{"id":"68875e44-316f-46cb-8cbd-bc9f8295002c","arxiv_id":"2507.21158","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conceptual framework uses implicit physiological signals to adapt AI explanations and support swift trust in high-stakes human-AI teams.","lead":"This paper proposes AXTF, a conceptual framework for adaptive explainability in human-AI teams that uses physiological signals such as EEG, ECG, and eye tracking to estimate trust and adjust explanations. The target setting is high-stakes, time-pressured environments like emergency response, where explicit feedback is impractical.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Internal contradiction in the fuzzy trust rule base (Table 3): Rule 6 forces Low trust whenever workload is High, conflicting with Rule 2's High output; the framework's central inference engine is underspecified.","rationale":"The reader identified the unvalidated sensor-to-state-to-trust mapping as the weakest assumption. That is a valid empirical concern. However, my stress-test pass finds a more fundamental and more easily checkable issue: the fuzzy rule base that is the paper's formal core is internally contradictory. This matters because the paper's stated contribution is a 'multi-objective, personalized trust estimation model' whose outputs drive all adaptive explainability. If the inference rules can fire conflicting trust levels for the same input state, the model does not yet constitute a coherent estimator, regardless of sensor accuracy. This is not a disagreement with the research direction; the AXTF concept is reasonable and the paper is honest that implementation and evaluation are future work. The contradiction can likely be fixed by adding rule priorities, fuzzy aggregation with defuzzification, or by revising Rule 6, and the paper should be accepted with that condition. Since the reader already gave CONDITIONAL, my analysis does not change the verdict, but it sharpens the condition: the framework needs a consistent inference engine and an explicit conflict-resolution mechanism before it can support the claimed closed-loop trust adaptation.","tokens_in":13954,"tokens_out":2730,"duration_ms":29930,"concrete_test":"Implement the Table 3 rule base exactly as written using the membership functions in Section 3.2 and enumerate all combinations of W in {Low, Medium, High}, S in {Low, Medium, High}, E in {Negative, Neutral, Positive}, and P in {Low, Medium, High}. Record each input where two or more fired rules assign different T values; the state (W=High, S=Low, E=Positive, P=High) should fire Rule 2 (T=High) and Rule 6 (T=Low). If such conflicts are found, specify and implement an explicit conflict-resolution and defuzzification scheme, then verify that the resulting trust estimates are unique and continuous. Without such a scheme, the central trust model is not well-defined even before empirical validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AXTF produces dynamic trust estimates guiding explanation adaptation depends on the trust inference engine being coherent. In Section 3.2, Table 3, Rule 6 states 'IF W = High OR E = Negative THEN T = Low', while Rule 2 states 'IF S = Low AND E = Positive AND P > 0.8 THEN T = High'. For any state with W=High, S=Low, E=Positive, P>0.8, Rule 6 forces T=Low and Rule 2 forces T=High simultaneously. The paper provides no conflict resolution, defuzzification, or rule weighting scheme. Thus the multi-objective trust model can output contradictory trust levels for a single input state, leaving downstream explanation adaptation undefined at those points. This is an internal consistency problem, not merely missing empirical validation: even perfect sensor-to-state classification does not resolve the contradiction. Additional underspecification: Rule 4 uses a crisp threshold P>0.6 while other rules refer to fuzzy performance membership, and Rules 1 and 6 overlap; the aggregation of multiple fired rules is never described.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AXTF, a conceptual framework for adaptive explainable AI in high-stakes human-AI teams. It proposes a closed-loop architecture in which physiological and behavioral signals (EEG, ECG, GSR, eye tracking) are classified into workload, stress, and emotional valence; a fuzzy rule-based engine (Table 3) then estimates trust (Low/Medium/High) from these states and a system performance score; and the trust estimate modulates explanation features such as timing, duration, granularity, content, transparency, adaptability, and mode of delivery. The stated goal is to promote swift trust, reduce cognitive overload, and improve decision-making. The manuscript is explicitly a conceptual proposal: it contains no implementation, simulation, or empirical study, and Section 4 acknowledges that validation in interactive environments remains future work.","tokens_in":14184,"tokens_out":3915,"duration_ms":43234,"significance":"The paper addresses a genuine and timely problem: static, one-size-fits-all XAI and the reliance on explicit human feedback are ill-suited to high-pressure, time-critical human-AI collaboration. The idea of using implicit physiological signals to drive trust-calibrated explanation adaptation is well motivated and the related-work synthesis is useful. If the framework were formally specified and empirically validated, it could provide a valuable foundation for affective, situated XAI. However, the contribution as it stands is a conceptual sketch rather than a working model: the central trust inference engine has an internal inconsistency, the causal claims in the abstract and Section 3 are not supported by any evidence in the paper, and key inference-layer