{"id":"8a9e24db-4abf-4918-b5c1-694f0b78bddc","arxiv_id":"2504.17170","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Human-AV communication should adapt to context, rider traits, and goals, as shown by two experiments and a machine-learning trust prediction study.","lead":"This dissertation reports three studies on how self-driving cars should talk to passengers. It finds that communication errors hurt trust even when driving is perfect, and that personal traits and driving context change what riders need.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Chapter 3's error manipulation is not communication-only: incorrect 'what'/'why' statements imply faulty perception or planning, so the 'even if driving is perfect, explanations alone reduce trust' claim is not cleanly supported.","rationale":"The reader identifies Chapter 2's practice-effect confound as the weakest assumption. That is a real limitation, and the authors candidly flag it in §2.5.2. However, it is not the most load-bearing for the dissertation's central claim: the between-group comparisons for racing-line distance (Group 2 vs Group 1, Group 4 vs Group 1) already hold practice constant because Group 1 is an observation-only control. The all-groups pre-post gains are contaminated by practice, but those are not the only—or primary—evidence for coached improvement. In contrast, Chapter 3 is the chapter that directly supports the 'even with perfect driving, inaccurate explanations hurt adoption' claim, which the abstract and conclusion elevate to a general design principle. The manipulation in that chapter changes the AV's apparent beliefs about the world, not just the accuracy of its communication channel: a car that says 'merging right' while braking, or names the wrong obstacle, is a car whose perception or planning appears faulty. The drop in 'confidence in driving ability' is therefore an expected consequence of inferred incompetence, not a demonstration that users separate explanation quality from driving quality. This leaves the central claim less secure than the prose suggests. A replication or secondary analysis that separates diagnostic from non-diagnostic explanation errors would settle the point. If the effect only appears for diagnostic errors, the thesis needs to be narrowed to 'communication errors that reveal underlying system faults,' which is a weaker but still useful claim. The current reader verdict of CONDITIONAL remains appropriate; no verdict change is needed, but the condition should explicitly include this confound.","tokens_in":45695,"tokens_out":8789,"duration_ms":93817,"concrete_test":"Run a preregistered replication of Chapter 3 with two error types: (a) diagnostic errors that imply wrong perception/planning, as in the current design, and (b) communication-only errors that are semantically detached from the driving decision, e.g., the AV brakes correctly for a pedestrian but says it is braking because of an unrelated event or states the wrong street name. Keep the identical videos and match assignment between conditions. If confidence in driving ability drops only for diagnostic errors, the crossover effect is driven by perceived system incompetence, not by explanation accuracy per se, and the headline claim must be qualified. If it drops for both, the 'even if driving is perfect' claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that AV communication must be accurate independently of driving quality—rests on Chapter 3's crossover result: confidence in the AV's driving ability dropped across error conditions despite identical trajectories (Table 3.3; §3.5.1). But the manipulation does not isolate explanation accuracy from perceived system competence. In the 'Low' condition, the AV says 'Braking, a cyclist is crossing the road' when the obstacle is a pedestrian; in the 'High' condition it says 'Merging right' while braking (Table 3.2). These are not merely inaccurate descriptions of a communication channel; they are incorrect claims about the AV's perceptual beliefs and intended actions. A participant who hears 'a cyclist is crossing' while seeing a pedestrian can rationally infer that the AV's object-recognition or planning model is wrong, and therefore that its future driving may be unsafe. The observed reduction in 'confidence in driving ability' may then be a sensible update about the AV's competence, not evidence that explanation quality is evaluated independently of driving quality. Because the dissertation explicitly frames the result as 'despite identical driving' (§3.1, §3.5.1), the inference from this study overreaches if the errors are diagnostic of the driving system. The within-subjects presentation of the same scenario three times also invites contrast effects that can inflate the apparent crossover, although this is secondary. The strongest empirical pillar for the thesis therefore needs a cleaner operationalization before the 'even if driving is perfect' claim is secure.