Trust in AI: Progress, Challenges, and Future Directions

Motivation

• AI has rapidly diffused across everyday applications and outperforms conventional solutions in domains including health, transportation, security, and entertainment –.Its expanding role makes trust consequential for behavior, interaction, and technology acceptance.Source: §1 · p. 2
• Trust enables people to accept, rely on, support, and adopt AI, whereas distrust reduces diffusion and prevents stakeholders from realizing AI’s potential,.Trust depends on anticipating whether AI behavior matches users’ desires.Source: §1 · p. 2

Scope of Trust

• AI trustworthiness depends on technical factors such as accuracy, reliability, transparency, and explainability, alongside ethical, legal, human, and contextual considerations,.Unlike rule-based automation, AI can learn and make proactive, unexpected, or incomprehensible decisions from training data.Source: §1 · p. 2

Trust Barriers

• Black-box behavior, risk aversion, and regulatory, medicolegal, or ethical concerns have hindered AI adoption in manufacturing, medical imaging, autonomous vehicles, finance, and other domains –.Accordingly, trust drives AI usage while distrust acts as a barrier to development and application.Source: §1 · p. 3

Review Contribution

• The review synthesizes trust and distrust conceptions, models, impacts, technical and axiological metrics, distrust sources such as autonomy and dignity threats, and solutions for trustworthy and responsible AI.It addresses the literature gap through a systematic review of relevant theories, frameworks, measurements, and proposed resolutions.Source: §1 · p. 3

Review protocol

• The review followed a PRISMA-based protocol because no dedicated database for trust in AI existed.The protocol was developed for an inclusive systematic review of academic papers, reports, case studies, and AI trust frameworks written in English.Source: §2 · p. 5, Figure 1 · p. 5

Search strategy

• Searches combined manual identification of 19 papers with keyword-based Google Scholar searches covering trust, trustworthy AI, distrust, and related trust considerations.The keyword strategy included explainability, transparency, interpretability, empathy, privacy, fairness, accountability, safety, accuracy, and robustness; searches stopped when later pages yielded few new relevant papers.Source: §2 · pp. 5–6

Search and selection

• 329 target papers were selected from 336 relevant papers identified among 1,205 search results after duplicate removal and screening for journal publication and substantive focus on trust in AI.The dominant topic was assessed from the papers’ main sections rather than titles and keywords alone.Source: §2 · p. 6

Qualitative analysis

• Four researchers critically analyzed selected papers and developed eight major codes for categorizing the review findings.The codes formed the basis of the review’s results categorization in Table 2.Source: §3 · p. 6, Table 2 · p. 6

Trust in AI interactions

• The reviewed literature examines trust across human-machine, machine-human, and machine-machine interactions, including robo-advisors, autonomous vehicles, adversarial attacks, and governance mechanisms.Governance mechanisms include self-imposed standards, certification, corporate guidelines, and governmental regulations.Source: §3 · p. 7

Technology acceptance

• Trust and distrust affect AI acceptance across economic, healthcare, investment, assistant, and chatbot domains, with cognitive and emotional trust alongside deception and bias concerns.Examples include economic output, physician acceptance, reduced healthcare waiting time, algorithmic investment advice, and managers’ endorsement.Source: §3 · p. 7

Trustworthiness dimensions

• Major trustworthiness themes include transparency, explainability, interpretability, empathy, privacy, fairness, accountability, and technical safety, accuracy, and robustness.The literature also addresses opaque algorithms, privacy-by-design, algorithmic discrimination, reliability, security, lineage, and high-stakes decision-making.Source: §3 · p. 7

Metrics and frameworks

• Trustworthiness is evaluated through psycho-physiological, empirical, theoretical, qualitative, experimental, and multidimensional approaches, including certification, auditing, risk assessment, and performance metrics.Frameworks address ethics, fairness, accountability, human agency, privacy, data governance, technical robustness, safety, and societal and environmental wellbeing.Source: §3 · pp. 7–8

Trust breakers and makers

• Trust breakers include surveillance, manipulation, privacy loss, bias, autonomy and dignity threats, unpredictability, and loss of human control, while trust makers include transparency, explainability, certification, oversight, and value-based ethics.The literature also considers anthropomorphism, reputation, context, expertise, culture, marketing, performance, accountability, and risk-mitigating practices.Source: §3 · p. 8

