SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations

Jina SuhLindy LeErfan ShayeganiGonzalo RamosJudith AmoresDesmond C. OngMary CzerwinskiJavier Hernandez

article2026IEEE Transactions on Affective Computing10 citations

Introduces the SENSE-7 taxonomy and dataset of real-world human-AI conversations with turn-by-turn user annotations, providing a practical foundation for measuring and modeling context-dependent empathy in conversational agents.

Listen

As artificial intelligence conversational agents are increasingly deployed in workplace and personal settings, digital empathy has become a central focus for user engagement and satisfaction. However, conventional engineering approaches often attempt to simulate internal, humanlike emotional states or rely on third-party crowdworker ratings, ignoring the subjective, dynamic, and relational nature of how users actually perceive empathy during real-world interactions.

The article aims to introduce a human-centered, seven-dimension behavioral taxonomy of digital empathy and evaluate how real users perceive these empathic behaviors across sustained, multi-turn human-AI conversations.

To conduct this evaluation, the researchers developed the SENSE-7 framework and gathered 695 naturalistic conversations from 109 information workers interacting with four distinct large language model configurations across informational, personal, and work-related tasks. Participants provided per-turn and post-conversation empathy ratings alongside psychometric profiles, contextual goals, and mood indicators. The authors then performed mixed-effects statistical modeling, qualitative feedback analyses, and an automated classification experiment on a fully anonymized 672-conversation subset using an advanced language model.

The analysis yielded several key findings: First, empathy perception is highly fragile; even a single poor turn substantially reduced overall conversation empathy ratings and significantly harmed user engagement, task success, and future adoption intent. Second, user expectations and individual traits heavily dictate desired empathy; participants tackling personal or work issues desired significantly higher empathy, whereas individuals with high cognitive reappraisal skills desired less empathy when handling personal or work problems. Third, users prioritized cognitive understanding and response appropriateness over superficial affective displays, frequently expressing frustration with unsolicited advice or generic bulleted lists. Finally, baseline automated classification using an advanced language model achieved an accuracy of 48.7% and exact-or-within-one-level accuracy of 96.0% (Spearman rank correlation of 0.369), demonstrating the feasibility of automated empathy measurement while highlighting remaining technical challenges.

These findings indicate that empathy in conversational systems should be treated as an interactional, context-sensitive behavior rather than a simulated internal state. For enterprise deployments, uncalibrated or generic emotional expressions risk appearing superficial, invalidating user feelings, and eroding trust. Conversely, systems that appropriately adapt their communication style to user context and task type can enhance engagement and user satisfaction.

Organizations developing or integrating conversational agents should implement dynamic empathy calibration that adjusts responses based on task context and explicit user preferences, allowing users to opt out of empathic styling when only factual assistance is required. AI architectures should incorporate conversational repair mechanisms to detect and recover from misaligned turns, while expanding memory mechanisms to preserve relational continuity across multi-turn exchanges.

These conclusions should be interpreted in light of certain limitations: the participant sample was restricted to information workers at a single technology company with generally positive attitudes toward AI, and findings were derived exclusively from text-based models. While confidence in the seven-dimension behavioral taxonomy and empirical findings is high within this operational setting, cautious validation is recommended before generalizing to non-technical demographics or high-stakes contexts such as specialized mental healthcare.

Microsoft University of California, Riverside University of Michigan University of Texas at Austin University of Washington classifiers communication language model SENSE-7 human-AI conversations artificial empathy

arXiv: 2509.16437 microsoft/sense-7

Paper: Can Large Language Models be Good Emotional Supporter? Mitigating Preference Bias on Emotional Support Conversation, Dongjin Kang et al. (2024). This study analyzes how large language models navigate multi-stage emotional support conversations and suffer from strategy bias, providing key background for evaluating observable empathic behaviors.
Paper: Evaluating Very Long-Term Conversational Memory of LLM Agents, Adyasha Maharana et al. (2024). This paper establishes evaluation frameworks for long-term conversational memory in AI agents, contextualizing how continuity failures disrupt sustained human-AI relationships.
Paper: The Past, Present and Better Future of Feedback Learning in Large Language Models for Subjective Human Preferences and Values, Hannah Kirk et al. (2023). This survey examines the challenges of learning from subjective human preferences and values, grounding SENSE-7's approach to user-perceived and individualized empathy.
Paper: MT-Bench-101: A Fine-Grained Benchmark for Evaluating Large Language Models in Multi-Turn Dialogues, Ge Bai et al. (2024). This work establishes a fine-grained evaluation taxonomy for multi-turn conversational competencies in large language models, setting the methodological baseline for multi-turn dialogue assessment.
Paper: How NOT To Evaluate Your Dialogue System: An Empirical Study of Unsupervised Evaluation Metrics for Dialogue Response Generation, Chia-Wei Liu et al. (2016). This foundational paper demonstrates the discordance between automated metrics and human judgments in dialogue systems, motivating human-grounded evaluation datasets.
Paper: Personalizing Dialogue Agents: I have a dog, do you have pets too?, Saizheng Zhang et al. (2018). This landmark research introduces persona conditioning to maintain conversational consistency and user engagement over interactive dialogues.

Paper: Reasoning Models Generate Societies of Thought, Junsol Kim et al. (2026). This paper investigates how reasoning models internally simulate conversational and socio-emotional roles, offering a mechanistic perspective on generating the adaptive empathic behaviors highlighted by SENSE-7.
Paper: Metacognition in LLMs: Foundations, Progress, and Opportunities, Gabrielle Kaili-May Liu et al. (2026). This comprehensive volume surveys self-monitoring and cognitive regulation in language models, providing frameworks to build conversational agents that reliably adapt to user expectations.
Paper: Agentic Reasoning for Large Language Models, Tianxin Wei et al. (2026). This work details architectures for memory, feedback, and interactive reasoning in autonomous agents, providing the engineering principles necessary to implement sustained, context-sensitive agentic behaviors.

