REVIEW 3 major objections 5 minor 14 references
Emotion-sensitive Explanation Model
T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read Emotion-aware AI explanations trigger on user arousal
desk verdict A clear conceptual proposal whose main gap is that the emotional module emits only a binary alarm and never maps emotion to a strategy; still worth engaging for affective XAI researchers. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is the emotion-sensitive grounding hierarchy, a transfer of Clark-style grounding to affective human-agent interaction. It is realized by an emotional model that detects arousal via EmoNet facial-expression recognition and smartwatch heart-rate variability, using a rolling z-score anomaly detector (threshold 2.5 over a 500 ms window) to spot deviations; a cognitive model that performs risk assessment and hosts an LLM-based understanding dialog with the robot Floka; and a phase-control component that orchestrates the transitions from arousal to understanding to agreement, allowing extra phases to be inserted.
What would settle it
Run a controlled study in which users receive explanations while their facial expressions and heart-rate variability are recorded, and compare system-triggered clarification events against independently rated states of confusion or understanding; if the z-score detector's triggers do not coincide with genuine comprehension difficulties (measured by follow-up comprehension questions), or if triggering clarification dialogs yields no comprehension gain over a no-trigger baseline, the emotion-sensitive grounding loop has no empirical benefit.
Extended reading notes
Core claim
The central claim is that grounding theory, originally developed for spoken-language systems, can be transferred to human-explainer interaction and extended with emotion-sensitive steps. The authors adapt the eight-level grounding hierarchy of a spoken-language telephone system to the human explainee, adding three specific stages: emotional or epistemic arousal, understanding, and agreement. At the first stage, the system watches for micro-reactions such as irritation or surprise; when a deviation is detected, it initiates a clarification sub-dialog to verify understanding, and only after understanding is established does it assess whether the user agrees with the explanation. This makes the interaction mixed-initiative: the explaining system, not only the user, can start a clarification when it senses potential misunderstanding.
Load-bearing premise
The load-bearing assumption is that emotional or epistemic arousal can be reliably detected in real time from facial expressions and heart-rate variability by a rolling z-score anomaly detector with a 2.5 threshold over a 500 ms window, and that such detected deviations are valid triggers for clarification dialogs.
Editorial extensions
If this is right
- XAI systems could shift from static, one-size-fits-all explanations to adaptive strategies that respond to the user's real-time emotional state.
- Detecting subtle arousal signals lets the system initiate clarification dialogs on its own, enabling true mixed-initiative explanation rather than waiting for the user to ask.
- The three-stage structure gives dialog designers a principled ordering: secure understanding before seeking agreement, and treat emotional reactions as a trigger rather than noise.
- By tuning explanation delivery toward an optimal arousal window, the model provides a concrete route to mitigate the documented negative association between emotional intensity and explanation understanding.
- The agreement stage explicitly opens the door to counterfactual explanations—presenting decisions without a contested feature—so the system can address biased or outdated training data.
Reading between the lines
- A natural testable extension is to calibrate the arousal threshold empirically: compare the rolling z-score trigger against ground-truth labels of user confusion, which the paper does not provide.
- The model could be broadened beyond general arousal to distinguish epistemic emotions such as confusion, curiosity, and surprise, which likely require different explanatory responses than an undifferentiated arousal spike.
- The agreement stage implies a values-sensitive design: if a user objects to a feature like gender, the system can offer a counterfactual without it, effectively making the AI's value assumptions negotiable; the paper leaves this ethical dimension implicit.
- The architecture is modular enough that the phase-control component could insert additional phases (for example, a 're-explain after disagreement' loop) without changing the grounding hierarchy, a flexibility the paper notes but does not explore.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a three-stage emotion-sensitive explanation model for XAI: (1) detection of emotional or epistemic arousal, (2) verification of understanding, and (3) assessment of agreement. The model is intended to ground explanations in a multimodal interaction with a virtual robot, using facial expression and heart-rate signals to detect deviations via a rolling z-score anomaly detector, and to initiate clarification dialogues when deviations occur. The authors argue that this provides a conceptual basis for XAI systems that dynamically adapt explanation strategies to users' emotional states, ultimately supporting better and more user-centered decision-making.