1.1.1 Theories and definitions of trust in AI

• AI trust involves anticipating a system’s correctness or useful functionality under uncertainty, vulnerability, and risk, rather than necessarily anticipating its behavior or benevolence,,,,.Trust and trustworthiness can diverge: accurate systems may be distrusted without understandable logic, while poor systems may be trusted because of persuasive interfaces, [42].Source: §1.1.1 · pp. 9–10
• Observable cues—including interfaces, inputs, outputs, uncertainty, rationales, and company logos—shape perceived trustworthiness and can produce over-trust when engineered or misleading.Accurate assessment depends on cue relevance, availability, detection, and correct utilization, while trustworthy AI should address beneficence, non-maleficence, autonomy, justice, and explicability.Source: §1.1.1 · pp. 10–11

1.1.2 Trust in types of human-Machine interaction

• Human-machine, machine-human, and machine-machine interactions distribute trust across users, AI systems, data sources, and other agents, with privacy, adversarial attacks, authorization, and technical correctness shaping trust requirements,,,.AI-object interaction remains scarcely studied, including systems trusting signs, social-network data, theories, algorithms, institutions, and other complex objects.Source: §1.1.2 · pp. 11, 12, 15
• Trust calibration in safety-critical collaboration requires interpretable, uncertainty-aware systems that reveal limitations and trigger human control transfers when AI decisions are uncertain,,.Human traits, emotions, humanness, personality, and changing perceptions also affect trust, while bias, opacity, and poor explainability can expose users to unsafe or unethical decisions, [54],,,,.Source: §1.1.2 · pp. 14–15

1.1.3 Impact of trust/distrust on AI technology acceptance in different domains

• In healthcare, transparency and explainability support acceptance, whereas limited trials, opaque information flows, bias, regulatory uncertainty, low literacy, and unexplained AI diagnoses reduce trust and adoption,,,, [87],,.Users may value human doctors’ holistic and social roles, and explanation format and timing affect whether accurate AI diagnoses are accepted,.Source: §1.1.3 · p. 16
• Across finance, assistants, manufacturing, and robotics, adoption depends on trust shaped by human preference, empathy, transparency, managerial endorsement, robot characteristics, and cognitive-emotional evaluations,,,,,,,,.Only half of people trust algorithmic investment advice,; cognitive trust more strongly predicts recommendation-system adoption, while emotional and cognitive trust are correlated.Source: §1.1.3 · pp. 16–17

3.2. Trustworthy AI and its metrics: Technical, and non-technical metrics

• Trustworthy AI in healthcare requires reliable, safe, and ethical systems, with technical metrics complemented by ethical, legal, traceability, and performance-monitoring foundations.Clinical applications require attention to trust in AI-supported diagnosis and rehabilitation, while reliability limitations can constrain medical use,,,,.Source: §3.2 · p. 18
• The review identifies accuracy and safety as preliminary metrics, reliability as including fairness and robustness, and transparency and value alignment as essential for human-AI interaction.The associated ethics principles include privacy, fairness and justice, safety and reliability, transparency, and social responsibility and beneficence.Source: §3.2 · p. 18
• Trust-building approaches across autonomous robotics include standardized definitions, complaint systems, independent ratings, software validation, trust modeling, ethics guidelines, and human-AI interaction roadmaps,,, [110],,;;; [115];.Source: §3.2 · p. 18
• The review elaborates seven non-technical and three technical trustworthy-AI metrics, together with measurement models and trustworthiness frameworks.Source: §3.2 · p. 19

3.2.1. Trust & explainability / transparency / interpretability

• Opaque deep neural networks create black-box trust concerns, motivating Explainable AI to add explanations, transparency, and interpretation, which can increase end-user trust,,, [122], [123],,.Source: §3.2.1 · p. 19
• Transparency can avert overtrust, but manipulative justifications can promote it; excessive early emphasis may hinder innovation, and appropriate transparency varies across stakeholders and contexts,,,,,, [131],,,,.Stakeholder categories include developers, regulators, deployers, users, and society, while social geography and personality can further alter required transparency,,.Source: §3.2.1 · p. 19
• Explainability faces an accuracy-transparency tradeoff: highly accurate deep-learning models are often least explainable, whereas clearer models may perform worse; XAI seeks to mitigate this gap, [139],,.XAI has been applied across medical, robotic, transportation, and financial domains, while GDPR recognition of limits on solely automated decisions has further incentivized it,,,,,,,.Source: §3.2.1 · p. 20
• Interpretability makes input-output rationale understandable, whereas explainability also addresses internal machinery and training; methods span pre-model, intrinsic, and post-hoc stages, but consensus and quantification remain unresolved [150],, [152],,,,,,,,,,,,,, [166], [167],,.Interpretability levels should suit the entity interacting with the system, although a consensual definition and measurement approach has not yet been reached [167],,.Source: §3.2.1 · pp. 20–21

