From Explainable to Interactive AI: A Literature Review on Current Trends in Human-AI Interaction

Motivation and gap

• The review finds that most human-AI research remains centered on explanations and occasional contestation, providing users limited agency to control or adapt AI systems.The authors argue for balanced interaction that separates user and AI tasks and establishes shared responsibility while managing trade-offs involving performance, complexity, and user fallibility.Source: §1 · pp. 2–3

Review contribution

• The review systematically consolidates and critically analyzes Human-Centered AI, Explainable AI, Interactive AI, and related fields addressing explicit, intentional, and informed user interaction.Its analysis evaluates literature across AI users, implementations, and goals, while examining implications for trust, acceptance, and utilization.Source: §1 · pp. 3–4

Key findings

• Many studies exclude end-users from AI co-design or even basic interaction, while active-interaction research concentrates on low-risk domains and trivial tasks rather than healthcare and security.The review also finds that most systems prioritize user experience and acceptance over broader forms of agency.Source: §1 · p. 3
• Only a handful of studies, particularly in Interactive Machine Learning, allow users to modify AI mechanics, indicating an opportunity to extend agency and control beyond system understandability.Across users, implementations, and goals, a significant share of the literature proposes theoretical rather than practical solutions, motivating more experimentation in Interactive AI.Source: §1 · p. 3

Recommendations

• The review synthesizes challenges and proposes future-design guidelines that balance system autonomy, user agency, and user needs through participatory interaction design and reflective stakeholder evaluation.The guidelines address user experience, transparency, interaction, and augmentation.Source: §1 · p. 4

2. Background

• AI’s emphasis on predictive accuracy and growing model complexity can overlook human needs while making decisions harder to understand, trace, and contest in high-stakes domains.Complex systems such as deep neural networks obscure internal decision processes, creating ethical and adoption challenges in medicine, law, finance, and security.Source: §2 · pp. 5–6

2.1. Explainable AI (XAI)

• XAI improves understanding through transparent models or post-hoc explanations, but simpler models may lose performance while opaque-model explanations remain difficult for end-users to interpret [42].Post-hoc methods can trace outputs through techniques such as input perturbation, yet these traces may remain too complicated for concrete user understanding.Source: §2.1 · p. 6
• XAI research still insufficiently examines users’ ability to interact with explanations across stakeholder contexts, offering end-users limited opportunities to adapt systems or integrate them into established processes [48].Different stakeholders have distinct goals and explanation needs, so explanation mechanisms should adapt to usage contexts and involve users more actively in design and evaluation.Source: §2.1 · pp. 6–7

2.2. Human-AI Beyond Explainability

• Human-AI interaction extends from humans-in-the-loop to end-users as co-creators of Interactive AI, but collaboration is often constrained by expertise, fallibility, socio-technical factors, and users’ limited roles as data providers or feedback agents.Improper integration can undermine interaction, although human involvement can improve decision-making, trust, and acceptance.Source: §2.2 · pp. 7–8
• Beyond explainability, the field calls for human-AI interaction that gives users agency to contest decisions, adapt and control systems, and participate actively in their co-creation.A systematic literature evaluation is needed to assess whether current HCI practices support interactive systems and user-centered adaptation or control.Source: §2.2 · pp. 8–9

3. Methodology

• The review applies a method-based developmental evaluation and PRISMA protocol to systematically collect and assess primary studies.Keywords were identified through initial evaluation, followed by database searches and snowballing.Source: §3 · p. 9

3.1. Inclusion and Exclusion Criteria

• Studies were included when human-AI interaction was explicit, intentional, and informed, requiring visible system impact, user agency and control, and awareness of the interaction.Passive intent inference and explanation-only systems were excluded because they lack sufficient user interaction.Source: §3.1 · pp. 9–10
• Search strings were developed through explanatory searches across major scholarly libraries and covered explainable, contestable, collaborative, hybrid, interactive, and human-centered AI terms.The strategy also included Interactive ML and established abbreviations such as XAI, AI, and ML.Source: §3.1 · p. 10

3.2. Search and Data Collection

• Google Scholar and Scopus supported diverse searches, while the review prioritized post-2003 articles, conference papers, reviews, and short surveys and excluded non-research publication types.Results were gathered from ACM, Springer, ScienceDirect, Taylor and Francis, and other sources to reduce venue bias.Source: §3.2 · p. 11

3.3. Study Eligibility Evaluation

• Three authors screened titles, venues, abstracts, and full texts, with conflict resolution producing excellent categorization reliability (intraclass correlation coefficient 0.94).Figure 1 presents the PRISMA-adapted review workflow.Source: §3.3 · pp. 11–12, Figure 1 · p. 12
• From 268 initially identified studies, 39 primary studies met the interaction criteria, and one-round forward and backward snowballing added 15, yielding 54 studies overall.The eligibility process used five relevance levels, with Level 4 requiring active and explicit user interaction beyond data provision.Source: §3.3 · pp. 11–13, Table 1 · p. 13