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The dissertation argues that human-autonomous-vehicle (AV) interaction should be understood as a complex system in which successful communication depends on the driving context, the interaction goal, human traits, and AV traits. Three empirical studies support the argument. Chapter 2 reports a pre-post driving-simulator experiment (n=41) evaluating an AI race-car driving coach that provides 'what' and 'why' explanations in visual or auditory modalities; it concludes that AI coaching can improve novice performance-driving skills and that modality-appropriate, task-sensitive explanations reduce cognitive load and uncertainty. Chapter 3 reports a within-subjects online video study (n=232) in which AV explanation errors were manipulated while the AV's driving was identical across conditions; it finds that explanation errors reduce comfort, reliance, satisfaction, and confidence in the AV's driving ability, with context (perceived harm, difficulty) and personal traits (initial trust, expertise) moderating these outcomes. Chapter 4 applies machine learning to survey data from young adults to predict trust in AVs from psychosocial traits, risk-benefit attitudes, and driving factors, reporting that risk and benefit perceptions are the strongest predictors. The dissertation concludes with design implications for transparent, adaptable, and personalized AV communication.","tokens_in":45998,"tokens_out":5399,"duration_ms":54030,"significance":"If the central claims hold, the dissertation makes a useful contribution to human-AI interaction and AV human factors. Its strengths include three distinct empirical studies, the use of linear mixed-effects models with random effects for participant and scenario, and a relatively large Chapter 3 sample. The chapters have already appeared in or been accepted for peer-reviewed venues, which speaks to the maturity of the individual studies. Chapter 4's machine-learning pipeline with SHAP values and ablation experiments provides a concrete, reproducible method for studying trust prediction. The dissertation is also unusually candid about its limitations, particularly the practice-effect confound in Chapter 2 and the post hoc scenario exclusions in Chapter 3. The proposed complex-systems framing is generative and practical. However, the strongest interpretive claim, that explanation errors matter 'even if the AV's driving performance is perfect,' is not cleanly supported by the Chapter 3 manipulation, because the errors used there are diagnostic of the AV's perceptual and planning competence.","major_comments":[{"comment":"The crossover claim that explanation errors reduce 'confidence in driving ability' despite 'identical driving' is not cleanly supported, because the error manipulation does not isolate communication accuracy from perceived system competence. In the Low condition the AV says 'Braking, a cyclist is crossing the road' while the video shows a pedestrian; in the High condition it says 'Merging right, a cyclist is crossing the road' while the vehicle is slowing and the obstacle is a pedestrian. These are not merely inaccurate descriptions of an independent communication channel; they are incorrect statements about the AV's perceptual beliefs and intended actions. A participant can rationally infer that the AV's object recognition or planning model is faulty and update their confidence in the AV's driving ability accordingly. The dissertation's strong conclusion in Section 3.6 ('even if the AV's driving performance is perfect ... people may still refuse to adopt AV technology') therefore overreaches. The study remains informative about holistic judgments of AV competence, but the claim that explanation quality is evaluated independently of driving quality must be reframed, or the design needs errors that cannot be diagnostic of perception or planning (and a manipulation check confirming that participants perceive the errors as communication-only). The within-subjects repetition of the same scenario three times is a secondary source of contrast effects and should be acknowledged in the interpretation.","section":"Section 3.4.3, Table 3.2; Section 3.5.1; Section 3.6"},{"comment":"The post hoc removal of three scenarios because the AV's driving was deemed 'imperfect' is a defensible quality check, but it means the 'identical driving' claim is defined only on the retained subset of 24 scenarios. Please report the main analyses both with and without the removed scenarios, and justify that the exclusion rule was applied independently of the outcome measures. Relatedly, the binary reliance measure ('rely on AV' vs. 