Trust & empathy in AI

• Empathy builds trust by combining empathic accuracy with supportive responses, enabling efficient communication, social bonding, and behavior that meets users’ expectations;.Empathy also influences credibility and trust through empathic accuracy.Source: §3.2.2 · p. 21
• Cultural-aware empathic agents should explain behavior and motivation, detect and interpret social signals, act respectfully, and adapt to individual differences.These requirements include sensitivity to values, norms, hearing levels, and dialects.Source: §3.2.2 · p. 21
• Empathy can indirectly support trust through accountability and user similarity, as people trust agents that consider stakeholder needs or share their values and preferences.The relationship between empathy and trust also requires reconsideration in AI-based treatments and patient online communities (,), and has been examined for online or AI-based services.Source: §3.2.2 · pp. 21–22

3.2.3. Trust and privacy

• Higher privacy concerns are associated with lower trust, while trust increases more accurate information sharing and decreases perceived risk;;;;,.Privacy includes control over identity, location, communication, access, and data-processing information, while self-efficacy in protecting data also affects trust,,.Source: §3.2.3 · p. 22
• Privacy concerns can make users reluctant to share detailed or accurate personal information and avoid online shopping, reducing the value companies gain from customer data,,.Customer privacy empowerment addresses this tension by giving users control over how, when, and how much personal information is used.Source: §3.2.3 · p. 22
• Transparency, explanations, privacy-by-design, and external legal guarantees can strengthen privacy protection and trust by improving control, accountability, and data confidentiality,,,.Privacy-by-design measures include data minimization, controllability, transparency, usable privacy functions, confidentiality, technical data quality, and limited data use; technical proposals include homomorphic encryption, blockchain, cooperative-information models, and anonymous authentication,,,,,.Source: §3.2.3 · pp. 22–23

3.2.4. Trust and fairness in AI

• Algorithmic discrimination and bias hinder trust, whereas perceived fairness enhances trust by assuring users that AI treats groups justly and without bias,,.Fairness concerns can arise from unrepresentative training data or unfair model design, such as insufficient data from Black-skinned and minority populations in medical diagnosis,.Source: §3.2.4 · p. 23
• Perceived fairness is positively related to induced fairness, and both are associated with greater user trust in AI.Human decisions are generally perceived as fairer and more trustworthy than algorithmic decisions for human tasks, except among minority and marginalized groups, who view algorithmic decisions as comparably trustworthy;,.Source: §3.2.4 · p. 23
• Perceptions of harm, injustice, and wrongdoing increase uncanniness, which negatively influences trust in AI agents, making rules and regulations necessary for fair trustworthy systems,.User biases also produce different trust levels for explainable AI diagnoses, including malignant and benign cases.Source: §3.2.4 · p. 23

Trust and accountability in AI

• Public trust requires improving individual models and recommendations while establishing minimum standards, verification, accountability, and transparent explanations.Source: §3.2.5 · p. 24
• Accountability research defines human stakeholders’ rights, AI developers’ obligations, and enforcement by governments, developers, and watchdogs.Source: §3.2.5 · p. 24
• The “right to explanation” treats accountability as answerability, requiring AI systems to justify their actions through suitable explanations.Source: §3.2.5 · p. 24