Cover for SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations

Abstract

Empathy is increasingly recognized as a key factor in human-AI communication, yet conventional approaches to "digital empathy" often focus on simulating internal, human-like emotional states while overlooking the inherently subjective, contextual, and relational facets of empathy as perceived by users. In this work, we propose a human-centered taxonomy that emphasizes observable empathic behaviors and introduce a new dataset, Sense-7, of real-world conversations between information workers and Large Language Models (LLMs), which includes per-turn empathy annotations directly from the users, along with user characteristics, and contextual details, offering a more user-grounded representation of empathy. Analysis of 695 conversations from 109 participants reveals that empathy judgments are highly individualized, context-sensitive, and vulnerable to disruption when conversational continuity fails or user expectations go unmet. To promote further research, we provide a subset of 672 anonymized conversation and provide exploratory classification analysis, showing that an LLM-based classifier can recognize 5 levels of empathy with an encouraging average Spearman $\rho$ =0.369 and Accuracy=0.487 over this set. Overall, our findings underscore the need for AI designs that dynamically tailor empathic behaviors to user contexts and goals, offering a roadmap for future research and practical development of socially attuned, human-centered artificial agents.

Topics

classifiers communication language model large language models taxonomy user context dialogue user information classification models

Keywords

SENSE-7 human-AI conversations artificial empathy empathy perception empathy assessment index empathic responses affective empathy conversational agents empathic concern empathic accuracy

I Introduction
II Related Work
III AI Empathic Behavior Scale
III-A Affective Understanding
III-B Cognitive Understanding
III-C Response Appropriateness
III-D Prosocial Expression
III-E Interest
III-F Contextual Understanding
III-G Relational Continuity
IV Human-Centered Empathic Conversation Data Collection
IV-A Study Procedure
IV-A1 Intake survey
IV-A2 AI assignment
IV-A3 Conversation task
IV-A4 Exit survey
IV-B Conversational AI agent
IV-C Study population
IV-D Analysis
IV-D1 Quantitative
IV-D2 Qualitative
IV-D3 Perceived Empathy Recognition
V Results
V-A What factors influence perceptions of empathy?
V-A1 Conversation topics and user expectations
V-A2 Conversation length and turn-by-turn empathy ratings
V-A3 Post-task perceptions
V-B How do empathy dimensions show up in conversations?
V-B1 Good and bad demonstrations of empathy dimensions
V-B2 Empathy as establishing a shared understanding
V-B3 Empathy as adaptability beyond cognitive and affective recognition
V-C Can perceived empathy be automatically measured?
VI Discussion
VI-A Revisiting empathy as an interactional construction
VI-B Human-centered empathic AI design
VI-C Measuring human-centered digital empathy
VI-D Limitations and ethical considerations
VII Conclusion
References

Knowls

Knowl 1 — Seven-Dimensional Digital Empathy Taxonomy
definition
The digital empathy taxonomy reframes psychological empathy concepts into observable, interactional artificial intelligence (AI) behaviors across three core functional categories (Affective, Cognitive, and Motivational) and seven operational dimensions:
1. Affective Understanding (Affective category): The AI agent demonstrates the ability to accurately detect, recognize, and comprehend the user's emotional state, feelings, and affective nuances (e.g., validating frustration or anxiety).
2. Cognitive Understanding (Cognitive category): The AI agent demonstrates comprehension of the user's mental perspective, intentions, points of view, and explicit goals (e.g., reflecting the underlying logic or objectives of the user's situation).
3. Response Appropriateness (Motivational category): The AI agent adapts its response dynamically to user context and experiences, tactfully deciding when to offer solutions, provide advice, validate feelings, or simply listen.
4. Prosocial Expression (Motivational category): The AI agent explicitly conveys warmth, care, supportive intent, and an active desire to assist and promote user well-being.
5. Interest (Motivational category): The AI agent exhibits curiosity, engagement, and active listening by asking relevant follow-up questions and encouraging the user to guide the interaction.
6. Contextual Understanding (Motivational category): The AI agent integrates the user's unique background, personal history, cultural circumstances, preferences, and external societal constraints into the exchange.
7. Relational Continuity (Motivational category): The AI agent establishes sustained rapport across turns and multi-session interactions by recalling, referencing, and weaving past conversational details into current exchanges.
Knowl 2 — SENSE-7 Dataset and Multi-Turn Annotation Protocol
experimental setup
The SENSE-7 (Subjective Empathy in Natural Sustained Exchanges; 7 dimensions) dataset consists of multi-turn conversational interactions between 109 information workers (55 men, 54 women) and conversational AI agents in real-world settings over a 4-week period. The collection protocol captures first-person, second-person (actual interactant), and per-turn annotations structured as follows:
1. Intake Assessment: Evaluates user demographics, trait empathy via the Single Item Trait Empathy Scale (SITES) and Toronto Empathy Questionnaire (TEQ), emotion regulation via the Emotion Regulation Questionnaire (ERQ: cognitive reappraisal and expressive suppression), and attitudes toward AI via the AI Attitude Scale (AIAS-4) and General Attitudes towards Artificial Intelligence Scale (GAAIS).
2. AI Assignment: Participants are block-randomized across four single-blind LLM conversational agents: standard GPT-4 (GPT4), GPT-4 prompted with a 7-dimension empathy system prompt (GPT4-empathy), Llama2-Chat-70B (Llama2-70b), and GPT-3.5-turbo augmented with intent extraction and context memory (IC).
3. Pre-Conversation Survey: Measures selected topic category (Information Task, Personal Issue, or Work Issue across 9 specific topics), perceived task importance ( $1$ to $5$ ), desired empathy level ( $1=\text{minimal}, 2=\text{moderate}, 3=\text{high}$ ), and baseline affect via the 10-item International Positive and Negative Affect Schedule Short Form (I-PANAS-SF).
4. Interaction & Per-Turn Labeling: Users conduct open-ended text dialogue (mean $10.0 \pm 6.2$ turns). Following dialogue completion, users label every AI turn on a 1–5 scale (Very Poor to Very Good, or N/A) for overall empathy and each of the 7 individual empathy dimensions.
5. Post-Conversation Survey: Captures post-task I-PANAS-SF mood, conversational engagement, task success, interaction positivity, reuse intention (1–5 agreement scales), and overall 7-dimension perceived empathy ratings.
The collected corpus contains 695 conversations (6,997 turns), with a fully anonymized public subset of 672 conversations.

