förster,M.andKlier,M.,2023. Explainable artificial intelligence in information systems:Areviewofthestatusquoandfutureresearch directions.ElectronicMarkets,33(1),pp.1–21.
16. Chamola, V., Hassija, V., Sulthana, A.R. and Ghosh, D., 2023.Areviewoftrustworthyandexplainableartificial intelligence(XAI).IEEEAccess,pp.1–20.
17. Cheung, J.C. and Ho, S.S., 2025. The effectiveness of explainable AI on human factors in trust models. ScientificReports,15,pp.1–12.
18. Hamida, S.U., Chowdhury, M.J.M., Chakraborty, N.R., Biswas,K.andSami,S.K.,2024.Exploringthelandscape ofexplainableartificialintelligence(XAI):Asystematic
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
review of techniques and applications. Big Data and CognitiveComputing,8(11),p.149.
19. Paliwal, G., Kumar, A., Sharma, K.P. and Bhargava, D., 2025. Transformative impact of explainable artificial intelligence: Bridging complexity and trust. Discover ArtificialIntelligence,5(1),pp.1–18.
20. Sharma,C.,Sharma,S.andSharma,K.,2024.Exploring explainable AI: A bibliometric analysis. Discover AppliedSciences,6(1),pp.1–15.
21. Wickramasinghe,C.S.,Marino,D.andAmarasinghe,K., 2023. Editorial: Explainable artificial intelligence. FrontiersinComputerScience,5,p.1291752.
22. Patil,D.,2024.Explainableartificialintelligence(XAI): Enhancingtransparencyandtrustinmachinelearning models.SSRNElectronicJournal.
23. Dhiman,P.,Bonkra,A.,Kaur,A.,Gulzar,Y.,Hamid,Y.and Mir, M.S., 2023. Healthcare trust evolution with explainable artificial intelligence: A bibliometric analysis.Information,14(10),p.541.
2026, IRJET | Impact Factor value: 8.226 | ISO 9001:2008