3.4. Definitions and Terminology

• Human-AI is framed as an integrated perspective on technology, people, and policies that prioritizes human factors, agency, augmentation, and collaboration over technology-centered design.The review situates these concepts at the intersection of HCI and AI.Source: §3.4 · pp. 14–15

3.4.1. Interactivity and Agency

• Interactivity ranges from passive explanations to explicit input, feedback, contestation, and correction, with agency depending on how users interact, why they interact, and the methods provided.Explicit interaction can occur through interfaces or controls that directly accept user input or responses.Source: §3.4.1 · p. 15

4. Analysis of Dimensions

• Selective coding grouped study attributes by users, data modality, implementation, application area, and study goals, while cohesively examining interactivity, human involvement, adaptation, and agency.The identified dimensions and their sub-categorizations are summarized in table 3 and Figure 3.Source: §4 · p. 16
• The dimensions were evaluated against stated objectives and contributions related to interactive AI, with analysis limited by the information available in published content.Source: §4 · p. 16

4.1. AI Users

• User participation varies with recruitment methods, task type, and expertise, making clear target-user profiling essential before developing human-AI systems.Users are generally categorized as practitioners, domain experts, or novices according to their AI and domain knowledge.Source: §4.1 · pp. 16, 19
• 29 (53%) studies provide generic interaction adaptations, while 14 target domain experts and only 6 target AI/ML experts.Figure 4 shows that many studies do not specify target users, although interactive studies often focus on expert users.Source: Figure 4 · p. 20, §4.1 · p. 20
• User testing with actual target users remains uncommon, often relying on AI experts whose behavior may not represent end-users [120].This substitution may suit XAI systems intended for AI experts but is inadequate for most stakeholders.Source: §4.1 · p. 20
• Domain experts and novices may lack AI expertise and need interpretation or education, but studies largely assist tasks interactively without explaining how AI works.Mishra and Rzeszotarski teach non-experts to reuse AI for their work.Source: §4.1 · p. 21
• User exclusion reduces human control and agency, while increasing system complexity or reducing AI expertise lowers interactivity; restoring control supports fairness, acceptance, and learning.The survey concludes that generic inclusiveness efforts remain insufficiently targeted, and the overall trend resists opening AI to end-users.Source: §4.1 · p. 21

4.2.1. Implementation of the Solution

• Only 23 studies (43%) present systems or models for direct user interaction, while many collaborative and human-centered proposals remain theoretical or lack concrete implementation.Figure 5 shows that interactive implementations are more practically developed than collaborative, human-centered, and hybrid proposals.Source: §4.2.1 · p. 22, Figure 5 · p. 23
• Practical systems include Teachable Machine and What-If, which empower users to interact with AI, test solutions, and subsequently use them.Other implementations include collaborative systems, virtual-reality design tools, and explanatory interactions, although some studies rely on Wizard-of-Oz agents.Source: §4.2.1 · p. 23

4.2.2. Application Domain

• Only 4 of 54 studies address high-impact domains, while explicitly specified applications concentrate in education/training (10 of 30) and leisure (7 of 30).This distribution suggests that current interactivity is more tolerated in low-stakes settings than in consequential tasks such as medicine or fairness.Source: §4.2.2 · p. 24, Figure 6 · p. 24
• High-stakes interactive AI remains under-explored, despite examples supporting medical diagnosis, pathology, brain-injury investigation, and fairer collaborative recidivism-risk assessment.Groups relying on AI were more confident overturning incorrect recommendations, making fairer decisions, and crediting correct recommendations, although one study did not simulate real-world conditions.Source: §4.2.2 · pp. 24–25

4.2.3. Modalities of Data

• Text is the most common modality, appearing in around 33% of studies, while image and audio/video data primarily support recognition and classification tasks.Textual interaction is especially prevalent in conversational systems, including educational chatbots and voice-based interactive robots.Source: §4.2.3 · p. 25, Figure 7 · p. 26
• Studies often omit explicit data descriptions and rarely explore multimodal interaction beyond text and images, limiting practical experimentation in high-risk user contexts.Interactive machine learning provides more specific data details, while conversational, immersive, and parameter-based approaches support user involvement in learning and AI performance.Source: §4.2.3 · pp. 26–27

4.3. AI Goals

• AI-goal definitions vary from improving user experience to enabling active system modification, with study goals and recruitment shaped by audience, task complexity, and domain expertise.Not all reviewed studies directly include users in testing, creating a substantial barrier to making interactivity prevalent.Source: §4.3 · p. 27