Significance. If fully realized, the model would address a recognized gap in XAI research: most explanation frameworks assume a rational user and do not account for emotional or epistemic arousal during explanation processing. The grounding-based framing grounded in prior work on human communication is a useful conceptual starting point, and the explicit integration of both emotional and epistemic states is commendable. However, the current manuscript is a conceptual sketch rather than a validated model. The described implementation does not actually deliver the claimed dynamic adaptation, and key sensing details contain errors. The paper's value at this stage is as a proposal; its claims need to be either substantiated or substantially scaled back.
major comments (3)
- [Section 3, Phase 1 and Figure 2] The emotional module outputs only a binary deviation flag: a rolling z-score anomaly detector (threshold = 2.5, 500 ms window) either triggers the same clarification dialog branch or allows the system to continue to the next feature. No valence, arousal direction, or emotion category is encoded, and no mapping from any specific detected state to a particular explanation strategy is provided. The abstract's claim that the model "dynamically adapt[s] explanation strategies to users' emotional states" is therefore unsupported by the described implementation. The authors should either specify a mechanism by which the type or direction of arousal selects among the "range of different explanation strategies" mentioned in Section 2.1, or revise the claims to describe an arousal-triggered clarification dialog without the stronger adaptation claim.
- [Section 3, Emotional Model] The text states that real-time emotion recognition is performed via facial expressions detected by EmoNet [4]. However, reference [4] describes EmoNet as a speech emotion recognition framework, not a facial expression detector. This is a load-bearing mis-specification because Phase 1's sensing modality is a core part of the proposed model. The authors must correct the citation or the described method, or clarify which actual affect-recognition tool is intended for facial expression analysis.
- [Section 4, Discussion] The discussion invokes the inverted U-shaped relationship between arousal and cognitive performance, stating that the system "targets the optimal arousal window for effective understanding." But the anomaly detector uses a two-sided threshold (|z| > 2.5), which treats high and low arousal as the same trigger: both initiate the same clarification dialog. Thus the model cannot distinguish over-arousal from under-arousal and cannot selectively steer users toward an optimal window without further differentiation. The inverted-U framing is therefore inconsistent with the implemented detection logic. Please either add a mechanism that distinguishes arousal direction and types (e.g., separate thresholds or classification of the anomaly sign) or temper the discussion to reflect what the model actually does, which is to react to any deviation.
minor comments (5)
- [References] References [9] and [10] appear to refer to the same manuscript with slightly different author spellings and publication status; please reconcile and provide a single complete citation. Also, the grounding theory is attributed to "Clark" in the text without a full reference; please add a proper citation.
- [Introduction] There is a typo in the first paragraph: "purly logical considerations" should be "purely logical considerations." In Section 2, "so calledcognitive forcing functions" is missing a hyphen and space; it should be "so-called cognitive forcing functions."
- [Figure 2] In the text, "see fig 2" should be capitalized and formatted as "see Fig. 2." Also, the figures are referenced but not visible in the provided manuscript; please ensure they are legible and correctly placed.
- [Section 3, Emotional Model] The specific values of the z-score threshold (2.5) and window (500 ms) are presented as design choices without justification or citation. If these are provisional placeholders, please state so explicitly; otherwise, provide a rationale for their selection.
- [Abstract] The phrase "users emotional states" is missing an apostrophe; it should be "users' emotional states."
Circularity Check
No material circularity: the emotion-sensitive explanation model is a conceptual proposal motivated by prior empirical work, not a derivation from fitted inputs or self-cited theorems.
full rationale
The paper does not present a derivation chain in which an output is equivalent to an input by construction. The three-stage model (emotional/epistemic arousal, understanding, agreement) is a proposed architecture, explicitly grounded in external work such as Brennan and Hulteen's grounding hierarchy and EmoNet. The authors' own prior results ([9], [10]) are used as empirical motivation for why emotions matter in explanation, but they do not define the model's stages, and no fitted parameter is renamed as a prediction. The only quantitative component, the rolling z-score anomaly detector with threshold 2.5 over a 500 ms window, is a stated parameter choice rather than a parameter fitted to the claimed outcome of effective explanation. The Discussion explicitly defers evaluation ('Future work will evaluate the model's effectiveness in a controlled user study'), so there is no fitted-input-called-prediction loop. The skeptical concern that the binary arousal flag does not yet select among explanation strategies is a design or correctness gap, not circularity: the model is under-specified, but it is not equivalent to its inputs. Self-citations are present but are not load-bearing in the sense of substituting for an argument or importing a uniqueness theorem. The central claim has independent conceptual content.
Assumptions & free parameters
free parameters (2)
- rolling z-score threshold =
2.5
- anomaly detection window =
500ms
assumptions (4)
- domain assumption Prior and task-generated emotions can negatively affect understanding of XAI explanations.
- domain assumption There is an inverted U-shaped relationship between arousal and cognitive performance.
- domain assumption Grounding hierarchies from human dialogue can be transferred to human-agent interaction for emotional states.