Knowl 3 — Post-Task Overall Perceived Empathy Linear Mixed-Effects Model

data/table

A linear mixed-effects model was fit to predict post-task overall perceived empathy (computed as the mean across the seven empathy dimensions) from model type, participant psychometrics, pre-task context, and turn quality, treating participant identity as a random effect (conditional $R^2 = 0.498$ , marginal $R^2 = 0.294$ , group variance $= 0.155$ ):

Parameter	Coefficient ( $\beta$ )	$P$ -value
Intercept	4.151	$<0.001$
Participant-Based
Scenario: GPT4 (Ref: GPT4-empathy)	-0.350	0.015
Scenario: IC	-0.412	0.007
Scenario: Llama2-70b	-0.469	0.001
Gender: Woman (Ref: Man)	0.108	0.339
Age: 26–35 (Ref: 18–25)	-0.185	0.361
Age: 36–45	0.167	0.440
Age: 46–55	0.060	0.772
Age: 56–65	-0.350	0.144
Age: Prefer not to answer	-0.167	0.649
TEQ	0.079	0.485
SITES	-0.010	0.904
ERQ: Cognitive Reappraisal	-0.028	0.632
ERQ: Expressive Suppression	-0.011	0.816
AIAS-4: AI Will Improve Life	-0.017	0.886
AIAS-4: AI Will Improve Work	-0.046	0.713
AIAS-4: Will Use AI	0.270	0.053
AIAS-4: AI Positive For Humanity	-0.132	0.135
GAAIS: Organizations Use AI Unethically	0.068	0.382
GAAIS: Interested Daily AI Use	-0.243	0.014
GAAIS: AI Dangerous	-0.086	0.259
GAAIS: Beneficial AI Applications	0.013	0.908
Would Disclose PII	0.010	0.867
Trust AI	0.199	0.029
Task-Based
Pre-Task Choice: Personal Issue (Ref: Information Task)	0.046	0.501
Pre-Task Choice: Work Issue	0.107	0.110
Desired Empathy For Task	0.041	0.325
Pre-Task Importance	-0.038	0.212
Pre-Task I-PANAS-SF: Positive	0.015	0.060
Pre-Task I-PANAS-SF: Negative	-0.004	0.698
Conversation Has Bad Turn: True (Ref: False)	-0.657	$<0.001$

The regression demonstrates that prompting GPT-4 with a multidimensional empathy system prompt (GPT4-empathy) yields significantly higher perceived empathy than vanilla GPT4, IC, and Llama2-70b. The occurrence of at least one poorly rated turn serves as the strongest negative determinant of post-conversation empathy.

Knowl 4 — Disproportionate Impact of Single 'Poor Turns' on Perceived Empathy and User Engagement
empirical result
The presence of a single turn rated as "poor" or "very poor" (score $< 3$ on any dimension) within a multi-turn conversation significantly degrades the overall post-conversation empathy score (linear mixed model $\beta = -0.657, p < 0.001$ , Cohen's $d = 1.142$ ).

While $45.9\%$ of turns across the 695 conversations were deemed applicable for empathy evaluation, only $12.7\%$ of turns contained at least one poor rating, and only $2.4\%$ had all seven dimensions rated poor. However, experiencing a poor turn showed substantial downstream effects on user engagement metrics:
1. Engagement Metric Correlations: Post-task perceived empathy correlated positively with perceived task success (Pearson $r = 0.619$ ), user conversational engagement ( $r = 0.632$ ), positive interaction experience ( $r = 0.657$ ), and willingness to reuse the agent for the same task ( $r = 0.602$ ).
2. Reuse Intention Penalty: The occurrence of a poor turn significantly diminished all four engagement variables. For participants aged 26–55, a poor turn reduced willingness to use the service again by $0.86$ to $0.97$ points on a 5-point Likert scale.
3. Dialogue Friction: Conversations containing a poor turn were significantly longer in total turns (mean $11.7$ vs. $9.5$ turns, $p < 0.001$ , $d = 0.362$ ) and exhibited longer user inputs per turn (mean $190.4$ vs. $140.8$ characters, $p = 0.002$ , $d = 0.276$ ), reflecting user attempts to correct or repair conversational breakdowns, while agent response length remained unchanged (mean $1359.0$ vs. $1372.0$ characters, $p = 0.812$ ).
Knowl 5 — User Importance Rankings Across Digital Empathy Dimensions
empirical result
Exit survey rankings by study participants evaluated the relative importance of the seven digital empathy dimensions (ranked 1 to 7, where 1 denotes highest importance):
1. Cognitive Understanding: Mean rank $= 2.34$ ; ranked as the #1 most critical factor by $28.1\%$ of participants ( $n = 54$ ).
2. Response Appropriateness: Mean rank $= 3.28$ ; ranked #1 by $19.3\%$ of participants ( $n = 37$ ).
3. Affective Understanding: Mean rank $= 3.47$ ; ranked #1 by $19.8\%$ of participants ( $n = 36$ ).
4. Contextual Understanding: Mean rank $= 4.05$ .
5. Relational Continuity: Mean rank $= 4.60$ .
6. Prosocial Expression: Mean rank $= 4.83$ .
7. Interest: Mean rank $= 5.43$ .
Participants identified Cognitive Understanding as a foundational prerequisite for all problem-solving scenarios, whereas Affective Understanding was prioritized primarily when addressing personal issues under high emotional distress. Response Appropriateness was valued over ungrounded emotional expression to prevent unsolicited advice or superficial sentiment.
Knowl 6 — Predictors of Conversational Topic Selection and Desired Empathy
empirical result
Generalized Estimating Equations (GEE) modeling revealed distinct user trait and contextual drivers of conversation topic choice and pre-task empathy expectations:
1. Personal Issues ( $30.4\%$ of conversations): Selection was positively predicted by higher interest in daily AI use ( $\beta = 0.27, p = 0.021, \text{Odds Ratio (OR)} \approx 1.31$ ) and higher trait empathy ( $\beta = 0.30, p = 0.042, \text{OR} \approx 1.35$ ), and negatively predicted by viewing organizational AI use as unethical ( $\beta = -0.19, p = 0.016, \text{OR} \approx 0.82$ ).
2. Work Issues ( $25.9\%$ of conversations): Selection was positively predicted by age 46–55 ( $\beta = 0.68, p = 0.047, \text{OR} \approx 1.96$ ), viewing organizational AI use as unethical ( $\beta = 0.49, p < 0.001, \text{OR} \approx 1.64$ ), and planning to use AI ( $\beta = 0.58, p = 0.002, \text{OR} \approx 1.78$ ), but negatively predicted by higher trait empathy ( $\beta = -0.58, p < 0.001, \text{OR} \approx 0.56$ ).
3. Desired Empathy Levels: Selecting a personal issue ( $\beta = 0.679, p < 0.001$ ) or a work issue ( $\beta = 0.408, p < 0.001$ ) elicited significantly higher pre-task desired empathy compared to information tasks ( $43.7\%$ of conversations).
4. Cognitive Reappraisal Moderation: While high cognitive reappraisal on the Emotion Regulation Questionnaire (ERQ) generally increased desired baseline empathy across tasks ( $\beta = 0.107, p = 0.037$ ), it exhibited significant negative interaction effects when handling personal issues ( $\beta = -0.151, p = 0.024$ ) or work issues ( $\beta = -0.160, p = 0.021$ ), indicating that individuals with stronger emotional reappraisal skills desire lower empathy and more concise, solution-oriented assistance during stressful tasks.
Knowl 7 — Empathy System Prompt Architecture for LLM Conversational Agents
model/method
To enhance LLM empathic performance, an empathy system prompt was designed around the seven-dimensional digital empathy taxonomy. The prompt architecture configures the LLM into an assistant grounded in counseling and motivational interviewing principles through three main sections:
1. Role Description: Explicitly tasks the agent with understanding and connecting with users on cognitive and emotional levels to deliver personalized, considerate interactions.
2. High-Level Goals:
  - Understand and relate: Deeply comprehend user emotions, perspectives, and experiences.
  - Provide appropriate responses: Align response tone and substance with user needs without unsolicited solution pushing.
  - Foster a supportive environment: Maintain a safe, non-judgmental dialogue space.
  - Encourage ongoing engagement: Maintain an evolving, continuous relationship leveraging historical context.
3. Reflective Self-Querying Pre-Conditions: Requires the model to internally reason over 22 specific self-reflection questions prior to generation (e.g., assessing emotional recognition accuracy, determining whether to offer advice vs. listen, checking if permission was asked before advising, exploring context/background, and incorporating prior conversational details) without outputting the prompts or questions directly to the user.
In single-blind evaluations, this prompting architecture applied to GPT-4 (GPT4-empathy) produced significantly higher perceived empathy than unprompted GPT-4 ( $\beta = 0.350, p = 0.015$ ).
Knowl 8 — LLM-Based Automated Perceived Empathy Classification Performance
empirical result
Automated zero-shot conversation-level perceived empathy recognition was evaluated using GPT-4o (API version 2024-10-21, $\text{temperature}=0.0, \text{top\_p}=1.0, \text{presence\_penalty}=0.0, \text{frequency\_penalty}=0.0$ ) across $N = 672$ anonymized multi-turn conversations over 10 repeated runs.