4.3.1. User Experience

• 39 studies (72%) advocate improving user experience, treating it as central to acceptability, collaboration, accessibility, understanding, and trust.User experience goals include evaluating behavior, following usability guidelines, guiding users, and providing explanations or feedback during interaction.Source: §4.3.1 · pp. 28–29, Figure 8 · p. 28

4.3.2. Transparency and XAI

• Only around 11% of studies appear to use transparent AI, while most do not specify transparency or rely on opaque, black-box implementations despite explainability efforts.Transparency supports trust, acceptance, satisfaction, and understanding, but its appropriate level varies with task complexity and stakeholder needs.Source: §4.3.2 · pp. 29–31, Figure 9 · p. 30

4.3.3. Interactivity

• Interactive AI commonly supports task completion and automation, but only a limited set lets users adapt underlying systems through feedback, parameter changes, corrections, or direct model alterations.Interactive machine-learning examples include users leading systems, correcting models, and adapting algorithms to preferred solutions, although the studied domains remain limited.Source: §4.3.3 · pp. 31–33

4.3.4. Algorithm Modification

• Only 5% of studies allow active user modification at the model or algorithm level, with barriers including limited expertise, implementation hurdles, task stakes, and potential bias.Most systems restrict interactivity to interaction or feedback rather than participation beyond feedback.Source: §4.3.4 · p. 33, Figure 11 · p. 33

5. Discussion and Directions

• The review maps current human-AI research, its empirical measures, and its movement beyond explainability toward approachable, interactive, user-centered systems.It highlights improved user experience, support for non-experts, reduced barriers, and research questions for practical implementation.Source: §5 · p. 35

5.1. Purpose of Interactive AI

• Interactive AI builds systems around user needs, balancing stakeholder demands for human augmentation and business demands for automation that minimizes interruption.This balance creates a paradox between collaboration with AI and completing tasks through automation.Source: §5.1 · p. 35
• User-centered interactivity gives people control to participate in and contest decisions, while enabling AI to capture real-time feedback and adapt.The central question is who remains the ultimate decision-maker as AI advances in human activities.Source: §5.1 · p. 35

5.1.1. Collaboration

• Collaboration assigns tasks according to human and AI strengths, improving both system and user performance rather than replacing human work.Cooperative roles are especially important because adding AI may worsen complexities such as algorithmic bias.Source: §5.1.1 · pp. 35–36
• Interaction helps AI gain knowledge, receive performance feedback, correct decisions, and build user acceptability, trust, and adoption.Collaboration should prioritize human augmentation, including creativity, while recognizing that excessive interactivity can inhibit acceptance.Source: §5.1.1 · p. 36

5.1.2. Augmentation

• Augmentation should alleviate human task difficulties through a practical human-AI loop, balancing human and AI needs while minimizing unwanted feedback and input errors.Users require fine-grained specification tools, control over preferences, and the ability to initiate critical actions without unwanted interruptions.Source: §5.1.2 · pp. 36–37
• Augmentation requires carefully balancing human intervention: systems may restrict intervention to low-confidence decisions, yet still depend on feedback to improve efficacy.The discussion connects this feedback loop to unresolved questions about continuously evolving human knowledge.Source: §5.1.2 · p. 37

5.2. Interactive AI in Practice

• Users mostly evaluate AI systems curated by experts, while their feedback is stored for later improvement rather than incorporated into the system; interaction design should close this gap.Explanations are treated as only a starting point, and users generally have little control over AI behavior.Source: §5.2 · pp. 37–38

5.2.1. Human-AI Guidelines

• Effective interaction requires identifying user needs, knowledge, and task relevance; domain experts appear more beneficial for system improvement than AI experts or novices, though task-focused study prevalence may bias this finding.Guidelines for user-centered AI are advocated, while many studies emphasize data over user expertise. Not every user needs interaction or is suitable for participation because users add differing complexity.Source: §5.2.1 · p. 38

5.2.2. Technical Challenges

• Interactive AI faces technical and multidisciplinary challenges, with high-stakes tasks tending toward lower interactivity and current applications concentrated in assistive tools and language-model assistants.Many proposed solutions remain theoretical, while deployed systems must support interaction at scale and rapid model adaptation despite complex training and optimization cycles.Source: §5.2.2 · pp. 38–39

AI and User Agency

• Interactive AI can give users agency to modify system mechanics, but complex autonomous systems make such agency difficult and remain prevalent across high-stakes domains.The section contrasts growing interaction with traditional AI’s isolated, autonomous control and emphasizes the need for agency in health, legal, and privacy contexts.Source: §5.3 · p. 39
• Future interactive AI should increase user agency while optimizing the balance between system autonomy and user control.Greater agency is presented as important for expert evaluation and the development of more interactive systems.Source: §5.3 · pp. 39–40
• Active algorithm alterations can improve acceptance of imperfect AI, although research must determine whether output correction, parameter updates, or feedback provide sufficient user control [135, 18, 136].Current deployments commonly support post-deployment interaction and feedback, while access to underlying models remains uncommon.Source: §5.3 · p. 40