'take control myself') is scaled from 0-1 to 0-10 and entered into linear mixed-effects models as if it were continuous. A mixed-effects logistic model or ordinal model would be more statistically appropriate. This does not necessarily change the direction of the reported effects, but a dissertation-level analysis should address the distributional issue rather than only noting that scaling does not affect interpretation.","section":"Section 3.4.2 and Section 3.4.4"},{"comment":"The pre-post design has no practice-only control group for the all-groups-combined improvements, and the manuscript explicitly acknowledges this limitation. Because the chapter's broad claim that 'AI coaching can effectively teach performance driving skills' partly relies on these combined improvements, the practice confound is load-bearing; the credible between-group evidence is essentially limited to racing-line distance in a sample of n=41. In addition, the analysis runs numerous linear mixed-effects models across multiple performance and self-report outcomes without correction for multiple comparisons. I recommend reporting effect sizes and confidence intervals, and either correcting for multiple comparisons or explicitly labeling secondary analyses as exploratory. This does not undermine the chapter's contribution, but it should be reflected in the strength of the conclusions.","section":"Chapter 2, Sections 2.4.2 and 2.5.2"}],"minor_comments":[{"comment":"The caption contains a typo: 'Presentaton' should be 'Presentation'.","section":"Table 2.1 caption"},{"comment":"The text says 'Individual comparisons between each condition ... show highly significant effects ... (Figure 3.4)' but the pairwise condition comparisons are reported in Table 3.4; Figure 3.4 is the later chart of reliance-decision factor importance. Please correct the cross-reference.","section":"Section 3.5.1"},{"comment":"The model-development section would benefit from explicit hyperparameter settings, software versions, and the exact train/validation split used, so that the SHAP analyses are reproducible from the text.","section":"Section 4.4.3"},{"comment":"The interaction-effect discussion distinguishes significant, trending, and non-significant results in the text, but the tables only use significance stars; adding standard errors or confidence intervals to the text summary would help readers calibrate the strength of the moderation effects.","section":"Chapter 3, Section 3.5.2"},{"comment":"The interview quotes are illustrative but are not accompanied by a formal coding scheme or inter-rater reliability; a sentence describing how themes were derived would strengthen the mixed-methods claims.","section":"Chapter 2, Section 2.4.5"}],"recommendation":"major_revision","confidential_remarks":"The dissertation is largely assembled from already-published or forthcoming peer-reviewed papers, and the complex-systems framing is synthetic rather than a new derivation. That is not itself a problem for a dissertation archive, but the editorial framing in Chapter 1 should not oversell the Chapter 3 result. If the authors reframe the 'even if driving is perfect' conclusion as a claim about perceived competence rather than a clean separation of explanation quality from driving quality, I would be supportive of eventual acceptance. The main risk is the overreach identified in my first major comment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The dissertation is a solid collection of three empirical studies, each already peer-reviewed in good venues. The genuinely new work is Chapter 3: a controlled, well-powered (n=232) test of how AV explanation errors affect trust, reliance, satisfaction, and confidence. Showing that errors move all four outcomes, and that the effect scales with error severity, is a real contribution. Chapter 2 is a smaller (n=41) simulator study of an AI driving coach; it has a plausible design and honest limitations, but the lack of a practice-only control really does limit the causal claim, exactly as the author says in Section 2.5.2. Chapter 4 is a competent ML prediction exercise; the SHAP analyses are transparent, and the finding that risk/benefit attitudes beat demographics is believable. The framework in Chapter 1 is more of a useful organizational lens than a rigorous theory, and it leans on the author's own prior work, but the empirical chapters stand independently and the self-citation is not circular.\n\nThe soft spot worth flagging is the interpretation of Chapter 3. The stress-test concern lands: the 'low' and 'high' error conditions are not purely communication failures. When the AV says 'a cyclist is crossing' while a pedestrian is visible, or says 'merging right' while braking, the participant is rationally entitled to infer that the AV's perception or planning is faulty. So the drop in 'confidence in driving ability' is not clean evidence that explanation quality is evaluated independently of driving quality. The crossover is still interesting and important, but the strong claim—'even if driving is perfect, explanations alone reduce adoption'—is not fully supported by this manipulation. A cleaner test would use errors that are clearly external to the driving system, or would add a condition where the AV explicitly acknowledges uncertainty. The author does discuss the crossover, but the framing leans on the strong reading.