Ground-truth labels were computed by averaging the 7 dimension scores per turn, then averaging across all turns in a conversation on a continuous 1–5 scale, with discrete distribution: Very Poor ( $n=5$ ), Poor ( $n=15$ ), Neutral ( $n=48$ ), Good ( $n=290$ ), and Very Good ( $n=314$ ).
1. Continuous Metric Evaluation:
  - Mean Absolute Error (MAE): $0.551 \pm 0.007$
  - Spearman rank correlation ( $\rho$ ): $0.369 \pm 0.015$ ( $p < 0.001$ )
2. Discrete 5-Class Evaluation:
  - Accuracy: $0.487 \pm 0.011$ (majority class baseline $= 0.467$ , random baseline $= 0.200$ )
  - Macro Sensitivity: $0.286 \pm 0.038$
  - Macro Specificity: $0.827 \pm 0.008$
  - Macro F1: $0.272 \pm 0.041$
3. Cumulative Absolute Error Distribution $F(k) = P(|\hat{y} - y| \le k)$ :
  - Exact agreement $F(0)$ : $0.486 \pm 0.011$
  - Within-1 Level Accuracy $F(1)$ : $0.960 \pm 0.004$
  - Within-2 Level Accuracy $F(2)$ : $0.996 \pm 0.002$
The cumulative error distribution demonstrates that misclassifications predominantly occur at adjacent rating boundaries rather than large ordinal errors.
Knowl 9 — Two-Stage Privacy Anonymization Workflow for Conversational LLM Corpora
model/method
To preserve conversational semantics and empathy dynamics while mitigating privacy risks in open-ended workplace and personal dialogues, a two-stage anonymization workflow was implemented:
Input: Set of raw multi-turn human-AI conversations $C = \{c_1, c_2, \dots, c_N\}$
Output: Set of anonymized conversations $C_{\text{anon}}$
Initialize $C_{\text{anon}} \leftarrow \emptyset$
for each conversation $c \in C$ do
    // Stage 1: Automated LLM Sensitive Entity Masking
     $c' \leftarrow \text{PromptGuidedLLM}(c)$
    // Masks PII, names of secondary individuals, workplace identifiers,
    // employer-revealing details, and proprietary data using tokens
    // (e.g., [NAME], [COMPANY], [LOCATION]) while preserving wording and errors.

    // Stage 2: Dual Independent Author Review
     $R_1 \leftarrow \text{ReviewerA}(c', c)$
     $R_2 \leftarrow \text{ReviewerB}(c', c)$
    $c_{\text{final}} \leftarrow \text{ConsensusResolve}(R_1, R_2)

    if IsSufficientlyAnonymizedWithoutContextDegradation( $c_{\text{final}}$ ) then
         $C_{\text{anon}} \leftarrow C_{\text{anon}} \cup \{c_{\text{final}}\}$
    end if
end for
return $C_{\text{anon}}$
Applying this procedure to the 695 raw conversations resulted in the exclusion of 23 conversations ( $3.3\%$ ) that could not be fully scrubbed without severe context degradation, producing a final privacy-compliant research dataset of 672 conversations.
Knowl 10 — Limitations and Ethical Boundaries of SENSE-7 and Empathic Conversational AI
limitation
The methodology and findings of the SENSE-7 framework have several key limitations:
1. Sample Homogeneity: The study cohort was drawn from information workers at a single large technology firm with high digital literacy and positive baseline attitudes toward AI (mean $3.78$ out of $5$ ), potentially limiting generalizability across diverse cultural, socio-economic, or linguistic populations.
2. Temporal Model Evolution: Evaluations were benchmarked on fixed model checkpoints (GPT-4-32k version 0613, GPT-3.5-turbo, Llama2-70b-Chat) and static prompt configurations, which may not capture newer model capabilities or multi-modal affordances (e.g., voice tone, facial expressions, physiological signals).
3. Risk of Empathic Simulation and Dependency: AI empathy represents observable functional behavior rather than genuine internal emotional attunement. Over-simulating empathy in high-stakes domains (such as crisis counseling or mental health interventions) carries ethical risks of user deception, inappropriate anthropomorphism, and unhealthy emotional reliance on artificial systems.
4. Contextual Inappropriateness: High empathy is not universally desired across all conversational tasks; excessive emotional validation in technical, factual, or coding scenarios can be perceived as intrusive or unhelpful.