assumptions are unexamined. These issues are fixable within the manuscript's scope, but they currently prevent the paper from fully supporting its stated contributions.","major_comments":[{"comment":"The fuzzy rule base is internally inconsistent. Rule 2 states that IF S = Low AND E = Positive AND P > 0.8 THEN T = High; Rule 6 states that IF W = High OR E = Negative THEN T = Low. For an input state with W = High, S = Low, E = Positive, and P > 0.8, both rules fire and prescribe contradictory trust outputs (High and Low). The manuscript does not specify any conflict-resolution scheme, rule weighting, aggregation operator, or defuzzification method, so the trust estimate is undefined for such states. Because this rule base is the core inference engine of the framework, the model is not well specified. The authors should either revise the rules to eliminate overlap, add a priority ordering, or specify a concrete fuzzy inference procedure (e.g., Mamdani min-max inference with centroid defuzzification) that resolves conflicting rule outputs.","section":"Section 3.2, Table 3"},{"comment":"The fuzzy rules mix crisp thresholds on the performance variable with fuzzy membership functions. The text defines fuzzy sets Low, Medium, and High for system performance P with the displayed membership functions, but Rules 2, 4, and 5 use numeric thresholds (P > 0.8, P > 0.6, P < 0.4) rather than linguistic terms such as 'P = High' or 'P = Low'. This is formally inconsistent: either P should be fuzzified and the rules should refer to the fuzzy sets, or the threshold-based rules should be stated in crisp terms. The current presentation leaves the semantics of the fuzzy inference engine ambiguous and should be clarified.","section":"Section 3.2, membership functions and Table 3"},{"comment":"The paper makes causal claims about the framework's effects, stating that it 'reduces cognitive overload, enhances trust, and improves decision-making' and that it 'promotes swift trust.' No implementation, simulation, or empirical evaluation is provided, and Section 4 itself states that 'implementation and evaluation ... will be essential to validate the framework.' These claims should be reframed as design goals or hypotheses, with an explicit statement that they are not yet demonstrated. Otherwise the abstract and Section 3 overstate the evidence in a way that could mislead readers about the maturity of the work.","section":"Abstract and Section 3"},{"comment":"The framework assumes that physiological and behavioral signals can be non-intrusively classified in real time into workload, stress, and emotional valence with sufficient accuracy to drive trust estimation. The manuscript does not discuss the expected accuracy, latency, or robustness of such classifiers in high-stakes, dynamic environments, nor does it consider how classification errors would propagate into the trust estimates and explanation adaptations. Since this inference layer is the input to the entire framework, the paper should include a more critical assessment of the evidence for this premise and the conditions under which the framework would fail.","section":"Section 3.1"}],"minor_comments":[{"comment":"The footnote contains a malformed email address and an unusual symbol ('envel⌢pe-⌢pennlfernando11@gmail.com'); this should be corrected to a clean institutional or personal email.","section":"Title page footnote"},{"comment":"Several references are incomplete or have formatting issues, for example reference [13] has '????' in place of the publication year, and reference [20] has a duplicated 'doi: 10.1080/15472450.2022.2140046' string. The reference list should be cleaned up.","section":"References"},{"comment":"Figure 1 is referenced in the text but is not visible in the provided manuscript; the authors should ensure that the figure is included and that its callouts (e.g., 'closes the loop') are legible and consistent with the text.","section":"Figure 1"},{"comment":"There are numerous typographical errors and missing spaces, for example 'responsive teammate' should be 'a responsive teammate' and several sentences have missing words or broken spacing. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a workshop-style conceptual contribution, and its core idea is worth discussing. The main blockers are the internal inconsistency in the fuzzy rule base (Table 3) and the unsupported causal claims in the abstract and Section 3. If the authors specify the inference semantics, fix the rule conflicts, and temper the claims to match the conceptual nature of the work, the paper could be acceptable. The heavy reliance on self-citations is not problematic per se, but the reference list needs professional cleanup."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the Fernando et al. AXTF paper. My take: it is a genuinely useful conceptual blueprint, but the abstract overstates what is delivered, and the fuzzy rule base has a fixable internal contradiction that the authors should address.\n\nThe gap is real: static, uniform explanations with explicit feedback do not suit high-stakes, time-pressured human-AI teams. The paper's contribution is a specific closed-loop design—implicit physiological signals to a fuzzy trust inference engine to modulation of explanation timing, granularity, content, mode, etc.—with concrete rule tables. That integration is not present in the cited literature, and the tables make the proposal testable. The writing is clear and the related work is well anchored in the trust and XAI literature.