• Technical trustworthiness includes reliability, safety encompassing fairness and explainability, security, lineage, and accuracy, with accuracy influencing trust more strongly than explainability;;;; [235].Human-meaningful explanations can improve understanding and trust in system accuracy, but accuracy is reported as the more influential factor.Source: §3.2.6 · p. 25
• For high-stakes law, medical, and military decisions, interpretability, uncertainty awareness, comparative human-AI decision models, and adaptive calibration help address inappropriate trust, including over-trust;;.These approaches aim to identify when users should trust or distrust AI and calibrate reliance accordingly.Source: §3.2.6 · p. 25
• Trust can be strengthened through information-quality-aware decision models, supplier declarations, FactSheets, zero-touch security architectures, and diversity-based network robustness;;;;.These mechanisms promote transparency, accountability, secure communication, or resilience against failures and attacks.Source: §3.2.6 · p. 25
• Blockchain-based graph methods support trusted knowledge sharing by isolating malicious nodes and preventing knowledge pollution, while trust can improve AI design, development, and deployment despite cybersecurity limitations;.AI has also demonstrated increased efficiency in various tasks [e.g., Rahman et al., [245]], but cybersecurity does not guarantee trust.Source: §3.2.6 · pp. 25–26
• Mathematical models, weight of advice, and physiological signals examine trust robustness, topic-dependent reliance, cognitive load, and trust dynamics in human-AI and human-robot interactions;;;; [249];.Participants behaved similarly when receiving human or AI advice, but trust varied with the advice topic, while biosignals offered indicators of trust and cognitive load.Source: §3.2.6 · p. 26

Evaluation methods

• Trust assessment uses questionnaires, experimental protocols, qualitative evaluations, psycho-physiological approaches,, and empirical approaches,.These methods aim to assess human-AI trust and establish accuracy and safety guidelines for AI-assisted decision-making.Source: §3.2.7 · p. 26
• Questionnaires and protocols may omit trust factors, while diverse trust models, experimental constraints, and AI methods hinder comprehensive guidelines.The complexity of trust evaluation is further examined in.Source: §3.2.7 · p. 26

Trustworthiness frameworks

• Ontology-based trustworthiness models vulnerabilities, errors, and their relationships to establish confidence thresholds, while system evaluation incorporates vulnerability and risk assessments.Trust in manufacturing AI can also be categorized across organization, group, and individual levels.Source: §3.2.7 · p. 27

Trust assessment findings

• Trust differs between human-human and human-AI interaction in degraded scenarios, with human-human trust increasing while human-AI trust does not.Behavioral and physiological measures include individual and team performance, team situation awareness, and process measures.Source: §3.2.7 · p. 27

Trust metrics

• Recommended trust scales emphasize predictability, reliability, efficiency, and believability, while distinguishing trust in outputs from reliance on machine advice.Multi-dimensional trust scores can include satisfaction, job efficiency, understanding, control, and data protection.Source: §3.2.7 · p. 27

Trustworthiness certification

• AI certification can incorporate ethical principles, mandatory conformity assessment, algorithmic auditing, customization, education, and ISO-based quality and security management [266].Value-based trust may derive from algorithmic decision implications such as diagnosis accuracy and safeguarding [266].Source: §3.2.7 · pp. 27–28

Ethical foundations

• Ethical frameworks center trustworthy AI on accountability, privacy, ethics by design, and principles including transparency, fairness, non-maleficence, responsibility, autonomy, and explicability,,,, (Jobin et al., 2019),,.These approaches also examine algorithmic bias, value, fairness, and trust, including in global health.Source: §3.2.8 · p. 28

Explainability and social trust

• Explainable AI frameworks use guidelines, user feedback, and metrics such as fairness, accountability, transparency, and ethics to strengthen trust in air traffic management, public health, medicine, and data governance, (Shaban-Nejad et al., 2021a),.The stated aim is to bridge research transparency with practical explainability and obtain social license for data.Source: §3.2.8 · p. 28

Human-AI interaction

• Trust models for human-AI and conversational interactions examine decision similarity, behavioral patterns, contextual trust antecedents, cultural factors, user training, and reliance rather than trust,,,,.The frameworks span Naïve Bayes classification, conversational AI, industrial negotiations, technology acceptance, and robot autonomy.Source: §3.2.8 · pp. 28–29

Risk and consent

• Risk-management frameworks provide shared practices for managing AI risks across the organizational life cycle, while consent-based scenarios address trustworthy decisions in contact tracing, big-data analysis, and public-health emergencies,,,.These proposals respond to difficulties in explaining AI outcomes and deciding whether to accept them.Source: §3.2.8 · p. 29

Inspection and traceability

• Pattern identification, result inspection, and thorough examination are proposed to build trust when AI processes are not understandable or traceable, demonstrated with urban-logistics and heart-arrhythmia datasets.The framework links improved trust in human-AI interaction to making relevant patterns visible and inspectable.Source: §3.2.8 · p. 29