Coverage note — No substantial contributed material was omitted; all 7 dimensions of the AI Empathic Behavior Scale, the SENSE-7 dataset collection protocol and participant characteristics, regression models, turn-breakdown analysis, user dimensional rankings, system prompt architecture, anonymization workflow, automated classifier evaluations, and limitations were formalized as knowls.

References

1.M. H. Davis, “Measuring individual differences in empathy: Evidence for a multidimensional approach.” Journal of personality and social psychology, vol. 44, no. 1, p. 113, 1983.
2.B. M. Cuff, S. J. Brown, L. Taylor, and D. J. Howat, ‚Empathy: A review of the concept,” Emotion review, vol. 8, no. 2, pp. 144–153, 2016.
3.W. Ickes, ‚Empathic accuracy,” Journal of personality, vol. 61, no. 4, pp. 587–610, 1993.
4.J. Decety and W. Ickes, The social neuroscience of empathy. Mit press, 2011.
5.J. D. Wondra and P. C. Ellsworth, “An appraisal theory of empathy and other vicarious emotional experiences.” Psychological review, vol. 122, no. 3, p. 411, 2015.
6.R. W. Picard, Affective computing. MIT press, 2000.
7.J. W. Ayers, A. Poliak, M. Dredze, E. C. Leas, Z. Zhu, J. B. Kelley, D. J. Faix, A. M. Goodman, C. A. Longhurst, M. Hogarth et al., “Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum,” JAMA internal medicine, vol. 183, no. 6, pp. 589–596, 2023.
8.M. Rubin, J. Z. Li, F. Zimmerman, D. C. Ong, A. Goldenberg, and A. Perry, “Comparing the value of perceived human versus ai-generated empathy,” Nature Human Behaviour, 2025.
9.J. Z. Li, A. Herderich, and A. Goldenberg, “’Skill but not effort drive gpt overperformance over humans in cognitive reframing of negative scenarios,” Preprint at https://doi. org/10.31234/osf. io/fzvd8, 2024.
10.Y. K. Lee, J. Suh, H. Zhan, J. J. Li, and D. C. Ong, ‚‘Large language models produce responses perceived to be empathic,” arXiv preprint arXiv:2403.18148, 2024.
11.D. Ovsyannikova, V. O. de Mello, and M. Inzlicht, “Third-party evaluators perceive ai as more compassionate than expert humans,” Communications Psychology, vol. 3, no. 1, p. 4, 2025.
12.J. D. Wenger, D. Cameron, and M. Inzlicht, “The ai empathy choice paradox: People prefer human empathy despite rating ai empathy higher,” 2025.
13.Y. Yin, N. Jia, and C. J. Wakslak, “‘Ai can help people feel heard, but an ai label diminishes this impact,” Proceedings of the National Academy of Sciences, vol. 121, no. 14, p. e2319112121, 2024.
14.X. Wang, X. Li, Z. Yin, Y. Wu, and J. Liu, “’Emotional intelligence of large language models,” Journal of Pacific Rim Psychology, vol. 17, p. 18344909231213958, 2023.
15.V. Sorin, D. Brin, Y. Barash, E. Konen, A. Charney, G. Nadkarni, and E. Klang, “’Large language models and empathy: Systematic review,” Journal of Medical Internet Research, vol. 26, p. e52597, 2024.
16.A. Welivita and P. Pu, “‘Are large language models more empathetic than humans?” arXiv preprint arXiv:2406.05063, 2024.
17.J. S. Borg and H. Read, “‘What is required for empathic ai? it depends, and why that matters for ai developers and users,” in Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, vol. 7, 2024, pp. 1306–1318.
18.A. S. Raamkumar and Y. Yang, “’Empathetic conversational systems: A review of current advances, gaps, and opportunities,” IEEE Transactions on Affective Computing, vol. 14, no. 4, pp. 2722–2739, 2023.
19.H. Ma, B. Zhang, B. Xu, J. Wang, H. Lin, and X. Sun, “’Empathy level alignment via reinforcement learning for empathetic response generation,” IEEE Transactions on Affective Computing, pp. 1–12, 2025.
20.S. D. Preston and F. B. De Waal, “’Empathy: Its ultimate and proximate bases,” Behavioral and brain sciences, vol. 25, no. 1, pp. 1–20, 2002.
21.V. Scotti, R. Tedesco, and L. Sbattella, “’A modular data-driven architecture for empathetic conversational agents,” in 2021 IEEE International Conference on Big Data and Smart Computing (BigComp). IEEE, 2021, pp. 365–368.
22.A. Paiva, I. Leite, H. Boukricha, and I. Wachsmuth, “’Empathy in virtual agents and robots: A survey,” ACM Transactions on Interactive Intelligent Systems (TiiS), vol. 7, no. 3, pp. 1–40, 2017.
23.T. W. Bickmore, R. Fernando, L. Ring, and D. Schulman, “’Empathic touch by relational agents,” IEEE Transactions on Affective Computing, vol. 1, no. 1, pp. 60–71, 2010.
24.I. Leite, A. Pereira, S. Mascarenhas, C. Martinho, R. Prada, and A. Paiva, “The influence of empathy in human–robot relations,” International journal of human-computer studies, vol. 71, no. 3, pp. 250–260, 2013.
25.A. Tapus and M. J. Mataric, “‘Emulating empathy in socially assistive robotics.” in AAAI spring symposium: multidisciplinary collaboration for socially assistive robotics. California, 2007, pp. 93–96.
26.G. Castellano, A. Paiva, A. Kappas, R. Aylett, H. Hastie, W. Barendregt, F. Nabais, and S. Bull, “’Towards empathic virtual and robotic tutors,” in Artificial Intelligence in Education: 16th International Conference, AIED 2013, Memphis, TN, USA, July 9-13, 2013. Proceedings 16. Springer, 2013, pp. 733–736.
27.A. Sharma, I. W. Lin, A. S. Miner, D. C. Atkins, and T. Althoff, “‘Towards facilitating empathic conversations in online mental health support: A reinforcement learning approach,” in Proceedings of the Web Conference 2021, 2021, pp. 194–205.
28.K. Chawla, R. Clever, J. Ramirez, G. M. Lucas, and J. Gratch, “’Towards emotion-aware agents for negotiation dialogues,” 2021 9th International Conference on Affective Computing and Intelligent Interaction (ACII), pp. 1–8, 2021.
29.H. C. van Cuylenburg and T. N. D. S. Ginige, “‘Emotion guru: A smart emotion tracking application with ai conversational agent for exploring and preventing depression,” 2021 International Conference on UK-China Emerging Technologies (UCET), pp. 1–6, 2021.
30.A. N. Vaidyam, H. Wisniewski, J. D. Halamka, M. S. Kashavan, and J. B. Torous, “‘Chatbots and conversational agents in mental health: a review of the psychiatric landscape,” The Canadian Journal of Psychiatry, vol. 64, no. 7, pp. 456–464, 2019.
31.T. W. Bickmore, S. E. Mitchell, B. W. Jack, M. K. Paasche-Orlow, L. M. Pfeifer, and J. O’Donnell, “Response to a relational agent by hospital patients with depressive symptoms,” Interacting with computers, vol. 22, no. 4, pp. 289–298, 2010.