\n\nThat said, the paper is honest, the analyses are mostly appropriate, and the limitations are stated rather than hidden. The Chapter 3 result will be useful to anyone working on AV HMIs or trust in automation, even if the precise mechanism needs follow-up.\n\nRecommendation: this deserves a serious referee. It is a degree dissertation whose chapters have already passed CHI/Nature-level review; the compilation adds a coherent systems framing. I would not desk-reject it. I'd send it out, with an instruction to the author to soften the Chapter 3 causal language or add a disambiguation study.","headline":"A useful, honest dissertation whose Chapter 3 provides a real but over-stated result: explanation errors hurt trust even with identical driving, but the errors also leak information about perceived driving competence.","tokens_in":46464,"tokens_out":1020,"would_cite":true,"duration_ms":11736,"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 dissertation argues that AV communications must adapt to rider, goal, and context; three studies show explanations can teach, errors can erode trust even when driving is flawless, and risk-benefit attitudes best predict trust.","keywords":["autonomous vehicles","explainable AI","human-vehicle trust","multimodal explanations","situational awareness","driving simulation","personalized interfaces","risk-benefit attitudes"],"falsifier":"Run the same full-motion simulator protocol with an added control group that drives the same pre and post laps with the same time-on-task but no AI coach; if that group shows pre-post gains equal to the coached groups, the claim that AI explanations drive learning is falsified.","tokens_in":45472,"feed_emoji":"🚗","tokens_out":6776,"duration_ms":62822,"temperature":0.7,"pith_summary":"This dissertation argues that human-autonomous-vehicle (AV) interaction fails when design treats all riders and situations the same, and succeeds when communication adapts to the person, the goal, the driving context, and the AV's own traits. In a driving-simulator experiment, an AI coach that explained what it was doing and why improved novices' racing-line performance, with visual 'what' cues plus auditory 'why' cues outperforming all-auditory explanation. In a 232-person video study, inaccurate AV explanations reduced riders' comfort, reliance, satisfaction, and even their confidence in the AV's driving ability, even though the driving shown was identical and flawless across conditions. A machine-learning survey study found that perceived risks and benefits of AVs are the strongest predictors of young adults' trust, outweighing other personal characteristics. Together the studies support transparent, adaptable, personalized AV communication.","feed_headline":"Perfect driving still fails if AV explanations err","feed_subtitle":"In a 232-rider simulation, wrong explanations cut comfort, reliance, and confidence despite flawless driving.","key_machinery":"The central object is the human-AV complex system viewed as joint action: a multi-level success function in which high-level goals like safe transport or learning have subgoal criteria, and those criteria shift when components of the system change. Situational awareness—perception, comprehension, and projection—is the mechanism by which the team coordinates, and explanatory communication ('what' the AV is doing and 'why') is the main lever for building it. The dissertation manipulates that lever across three studies: explanation type and modality, explanation errors, and personal predictors of trust.","core_discovery":"In its own terms, the dissertation's central claim is that a human and an autonomous vehicle form a joint-action team whose success depends on filling situational-awareness gaps, and that these gaps change with the human's traits, the AV's traits, the goal, and the driving context. The sharpest evidence comes from the explanation-error study: riders saw identical, flawless driving, yet when the AV's spoken and on-screen explanations were inaccurate, comfort, reliance, satisfaction, and confidence in driving ability all fell significantly, with bigger falls for more severe errors. The AI-coach study shows the same communication machinery can be tuned for learning: visual 'what' guidance plus auditory 'why' rationale improved novices' racing-line performance without the overload produced by all-auditory explanations. The trust-prediction study shows that perceived risks and benefits of AVs dominate other personal factors in predicting young adults' trust. The