- domain assumption Real-time emotional arousal can be reliably measured via facial expressions (EmoNet) and heart-rate variability from a smartwatch.
invented entities (1)
-
Emotion-sensitive grounding hierarchy
Cite this review
Pith. "Pith review of Emotion-sensitive Explanation Model." pith.science (2026). https://pith.science/paper/NUYDMDFV
@misc{pith2026250510454,
author = {Pith},
title = {Pith review of: Emotion-sensitive Explanation Model},
year = {2026},
howpublished = {\url{https://pith.science/paper/NUYDMDFV}},
note = {Machine review of arXiv:2505.10454}
}
read the original abstract
Explainable AI (XAI) research has traditionally focused on rational users, aiming to improve understanding and reduce cognitive biases. However, emotional factors play a critical role in how explanations are perceived and processed. Prior work shows that prior and task-generated emotions can negatively impact the understanding of explanation. Building on these insights, we propose a three-stage model for emotion-sensitive explanation grounding: (1) emotional or epistemic arousal, (2) understanding, and (3) agreement. This model provides a conceptual basis for developing XAI systems that dynamically adapt explanation strategies to users emotional states, ultimately supporting more effective and user-centered decision-making.
Figures
Reference graph
Works this paper leans on
-
[4]
M. Gerczuk, S. Amiriparian, S. Ottl, and B. W. Schuller. Emonet: A transfer learning framework for multi-corpus speech emotion recog- nition.IEEE Transactions on Affective Computing, 14(2):1472–1487, 2021
work page 2021
-
[1]
D. Battefeld, S. Mues, T. Wehner, P. House, C. Kellinghaus, J. Wellmer, and S. Kopp. Revealing the dynamics of medical diagnostic reasoning as step-by-step cognitive process trajectories. InProceedings of the Annual Meeting of the Cognitive Science Society, volume 46, 2024
work page 2024
-
[2]
S. E. Brennan and E. A. Hulteen. Interaction and feedback in a spoken language system: A theoretical framework.Knowledge-based systems, 8(2-3):143–151, 1995
work page 1995
-
[3]
Z. Buçinca, M. B. Malaya, and K. Z. Gajos. To trust or to think: cognitive forcing functions can reduce overreliance on ai in ai-assisted decision-making.Proceedings of the ACM on Human-computer Inter- action, 5(CSCW1):1–21, 2021
work page 2021
-
[5]
L. Guerdan, A. Raymond, and H. Gunes. Toward affective xai: facial affect analysis for understanding explainable human-ai interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3796–3805, 2021
work page 2021
-
[6]
O. Lammert. Can ai regulate your emotions? an empirical investigation of the influence of ai explanations and emotion regulation on human decision-making factors. InWorld Conference on Explainable Artificial Intelligence. Springer, forthcoming, 2025
work page 2025
-
[7]
O. Lammert, B. Richter, C. Schütze, K. Thommes, and B. Wrede. Hu- mans in xai: increased reliance in decision-making under uncertainty by using explanation strategies.Frontiers in Behavioral Economics, 3: 1377075, 2024. doi: 10.3389/frbhe.2024.1377075
-
[8]
J. S. Lerner, Y . Li, P. Valdesolo, and K. S. Kassam. Emotion and deci- sion making.Annual review of psychology, 66(1):799–823, 2015
work page 2015
Show all 14 references
-
[9]
Richter, C
B. Richter, C. Schütze, A. Aksonova, and B. Wrede. Influence of prior and task generated emotions on xai explanation retention and under- standing, 2025. URL https://arxiv.org/abs/2505.10427
2025 arXiv
-
[10]
Richter, C
B. Richter, C. Schütze, A. Aksonovaa, and B. Wrede. Influence of prior and task generated emotions on xai explanation retention and under- standing. Manuscript submitted for publication, 2025
2025
-
[11]
Schmid and B
U. Schmid and B. Wrede. Explainable ai.KI-Künstliche Intelligenz, 36 (3):207–210, 2022
2022
-
[12]
Simkute, E
A. Simkute, E. Luger, B. Jones, M. Evans, and R. Jones. Explainability for experts: A design framework for making algorithms supporting ex- pert decisions more explainable.Journal of Responsible Technology, 7: 100017, 2021
2021
-
[13]
Thommes, O
K. Thommes, O. Lammert, C. Schütze, B. Richter, and B. Wrede. Hu- man emotions in ai explanations. In L. Longo, S. Lapuschkin, and C. Seifert, editors,Explainable Artificial Intelligence, pages 270–293. Springer, 2024. doi: 10.1007/978-3-031-63803-9_15
2024 doi
-
[14]
D. Wang, Q. Yang, A. Abdul, and B. Y . Lim. Designing theory-driven user-centric explainable ai. InProceedings of the 2019 CHI conference on human factors in computing systems, pages 1–15, 2019
2019
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.