32.F. Efthymiou and C. Hildebrand, “‘Empathy by design: The influence of trembling ai voices on prosocial behavior,” IEEE Transactions on Affective Computing, vol. 15, no. 3, pp. 1253–1263, 2024.
33.J. Zaki and K. N. Ochsner, “The neuroscience of empathy: progress, pitfalls and promise,” Nature neuroscience, vol. 15, no. 5, pp. 675–680, 2012.
34.A. Perry, “’Ai will never convey the essence of human empathy,” Nature Human Behaviour, vol. 7, no. 11, pp. 1808–1809, 2023.
35.C. Montemayor, J. Halpern, and A. Fairweather, “In principle obstacles for empathic ai: why we can’t replace human empathy in healthcare,” AI & society, vol. 37, no. 4, pp. 1353–1359, 2022.
36.S. Concannon and M. Tomalin, “Measuring perceived empathy in dialogue systems,” Ai & Society, vol. 39, no. 5, pp. 2233–2247, 2024.
37.O. N. Yalçın, “’How (not) to evaluate computational empathy: Testing the assumptions of the evaluation methods in a use-case,” in 2023 11th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW), 2023, pp. 1–7.
38.S. Concannon, I. Roberts, and M. Tomalin, “’An interactional account of empathy in human-machine communication,” Human-Machine Communication, vol. 6, 2023.
39.M. Inzlicht, C. D. Cameron, J. D’Cruz, and P. Bloom, “In praise of empathic ai,” Trends in Cognitive Sciences, vol. 28, no. 2, pp. 89–91, 2024.
40.E. Hadjiandreou, T. Lau, D. C. Ong, A. Perry, and C. Cameron, “‘What llmpathy can tell us about received empathy,” 2025.
41.K. Harrelson, “’Intention and empathy,” Philosophical Psychology, vol. 33, no. 8, pp. 1162–1184, 2020.
42.U. Genç and H. Verma, “’Situating empathy in hci/cscw: A scoping review,” Proceedings of the ACM on Human-Computer Interaction, vol. 8, no. CSCW2, pp. 1–37, 2024.
43.J. Decety and P. L. Jackson, “The functional architecture of human empathy,” Behavioral and cognitive neuroscience reviews, vol. 3, no. 2, pp. 71–100, 2004.
44.C. D. Batson, “These things called empathy: Eight related but distinct phenomena,” in The Social Neuroscience of Empathy. The MIT Press, 03 2009. [Online]. Available: https://doi.org/10.7551/mitpress/9780262012973.003.0002
45.E. Morgante, C. Susinna, L. Culicetto, A. Quartarone, and V. Lo Buono, “Is it possible for people to develop a sense of empathy toward humanoid robots and establish meaningful relationships with them?” Frontiers in Psychology, vol. 15, p. 1391832, 2024.
46.J. H. Eklund and M. S. Meranius, “‘Toward a consensus on the nature of empathy: A review of reviews,” Patient Education and Counseling, vol. 104, no. 2, pp. 300–307, 2021.
47.J. Zaki, “‘Empathy: a motivated account.” Psychological Bulletin, vol. 140, no. 6, p. 1608, 2014.
48.G. J. Depow, Z. Francis, and M. Inzlicht, “The experience of empathy in everyday life,” Psychological Science, vol. 32, no. 8, pp. 1198–1213, 2021.
49.M. Guthridge and M. J. Giummarra, “’The taxonomy of empathy: a metadefinition and the nine dimensions of the empathic system,” Journal of Humanistic Psychology, p. 00221678211018015, 2021.
50.R. N. Spreng*, M. C. McKinnon*, R. A. Mar, and B. Levine, “The toronto empathy questionnaire: Scale development and initial validation of a factor-analytic solution to multiple empathy measures,” Journal of personality assessment, vol. 91, no. 1, pp. 62–71, 2009.
51.D. D. Vachon and D. R. Lynam, “’Fixing the problem with empathy: Development and validation of the affective and cognitive measure of empathy,” Assessment, vol. 23, no. 2, pp. 135–149, 2016.
52.S. E. Decker, C. Nich, K. M. Carroll, and S. Martino, “Development of the therapist empathy scale,” Behavioural and cognitive psychotherapy, vol. 42, no. 3, pp. 339–354, 2014.
53.R. L. Reniers, R. Corcoran, R. Drake, N. M. Shryane, and B. A. Völlm, “The qcae: A questionnaire of cognitive and affective empathy,” Journal of personality assessment, vol. 93, no. 1, pp. 84–95, 2011.
54.K. E. Gerdes, E. A. Segal, and C. A. Lietz, “’Conceptualising and measuring empathy,” British Journal of Social Work, vol. 40, no. 7, pp. 2326–2343, 2010.
55.C. Breazeal, “‘Toward sociable robots,” Robotics and autonomous systems, vol. 42, no. 3-4, pp. 167–175, 2003.
56.A. Paiva, F. Correia, R. Oliveira, F. Santos, and P. Arriaga, “’Empathy and prosociality in social agents,” in The handbook on socially interactive agents: 20 Years of research on embodied conversational agents, intelligent virtual agents, and social robotics volume 1: methods, behavior, cognition, 2021, pp. 385–432.
57.S. Brave, C. Nass, and K. Hutchinson, “’Computers that care: investigating the effects of orientation of emotion exhibited by an embodied computer agent,” International journal of human-computer studies, vol. 62, no. 2, pp. 161–178, 2005.
58.T. Liu, S. Giorgi, A. Aich, A. Lahnala, B. Curtis, L. Ungar, and J. Sedoc, “’The illusion of empathy: How ai chatbots shape conversation perception,” arXiv preprint arXiv:2411.12877, 2024.
59.N. Eisenberg and R. Lennon, “’Sex differences in empathy and related capacities.” Psychological bulletin, vol. 94, no. 1, p. 100, 1983.
60.A. L. Guzman, “’Making ai safe for humans: A conversation with siri,” in Socialbots and their friends. Routledge, 2016, pp. 85–101.
61.H. Rashkin, E. M. Smith, M. Li, and Y.-L. Boureau, “‘Towards empathetic open-domain conversation models: A new benchmark and dataset,” in Annual Meeting of the Association for Computational Linguistics, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID:195069365
62.A. Sharma, A. S. Miner, D. C. Atkins, and T. Althoff, “‘A computational approach to understanding empathy expressed in text-based mental health support,” arXiv preprint arXiv:2009.08441, 2020.
63.Z. Xu and J. Jiang, “‘Multi-dimensional evaluation of empathetic dialog responses,” in Conference on Empirical Methods in Natural Language Processing, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:267751205
64.C. Ziems, W. Held, O. Shaikh, J. Chen, Z. Zhang, and D. Yang, “’Can large language models transform computational social science?” Computational Linguistics, vol. 50, no. 1, pp. 237–291, 2024.
65.A. Welivita and P. Pu, “‘A taxonomy of empathetic response intents in human social conversations,” in International Conference on Computational Linguistics, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:227230442
66.E. Svikhnushina, I. Voinea, A. Welivita, and P. Pu, “’A taxonomy of empathetic questions in social dialogs,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 2952–2973.