\n\nNow the soft spots. First, the abstract and Section 3 say the framework 'promotes swift trust,' 'reduces cognitive overload,' and 'improves decision-making.' There is no implementation or evaluation; the paper itself lists those as future work. As a design proposal, that is fine, but the claims should be 'designed to support' rather than 'promotes.' Second, the trust inference engine has an internal conflict. Rule 2 outputs High trust when S=Low, E=Positive, and P>0.8. Rule 6 outputs Low trust whenever W=High. So with high workload, low stress, positive emotion, and high performance, both rules fire with conflicting outputs. No aggregation or conflict resolution is specified. Rule 4 also uses a crisp P>0.6 while others use fuzzy membership. This matters because the trust estimate drives the entire adaptation loop. The rules can be refined, but as written the engine is underspecified. Third, the physiological-to-state classification premise is unvalidated, but the authors acknowledge it as future work, so I weigh that as a missing validation rather than a hidden flaw.\n\nBottom line: this is a worthwhile workshop paper that deserves serious review as a position paper, conditional on toning down the empirical claims and resolving the rule conflict. I would send it out rather than desk reject, and I'd mention the contradiction to the authors.","headline":"Useful conceptual blueprint for adaptive XAI, weakened by overstated claims and a fixable inconsistency in the fuzzy rule base.","tokens_in":14766,"tokens_out":3036,"would_cite":true,"duration_ms":32245,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a closed-loop framework that reads operator stress, workload, and emotion from physiological signals and uses the inferred trust state to modulate AI explanations in high-stakes teaming.","keywords":["Adaptive Explainability","Swift Trust","Human-AI Teaming","Implicit Feedback","Physiological Sensing","Fuzzy Trust Inference","Multimodal Feedback","High-Stakes Decision Making"],"falsifier":"Run a simulated emergency-response task with operators wearing EEG, ECG, and eye-tracking sensors, and compare AXTF-driven adaptive explanations against static explanations on measures of self-reported trust, cognitive load, and task performance; the framework is falsified if the adaptive condition does not improve trust or performance. A more targeted test collects concurrent physiological features, performance scores, and self-reported trust from operators during a high-pressure task and checks whether the Table 3 rule 'IF W=High AND S=High AND E=Negative THEN T=Low' systematically holds.","tokens_in":13684,"feed_emoji":"🧠","tokens_out":6616,"duration_ms":60968,"temperature":0.7,"pith_summary":"This paper argues that explainable AI in high-stakes settings, such as emergency response, should not hand every operator the same explanation; it should watch the operator and adapt. The proposed adaptive explainability trust framework (AXTF) reads physiological and behavioral signals—EEG, ECG/HRV, eye tracking, facial expressions—to infer workload, stress, and emotional valence in real time, then uses a fuzzy rule-based trust model to adjust explanation timing, duration, granularity, content, transparency, and delivery mode. The aim is to build and calibrate swift trust without demanding explicit feedback from an overloaded human. If the framework works as described, AI teammates could sense confusion or overload and proactively simplify, reassure, or detail their reasoning in the middle of a crisis. The paper is conceptual: it specifies the loop and its rules, but does not yet run an experiment.","feed_headline":"Biosignals would tailor AI explanations to operator stress","feed_subtitle":"A proposed loop reads workload, stress, and emotion, then adjusts what, when, and how an AI explains to build swift trust.","key_machinery":"The central object is the Adaptive Explainability Trust Framework (AXTF), a closed-loop pipeline with three coupled components: multimodal sensing (EEG, ECG/GSR/HRV, gaze, facial and vocal cues), a multi-objective neurofuzzy trust inference engine, and an explanation feature modulator. The trust engine maps four inputs—workload $W$, stress $S$, emotion valence $E$, and performance $P$—through triangular membership functions and a fixed rule table (Table 3) to a categorical trust estimate $T\\in\\{\\text{Low},\\text{Medium},\\text{High}\\}$. That estimate then selects among seven explanation features (timing, duration, granularity, content, transparency, adaptability, delivery mode), e.g., short proactive audio alerts for high stress and low trust, detailed interactive visualizations when trust is high and load is low.","core_discovery":"The paper's central claim is that explainability can be turned into a closed-loop, trust-calibrating mechanism: physical signals reveal the operator's cognitive and affective state, a multi-objective trust model converts those states into a dynamic trust estimate (Low, Medium, or High), and that estimate modulates which explanation the AI gives next. The paper encodes this in a fuzzy inference system whose rules, e.g., IF workload is High AND stress is High AND emotion is Negative THEN trust is Low, are grounded in literature on trust in automation. The framework reframes explanation not as a static artifact but as an adaptive communication strategy that balances transparency, cognitive efficiency, and trust calibration under time pressure.","pith_inferences":["A natural test bed is a simulated search-and-rescue drone task where operators wear EEG/ECG/eye trackers and explanations are toggled between AXTF-adaptive and static; the framework predicts lower self-reported cognitive load and higher behavioral trust in the adaptive condition.","The fuzzy rule table in Table 3 is falsifiable in the small: collecting ground-truth ratings of