Regulatory and measurement frameworks

• European trustworthy-AI guidance organizes trust around seven factors—human oversight, robustness and safety, privacy, transparency, fairness, societal well-being, and accountability—and proposes frameworks to measure them,.Figure 3 presents different frameworks that can be employed in trustworthy AI.Source: §3.2.8 · pp. 29–30, Figure 3 · p. 30

3.3. Distrust in AI and Scary AI

• Distrust in AI stems from AGI outcompetition and alignment fears, surveillance, privacy, hackability, autonomous systems, high-stakes decisions, and moral uncertainty,,,,,,, [298],,,.Distrust increases with decision stakes, while healthcare AI can sometimes receive more trust than human doctors, with gendered ethical implications,,,,.Source: §3.3 · pp. 31–32

3.3.1. Distrust makers in AI systems

• Three major distrust-maker classes are surveillance and manipulation, human autonomy and dignity threats, and unpredictable futures.These classes organize the section’s account of why AI systems become distrusted.Source: §3.3.1 · p. 32, Figure 4 · p. 32

3.3.2. Surveillance, and manipulation

• Surveillance distrust extends to developers, governments, and institutions, while hacked, incomplete, biased, or inaccurate datasets can produce harmful systematic misdiagnoses,,,,,.Trust therefore depends on both the AI system and the surrounding scientific, corporate, and governmental communities.Source: §3.3.2 · pp. 32–33

3.3.3. Human autonomy/dignity threat

• Autonomy and dignity concerns center on inherited social bias and AI supplanting human agency, including recidivism predictions that can reproduce racial inequities [227],,,.Dignity concerns intensify when patients trust algorithms more than human doctors, especially when that imbalance disadvantages female doctors,.Source: §3.3.3 · pp. 33–34

3.3.4. Distrust and unpredictable futures

• Unpredictability spans AGI alignment and containment risks, exploitative behavior in novel situations, privacy-insensitive data use, and autonomous-weapon escalation beyond timely human intervention,,,,,,,,,.Proposed mitigation includes AI boxing for large-scale risks, while local unpredictability includes dirty tricks and drone swarming.Source: §3.3.4 · p. 34

3.3.5. Challenges and barriers to breaking distrust

• Breaking distrust requires calibrated transparency and trustworthy-AI assessment, but explanations work best when timely, visual, and local, while users may still over-trust systems,,,, [227],,,, [313],.The challenge is finding both effective explanation conditions and an optimal trust level, as illustrated by an autonomous-car crash after the driver stopped monitoring the road.Source: §3.3.5 · p. 35

3.4.1. Factors that affect trust

• Trust varies with technical, human, and contextual factors, and mismatches between perceived and actual capability can produce harmful overtrust or under-trust.Performance, reliability, transparency, AI personality, expertise, culture, task risk, and reputation influence trust; higher task risk lowers trust, while branding or marketing can create overtrust when capabilities are misaligned,,,.Source: §3.4.1 · p. 36

3.4.2. Methods of Building trust in AI

• Trust-building combines technical safeguards such as accurate, reliable, explainable, transparent, and bias-free systems with axiological measures including documentation, regulation, and accountability,.Local explanations are especially effective during errors or critical situations, while global explanations improve later understanding; confidence reports, interactive visualizations, human-presented explanations, and visual example-based rationales can also increase trust,,, [328].Source: §3.4.2 · pp. 37–38
• Transparent documentation and external standards can support trust, but inaccessible metrics, insufficient regulation, and ethics-washing can limit or undermine it.Supplier declarations of conformity mainly assist expert users, whereas expert or regulatory endorsement, certification, accreditation, and enforceable rules can help non-experts; pretending to follow ethical guidelines damages trust [329],,,, [333],.Source: §3.4.2 · pp. 38–39

3.4.3. Case studies and items effects on building trust

• Domain-specific implementations materially affect trust: precise marketing information, agent-presented explanations, blockchain privacy guarantees, value similarity, and model-assessment reports each address different trust requirements,,,,,.Precise rather than imprecise recommendations increase trust, especially with high product quality or low information accuracy; virtual-agent explanations increase trust in speech recognition, while Cortex Certifai evaluates robustness, fairness, and interpretability through query access.Source: §3.4.3 · pp. 40–41