67.A. Kumar, N. Poungpeth, D. Yang, E. Farrell, B. Lambert, and M. Groh, “When large language models are reliable for judging empathic communication,” arXiv preprint arXiv:2506.10150, 2025.
68.S. Liu, C. Zheng, O. Demasi, S. Sabour, Y. Li, Z. Yu, Y. Jiang, and M. Huang, “’Towards emotional support dialog systems,” arXiv preprint arXiv:2106.01144, 2021.
69.J. Shen, D. DiPaola, S. Ali, M. Sap, H. W. Park, C. Breazeal et al., “‘Empathy toward artificial intelligence versus human experiences and the role of transparency in mental health and social support chatbot design: Comparative study,” JMIR Mental Health, vol. 11, no. 1, p. e62679, 2024.
70.Q. Zhu and J. Luo, “‘Toward artificial empathy for human-centered design: A framework,” ArXiv, vol. abs/2303.10583, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:257631572
71.C. A. Lietz, K. E. Gerdes, F. Sun, J. M. Geiger, M. A. Wagaman, and E. A. Segal, “‘The empathy assessment index (eai): A confirmatory factor analysis of a multidimensional model of empathy,” Journal of the Society for Social Work and Research, vol. 2, no. 2, pp. 104–124, 2011.
72.O. N. Yalcin and S. DiPaola, “’A computational model of empathy for interactive agents,” Biologically inspired cognitive architectures, vol. 26, pp. 20–25, 2018.
73.S. Fitrianie, M. Bruijnes, F. Li, A. Abdulrahman, and W.-P. Brinkman, “The artificial-social-agent questionnaire: establishing the long and short questionnaire versions,” in Proceedings of the 22nd ACM International Conference on Intelligent Virtual Agents, 2022, pp. 1–8.
74.A. Main, E. A. Walle, C. Kho, and J. Halpern, “The interpersonal functions of empathy: A relational perspective,” Emotion Review, vol. 9, no. 4, pp. 358–366, 2017.
75.J. A. Hall, R. Schwartz, and F. Duong, “‘How do laypeople define empathy?” The Journal of Social Psychology, vol. 161, no. 1, pp. 5–24, 2021.
76.J. L. Beredo and E. C. Ong, “‘A hybrid response generation model for an empathetic conversational agent,” in 2022 International Conference on Asian Language Processing (IALP). IEEE, 2022, pp. 300–305.
77.J. L. Krupnick, S. M. Sotsky, I. Elkin, S. Simmens, J. Moyer, J. Watkins, and P. A. Pilkonis, “’The role of the therapeutic alliance in psychotherapy and pharmacotherapy outcome: Findings in the national institute of mental health treatment of depression collaborative research program,” Focus, vol. 64, no. 2, pp. 532–277, 2006.
78.B. J. Kalisch, “‘What is empathy?” The American journal of nursing, pp. 1548–1552, 1973.
79.H. Kawamichi, K. Yoshihara, A. T. Sasaki, S. K. Sugawara, H. C. Tanabe, R. Shinohara, Y. Sugisawa, K. Tokutake, Y. Mochizuki, T. Anme et al., “‘Perceiving active listening activates the reward system and improves the impression of relevant experiences,” Social neuroscience, vol. 10, no. 1, pp. 16–26, 2015.
80.A. Carré, N. Stefaniak, F. d’Ambrosio, L. Bensalah, and C. Besche-Richard, “The basic empathy scale in adults (bes-a): factor structure of a revised form.” Psychological assessment, vol. 25, no. 3, p. 679, 2013.
81.S. M. Jones, “Supportive listening,” The Intl. Journal of Listening, vol. 25, no. 1-2, pp. 85–103, 2011.
82.J. Y. Jang, S. Shin, and G. Gweon, “’Minimal yet big impact: How ai agent back-channeling enhances conversational engagement through conversation persistence and context richness,” in Findings of the Association for Computational Linguistics: EMNLP 2024, 2024, pp. 14 509–14 521.
83.Y. Ma, K. L. Nguyen, F. Z. Xing, and E. Cambria, “‘A survey on empathetic dialogue systems,” Information Fusion, vol. 64, pp. 50–70, 2020.
84.T. Bickmore, D. Schulman, and L. Yin, “‘Maintaining engagement in long-term interventions with relational agents,” Applied Artificial Intelligence, vol. 24, no. 6, pp. 648–666, 2010.
85.J. Zaki, N. Bolger, and K. Ochsner, “‘It takes two: The interpersonal nature of empathic accuracy,” Psychological science, vol. 19, no. 4, pp. 399–404, 2008.
86.N. M. Thompson, A. Uusberg, J. J. Gross, and B. Chakrabarti, “‘Empathy and emotion regulation: An integrative account,” Progress in brain research, vol. 247, pp. 273–304, 2019.
87.Y. Wang, Y. Wang, K. Crace, and Y. Zhang, “‘Understanding attitudes and trust of generative ai chatbots for social anxiety support,” arXiv preprint arXiv:2501.15628, 2025.
88.H. Cramer, J. Goddijn, B. Wielinga, and V. Evers, “‘Effects of (in) accurate empathy and situational valence on attitudes towards robots,” in 2010 5th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE, 2010, pp. 141–142.
89.S. Konrath, B. P. Meier, and B. J. Bushman, “Development and validation of the single item trait empathy scale (sites),” Journal of research in personality, vol. 73, pp. 111–122, 2018.
90.J. J. Gross and O. P. John, “‘Individual differences in two emotion regulation processes: implications for affect, relationships, and wellbeing.” Journal of personality and social psychology, vol. 85, no. 2, p. 348, 2003.
91.S. Grassini, “’Development and validation of the ai attitude scale (aias4): a brief measure of general attitude toward artificial intelligence,” Frontiers in psychology, vol. 14, p. 1191628, 2023.
92.A. Schepman and P. Rodway, “‘Initial validation of the general attitudes towards artificial intelligence scale,” Computers in human behavior reports, vol. 1, p. 100014, 2020.
93.E. R. Thompson, “Development and validation of an internationally reliable short-form of the positive and negative affect schedule (panas),” Journal of cross-cultural psychology, vol. 38, no. 2, pp. 227–242, 2007.
94.S. B. Merriam et al., “‘Introduction to qualitative research,” Qualitative research in practice: Examples for discussion and analysis, vol. 1, no. 1, pp. 1–17, 2002.
95.Adobe, “’How chatgpt is changing the way we search,” https://www.adobe.com/express/learn/blog/chatgpt-as-a-search-engine, 7 2025, accessed: 2025-08-20.
96.J. Hernandez, J. Suh, J. Amores, K. Rowan, G. Ramos, and M. Czerwinski, “’Affective conversational agents: Understanding expectations and personal influences,” arXiv preprint arXiv:2310.12459, 2023.
97.A. Bhattacharjee, J. Suh, M. Ershadi, S. T. Iqbal, A. D. Wilson, and J. Hernandez, “‘Understanding communication preferences of information workers in engagement with text-based conversational agents,” arXiv preprint arXiv:2410.20468, 2024.

