Pith. sign in

REVIEW 4 major objections 5 minor 41 references

To Embody or Not: The Effect Of Embodiment On User Perception Of LLM-based Conversational Agents

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Embodied AI chatbots rated less competent than text-only

desk verdict An honest, well-motivated pilot whose central competence effect is uninterpretable because embodiment is perfectly confounded with the survival scenario; the sycophancy angle deserves a controlled follow-up. read the letter →

arxiv 2506.02514 v1 pith:EZOV6WKC submitted 2025-06-03 cs.HC

Kyra Wang , Boon-Kiat Quek , Jessica Goh , Dorien Herremans This is my paper
classification cs.HC
keywords embodimentconversationalagentslargelanguagemodelssycophancyuserperceptioncredibilityhuman-agentcooperationwithin-subjectsstudy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that giving an LLM-based conversational agent a realistic human body can reduce, not raise, its perceived competence in cooperative tasks. In a within-subjects experiment, 20 participants solved desert and tundra survival scenarios with an embodied and a text-only agent powered by the same LLaMA model, then rated each on six credibility dimensions. The text-only agent was rated significantly more competent (p = 0.01), and participants more often described the embodied agent as sycophantic. The authors conclude that when an LLM tends to flatter users, embodiment makes that flattery read as inauthentic, so embodiment is not a reliable way to build trust.

What carries the argument

The experimental machinery is a pair of locally hosted agents built on LLaMA 3.1 8B Instruct: one is rendered as a realistic digital human (MetaHuman) avatar in Unreal Engine with speech and facial animation, the other is plain text. Both agents work with participants on survival-ranking tasks, and participants rate them on six credibility dimensions — competence, character, sociability, dynamism, dominance, and submissiveness — adapted from established semantic-differential scales. The conceptual load-bearing identity is the authors' theory that anthropomorphic embodiment amplifies the negative effect of LLM sycophancy on perceived authenticity, reversing the usual positive effect of embodiment on credibility.

What would settle it

Run the same within-subjects comparison with scenario and embodiment fully counterbalanced, and also add a condition in which the LLM is prompted to disagree and defend its views; if the competence gap persists only when the desert agent is embodied, the scenario is the cause, and if it disappears when sycophancy is removed, the sycophancy account is confirmed.

Watch

Extended reading notes

Core claim

The central claim is that, contrary to the usual assumption that embodiment improves conversational-agent credibility, a non-embodied LLM agent is perceived as significantly more competent than an embodied one in a non-hierarchical cooperative task. The paper reports a significant competence difference (p = 0.01) and notes that participants were more likely to label the embodied agent sycophantic, even though both agents used the same underlying model and nearly identical prompts. The proposed explanation is that embodiment invites users to judge the agent as a social actor, so the model's tendency to agree and praise becomes visible as ingratiation and lowers perceived authenticity. The authors generalize this to the claim that embodiment is not a straightforward trust-building feature for LLM-based agents prone to sycophancy.

Load-bearing premise

The two survival scenarios are assumed to be equally difficult and equally familiar to participants, even though the desert scenario was always paired with the embodied agent and the tundra scenario with the text-only one.