4.1.Interaction of technical and non-technical factors of trust and trustworthiness in AI

• Trust drives AI adoption, but trustworthiness and trust differ: technical performance, reliability, and explainability interact with non-technical factors, while bias, discrimination, privacy, and domain-specific relationships remain unresolved.Understanding these relationships requires well-designed experiments and comprehensive quantitative and qualitative models of trust.Source: 4.1.Interaction of technical and… · pp. 41–42

4.2. Non-interchangeability of interpretability, explainability and transparency, and their classification

• Interpretability, explainability, and transparency are distinct concepts that should be classified by stakeholder information needs, because excessive transparency can confuse users, reduce trust, or enable system gaming.Interpretability links inputs to outputs, explainability addresses internal workings, and transparency makes explanations clear.Source: §4.2 · p. 42, §4.3 · p. 42

4.3. Trust as a two-way street

• Trust is reciprocal and fragile: users must trust AI while AI systems must be trusted by users, and destructive decisions by major companies motivate neutral oversight focused on safety and reliability.The discussion also distinguishes empathy-related trust from trust in human online exchanges and identifies privacy violations as a potential cost of empathy-based trust.Source: §4.3 · p. 43, §4.4 · p. 43, §4.5 · p. 43

4.6. The subjectivity of trust in AI vs. the objectivity of reliable AI

• Trust cannot be equated with reliability or guaranteed by a right to explanation: subjective trust may produce overtrust, objective reliability may not persuade users, and explanation can create incentives for simpler but suboptimal models.Transparency and explainability may bridge objective and perceived fairness, while privacy can require reciprocal disclosure of user and product information.Source: §4.6 · pp. 43–44, §4.7 · p. 44, §4.8 · p. 44

4.9. Development of direct AI accountability

• AI accountability remains underdeveloped because systems make decisions while developers bear responsibility; robust legal frameworks must align explanation and accountability and address whether AI can be directly accountable.Avoiding liability may distort development, and predictable accuracy alone does not establish trustworthiness without a reliable basis for behavior.Source: §4.9 · p. 45

4.10. Challenges of measuring trust and trustworthiness in AI

• Trust measurement must integrate culturally sensitive ethical principles with technical, interactional, performance, privacy, control, vulnerability, error, and risk metrics across individual, group, and organizational levels.The discussion proposes interconnected frameworks and a comprehensive reference of principles, protocols, roadmaps, and guidelines rather than a single universal framework.Source: §4.10 · pp. 45–46

4.11. Trust equity problem in AI

• Trust equity requires more than minimal adoption: unequal trust, distrust, or overreliance can intensify demographic and economic harms, motivating domain-specific interventions and an intersectional trust-equity framework.The proposed framework links conditions for trustworthiness and inappropriate trust through a feedback loop between equitable AI adoption and trust research.Source: §4.11 · pp. 46–47, Figure 6 · p. 48

Governance and standards

• Trustworthy AI requires pervasive governance that protects human care and satisfaction, standardizes laws and design, and continuously tracks algorithms, products, and codified principles.Trust cannot be achieved by considering only one or more individual metrics.Source: §5 · pp. 48–49

Trust metrics

• Trust metrics should adapt to technological and cultural dynamics, combining reliability, transparency, explainability, interpretability, accountability, fairness, security, privacy, provenance, automation, and vulnerability reduction.Explainability and interpretability are described as more important than accuracy, while flexibility can be informed by questionnaires, interviews, and surveys.Source: §5 · p. 49

Human and contextual factors

• Building trust requires evaluating AI-related, human-related, and context-related factors, while documentation helps users perceive system trustworthiness and axiological factors may increase trust without improving technical trustworthiness.Transparency, explainability, and performance are identified as especially important technical factors across application domains.Source: §5 · pp. 49–50

AI–AI interaction

• Future research should calibrate trust in AI–AI interaction, emphasizing security and reliability, robustness against adversarial attacks, malicious-resource detection, and dynamic calibration across multiple sources.Unlike human–AI interaction, transparency and explainability are described as having no impact on trust in AI–AI interaction.Source: §5 · p. 50

Research gaps

• Figure 7 identifies uneven attention across four major classes of trust-related AI research, leaving neglected areas as fertile directions and some questions open for the long term.The review bases this conclusion on the number of reviewed papers and quantitative analysis.Source: §5 · p. 50, Figure 7 · p. 50