Citation

MLA

Suh, J., et al. “SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations”. IEEE Transactions on Affective Computing, vol. 17, no. 2, 2026, pp. 2227–44, https://doi.org/10.1109/taffc.2026.3656531.

APA

Suh, J., Le, L., Shayegani, E., Ramos, G., Amores, J., Ong, D. C., Czerwinski, M., & Hernandez, J. (2026). SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations. IEEE Transactions on Affective Computing, 17(2), 2227–2244. https://doi.org/10.1109/taffc.2026.3656531

Chicago

Suh, J., L. Le, E. Shayegani, et al. 2026. “SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations”. IEEE Transactions on Affective Computing 17 (2): 2227–44. https://doi.org/10.1109/taffc.2026.3656531.

Harvard

Suh, J. et al. (2026) “SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations”, IEEE Transactions on Affective Computing, 17(2), pp. 2227–2244. Available at: https://doi.org/10.1109/taffc.2026.3656531.

Vancouver

1. Suh J, Le L, Shayegani E, Ramos G, Amores J, Ong DC, Czerwinski M, Hernandez J (2026) SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations. IEEE Transactions on Affective Computing 17:2227–2244

BibTeX

@article{Suh_2026, title={SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations}, volume={17}, ISSN={2371-9850}, url={http://dx.doi.org/10.1109/taffc.2026.3656531}, DOI={10.1109/taffc.2026.3656531}, number={2}, journal={IEEE Transactions on Affective Computing}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Suh, Jina and Le, Lindy and Shayegani, Erfan and Ramos, Gonzalo and Amores, Judith and Ong, Desmond C. and Czerwinski, Mary and Hernandez, Javier}, year={2026}, month=Apr, pages={2227–2244} }

Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

SENSE-7: Taxonomy and Dataset for Measuring User Perceptions of Empathy in Sustained Human-AI Conversations

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Seven-Dimensional Digital Empathy Taxonomy

Knowl 2 — SENSE-7 Dataset and Multi-Turn Annotation Protocol

Knowl 3 — Post-Task Overall Perceived Empathy Linear Mixed-Effects Model

Knowl 4 — Disproportionate Impact of Single 'Poor Turns' on Perceived Empathy and User Engagement

Knowl 5 — User Importance Rankings Across Digital Empathy Dimensions

Knowl 6 — Predictors of Conversational Topic Selection and Desired Empathy

Knowl 7 — Empathy System Prompt Architecture for LLM Conversational Agents

Knowl 8 — LLM-Based Automated Perceived Empathy Classification Performance

Knowl 9 — Two-Stage Privacy Anonymization Workflow for Conversational LLM Corpora

Knowl 10 — Limitations and Ethical Boundaries of SENSE-7 and Empathic Conversational AI

References

Citation

BibTeX

Source Code

Access the Paper