dissertation concludes that accurate, context-sensitive, personalized explanations are a necessary condition for AV acceptance.","pith_inferences":["Editorial: If explanation errors contaminate perceived driving ability, then a Level-3 conditional automation system could provoke unsafe takeovers after a communication failure; a takeover-timing study would test this.","Editorial: The dominance of risk-benefit perceptions implies trust interventions could work through benefit framing and risk communication rather than more transparency alone; the dissertation does not test this.","Editorial: The error-to-driving-confidence crossover may generalize to other XAI domains, where explanation quality shapes perceived model competence even when outputs are correct."],"forward_implications":["AV explanation systems should be held to accuracy standards comparable to driving-system safety, because wrong explanations reduce riders' willingness to rely on the AV even when the driving is perfect.","HMI design should match information to modality: visuospatial 'what' guidance works best as a visual cue, while 'why' rationale can be auditory, reducing uncertainty and cognitive overload.","Communication should adapt to context: high perceived harm amplifies the damage of explanation errors, while in very difficult situations riders may prefer to keep relying on the AV rather than take over.","Personalization of trust should target risk and benefit perceptions, since these predicted young adults' AV trust more strongly than other psychosocial traits in the machine-learning model.","Explanation systems should be deployed only when their accuracy is assured, since the studies indicate faulty communication can be worse than silence or observation alone."],"supporting_citations":[{"why":"Supplies the situational awareness lens that frames perception, comprehension, and projection as the informational foundation of successful human-AV teaming.","marker":"[63]"},{"why":"Provides the trust-in-automation model whose dispositional, situational, and learned trust factors organize the contextual and personal variables tested in Chapters 3 and 4.","marker":"[87]"},{"why":"Frames explainable AI as the mechanism for making black-box decisions understandable and trustworthy, which motivates studying explanation delivery and errors.","marker":"[77]"},{"why":"Grounds the 'what' and 'why' explanation distinction used to design and label the explanation manipulations.","marker":"[159]"},{"why":"Prior experimental work on how and why AV explanations affect driver attitudes and performance that this dissertation extends to the error case.","marker":"[122]"},{"why":"Prior evidence that AV-caused errors damage user trust more than external errors, serving as the baseline expectation for error effects in Chapter 3.","marker":"[148]"},{"why":"Supports the claim that trust declines after errors are long-lasting, motivating the study of explanation errors before deployment.","marker":"[201]"}],"fun_headline_variants":["Wrong AV explanations erode trust even with perfect driving","AV communication must adapt: one-size-fits-all fails riders","Personalized AV explanations beat generic ones for trust and learning","Flawless driving isn't enough: AV explanations make or break trust"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Chapter 2 claim that the AI coach caused the learning gains assumes the gains are larger than plain practice would produce, but the study has no practice-only control group, a gap the author acknowledges; if practice alone explains the gains, the coaching effect does not stand.","fun_headline_variants_meta":{"raw":{"variants":["Wrong AV explanations erode trust even with perfect driving","AV communication must adapt: one-size-fits-all fails riders","Personalized AV explanations beat generic ones for trust and learning","Flawless driving isn't enough: AV explanations make or break trust"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3234,"prompt_tokens":1023,"completion_tokens":2211,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":2141}},"tokens_in":639,"tokens_out":2211,"duration_ms":13386,"temperature":1.0,"reasoning_tokens":2141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:46:35.042821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same full-motion simulator protocol with an added control group that drives the same pre and post laps with the same time-on-task but no AI coach; if that group shows pre-post gains equal to the coached groups, the claim that AI explanations drive learning is falsified.","supporting_citations":[{"cited_title":"Differential impact of autonomous vehicle malfunctions on human trust","cited_arxiv_id":null,"evidence_quote":"Supports the claim that trust declines after errors are long-lasting, motivating the study of explanation errors before deployment."}],"review_version":1}