workload, stress, valence, performance, and trust during a simulated crisis would confirm or refute the seven rules, and could show that trust depends on factors beyond these four inputs.","Because the loop closes through the human, one subtle risk is a self-reinforcing cycle: an adaptation that reduces stress changes the physiology the system reads, so the trust estimate and subsequent explanation may chase the operator's state rather than track system performance; the paper does not analyze this feedback dynamics, but it is directly testable.","The framework could be extended to multi-human teams where the trust estimate is aggregated across operators, or to autonomous vehicles and medical triage, which face similar time-critical explainability demands."],"forward_implications":["In high-stakes human-AI teams, explanations should be delivered proactively or reactively based on the inferred state of the operator, not as a fixed text block.","Trust can be treated as a continuously estimated variable that drives explanation adaptation, enabling trust repair rather than just trust measurement.","Non-intrusive physiological sensing can replace explicit user feedback, making adaptive XAI feasible when the operator's hands, eyes, and voice are occupied.","Explanation granularity, duration, and modality should be tuned to the operator's current cognitive load, e.g., 2–3 second audio confirmations under high load, longer layered explanations when load is low.","If realized, the framework would support both trust calibration and situation awareness in time-sensitive domains like emergency response and mission-critical decision support."],"supporting_citations":[{"why":"Defines swift trust in temporary groups, the trust construct the framework is designed to foster.","marker":"[4]"},{"why":"Meta-analysis of trust factors in human-robot interaction; supplies the empirical link between workload, performance, and trust that the fuzzy rules encode.","marker":"[5]"},{"why":"Demonstrates that trust can be classified from EEG and GSR signals, grounding the physiological sensing component.","marker":"[8]"},{"why":"Shows explanation granularity and cognitive load interact, motivating granularity adaptation in the framework.","marker":"[15]"},{"why":"Argues transparency, explainability, and situation awareness jointly shape human-AI team performance; the basis for the explanation-modulation rules.","marker":"[16]"},{"why":"Provides the three-level model of human trust in automation that informs the trust taxonomy and the dispositional baseline.","marker":"[25]"},{"why":"Supplies the model-reconciliation technique that adaptive explanation adaptation builds on.","marker":"[28]"},{"why":"Relates estimated user trust to transparency behavior, a direct precursor of the trust-sensitive adaptation loop.","marker":"[29]"}],"fun_headline_variants":["Biosignals tune AI explanations to operator trust","AI adapts explanations from your stress and workload","Closed-loop XAI reads your stress and changes its advice","Adaptive AI uses heart and eye signals to build trust","Trust-calibrated AI: from biosignals to tailored explanations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole loop depends on the assumption that physiological signals such as EEG, ECG, and eye tracking can be classified in real time into workload, stress, and emotional valence accurately enough, and that these states actually drive trust in the way Table 3 encodes.","fun_headline_variants_meta":{"raw":{"variants":["Biosignals tune AI explanations to operator trust","AI adapts explanations from your stress and workload","Closed-loop XAI reads your stress and changes its advice","Adaptive AI uses heart and eye signals to build trust","Trust-calibrated AI: from biosignals to tailored explanations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000614,"raw_usage":{"total_tokens":2836,"prompt_tokens":912,"completion_tokens":1924,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":1844}},"tokens_in":528,"tokens_out":1924,"duration_ms":14554,"temperature":1.0,"reasoning_tokens":1844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:31:29.414066+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a simulated emergency-response task with operators wearing EEG, ECG, and eye-tracking sensors, and compare AXTF-driven adaptive explanations against static explanations on measures of self-reported trust, cognitive load, and task performance; the framework is falsified if the adaptive condition does not improve trust or performance. A more targeted test collects concurrent physiological features, performance scores, and self-reported trust from operators during a high-pressure task and checks whether the Table 3 rule 'IF W=High AND S=High AND E=Negative THEN T=Low' systematically holds.","supporting_citations":[{"cited_title":"Meyerson, K","cited_arxiv_id":null,"evidence_quote":"Defines swift trust in temporary groups, the trust construct the framework is designed to foster."},{"cited_title":"Paleja, M","cited_arxiv_id":null,"evidence_quote":"Shows explanation granularity and cognitive load interact, motivating granularity adaptation in the framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Argues transparency, explainability, and situation awareness jointly shape human-AI team performance; the basis for the explanation-modulation rules."},{"cited_title":"Sreedharan, T","cited_arxiv_id":null,"evidence_quote":"Supplies the model-reconciliation technique that adaptive explanation adaptation builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Relates estimated user trust to transparency behavior, a direct precursor of the trust-sensitive adaptation loop."}],"review_version":1}