• AI interfaces span conversational text, graphical manipulation and creation, sensory robotics, and AR/VR interactions, including combinations of modalities for richer user experiences.These interfaces differ substantially from traditional web and mobile applications, making practical user research increasingly important.Source: §5.4 · pp. 40–41
• Interfaces are organized into instructing, conversing, exploring, and manipulating, with conversing dominant across intelligent systems and generative AI such as ChatGPT.Conversational systems can support goal- or task-oriented interaction and may also impersonate experts.Source: §5.4 · p. 41
• Instructing interfaces such as Teachable Machines are comparatively usable, while exploring and manipulating remain less explored but let users inspect predictions, alter parameters, annotate data, and interact with models.Human-AI interfaces can also help users interpret explanations and build causal inferences.Source: §5.4 · p. 41
• Interactive interfaces provide feedback about users’ understanding of AI outputs, supporting continued interaction, system trust, and acceptance, although their capabilities depend heavily on practical implementation.Theoretical contributions remain important as foundations for experimentation, but without implementation interface capabilities cannot be evaluated in practice.Source: §5.4 · pp. 41–42

5.5. Addressing HCAI challenges: current trends and gaps

• HCAI research emphasizes transparency and explainability while increasingly pursuing rapid feedback, AI adaptation, error reduction, and human-led collaboration.The intended direction places users at the center of the human-AI loop, with AI complementing human skills and expertise.Source: §5.5 · p. 42
• User-centered design should incorporate human factors from the outset through feedback, adaptation, personalization, and customization that improve acceptance, trust, and perceived usability.Xu et al. call for HCAI research agendas to make user-centered AI design more prevalent and ensure AI enhances human capabilities.Source: §5.5 · p. 42
• AI interaction design must address diverse non-expert users through accessible interfaces and natural-language systems that account for socio-technical needs.Traditional systems have often focused on specialized user groups despite AI’s expanding role in everyday life.Source: §5.5 · p. 43
• Interactivity spans understanding, contestation, feedback, modification, task support, augmentation, and collaborative decision-making, but excessive interaction can reduce trust.Underlying-model modification remains largely limited to educational or experimental low-stakes applications, motivating future systems with greater user agency, explainable interaction, and alignment with user needs and values.Source: §5.5 · p. 43

Limitations

• The review includes theoretical interactivity studies without participant testing, so its findings require caution when compared with user-centered studies.Its scope may also omit relevant work because restrictive queries, the field’s rapid change, and coverage only through 2023 constrain the synthesis, despite extensive snowballing.Source: §5.6 · pp. 43–44

Future Directions

• Future research should examine how balanced user control and AI autonomy affect interactions, including collaborative and hybrid intelligence approaches.Interactive machine learning operationalizes intentional, informed, and explicit user interaction, while hybrid intelligence combines human and AI capabilities toward shared goals.Source: §5.6 · p. 44

Transparency and Explainability

• Practical transparency remains difficult because explainability needs vary by user role and lack defined benchmarks for evaluating AI outputs.These gaps are especially consequential in healthcare, justice, and security, where systems must support accountability, consequence anticipation, and tracing decisions.Source: §5.6 · pp. 44–45

Future Directions

• Future work should investigate interactivity in Generative AI and determine which interaction types are effective across different user contexts.Increasing model complexity makes interpreting generated content more demanding, while emerging creative uses create new possibilities for interaction.Source: §5.6 · p. 45

Conclusion

• The review systematically examines Interactive AI across users, implementations, and goals, identifying patterns, gaps, and opportunities beyond Explainable AI.Interactive AI is defined as explicit, intentional, and informed interaction with AI systems.Source: §6 · p. 45
• The primary contribution is advocating active user participation in the AI loop through feedback and co-creation, including adapting and co-designing AI mechanics.This prioritizes the user’s role beyond receiving explanations.Source: §6 · p. 46
• Current research mainly improves user experience and accessibility, while practical user-centered interaction, control, and adaptation remain under-explored.Emerging studies let AI adapt to user feedback, but concrete solutions remain scarce; work also includes expert interactive machine learning and task-focused assistants.Source: §6 · p. 46
• Only a few interactive studies let users modify algorithmic design parameters, leaving meaningful agency over AI decision-making largely unaddressed.Most systems provide feedback in ways that increase awareness of AI mechanisms without enabling complete user control.Source: §6 · p. 46