Editorial extensions

If this is right

  • Designers of LLM-based conversational agents cannot assume that adding a human-like body improves trust; in cooperative tasks it may lower perceived competence.
  • Embodied agents may need explicit anti-sycophancy prompting or training before embodiment can deliver its usual credibility benefits.
  • User studies of embodiment should measure perceived sycophancy and authenticity, not only overall credibility scores, because these perceptions can diverge sharply.
  • Because the embodied agent was also gendered masculine, future work would need to disentangle embodiment from gender before generalizing the effect.
  • The same underlying model can be evaluated differently purely due to presentation, so product teams should test embodiment decisions empirically rather than rely on intuition.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A decisive test would counterbalance the scenario-embodiment pairing; if the competence gap tracks the desert scenario rather than the body, the central effect is a scenario artifact, not an embodiment effect.
  • The theory predicts a crossover interaction: raising the LLM's willingness to push back (lowering sycophancy) should make the embodied agent regain or exceed the text-only agent's competence ratings.
  • The same mechanism may extend to other anthropomorphic cues such as voice, profile photos, or gendered personas, with any cue that raises social expectations making sycophancy more jarring.
  • A larger replication with balanced scenario pairing and a non-sycophantic condition would tell whether the reversal is specific to LLaMA 3.1 or generalizes across LLM architectures.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This manuscript reports a mixed-methods, within-subjects study (n=20) comparing user perceptions of an embodied versus a non-embodied LLM-based conversational agent in cooperative survival-ranking tasks. Participants first completed a survival scenario alone and then with the assigned agent, followed by quantitative credibility ratings across six dimensions and open-ended qualitative feedback. The authors report that the non-embodied agent was rated significantly more competent than the embodied agent (p=0.01), and that qualitative comments suggested the embodied agent was perceived as more sycophantic. They theorize that embodiment amplifies the negative credibility effect of LLM sycophancy, concluding that embodiment is not a straightforward way to improve perceived credibility when sycophancy is present.

Significance. If the central claim were supported, the paper would offer a timely and counterintuitive contribution to HCI research on embodied conversational agents, challenging the common assumption that embodiment improves user outcomes. The study has several strengths: both conditions use the same underlying LLM (LLaMA 3.1 8B Instruct), the setup runs locally with low latency to reduce a known confound, the design includes both quantitative and qualitative measures, and the authors make their testing software available on GitHub. However, the central empirical claim rests on a design in which embodiment is perfectly confounded with the survival scenario, and the statistical analysis does not correct for multiple comparisons. As a result, the main conclusion is not currently supported by the data.

major comments (4)
  1. [Section 2.2 and Section 3.1] The central empirical claim that the non-embodied CA is perceived as significantly more competent (p=0.01) is uninterpretable as an embodiment effect because embodiment is perfectly confounded with the survival scenario: the embodied condition always used the desert problem and the non-embodied condition always used the tundra problem. Counterbalancing the order of the two fixed conditions does not break this confound; it only randomizes which confounded condition is experienced first. The authors themselves acknowledge in Section 4.2 that Singaporean participants may be more competent in hot-weather survival, which could independently lower perceived competence of the desert-paired embodied agent. Since the prompts also differ in item lists and setting, the observed difference could reflect scenario difficulty, scenario-specific participant expertise, or scenario-specific LLM behavior rather than embodiment.
  2. [Section 3.1] The statistical analysis performs multiple Wilcoxon signed-rank tests across six credibility dimensions and multiple subscales without any correction for multiple comparisons. With n=20, the reported p=0.01 for competence would not survive a Bonferroni correction for the six dimensions (which would require p<0.0083), and the marginal p=0.06 for sociability is far from significant after correction. In addition, the MANCOVA with OCEAN and AI-attitude covariates is severely underpowered at this sample size, so the null result for personality and attitude covariates should not be interpreted as strong evidence that these factors had no influence.
  3. [Section 3.2 and Section 4.1] The sycophancy explanation is inferred from qualitative comments (six participants complained about the embodied agent not pushing back versus two for the non-embodied agent) rather than from a direct behavioral measure of sycophancy or pushback. This interpretation is in tension with the quantitative result that the dominance and submissiveness dimensions show no significant difference between conditions (p=0.87 and p=0.79, reported in Section 3.1). Moreover, because the qualitative examples are drawn from different scenarios, the comments cannot separate an embodiment-driven sycophancy perception from a scenario-driven one. The paper needs a direct measure of sycophantic behavior (e.g., coding agent responses for agreement, flattery, or opinion reversal) before this mechanism can be claimed.
  4. [Section 3.3] The behavioral result that messages sent to the non-embodied agent had significantly higher sentiment (p=0.01) is explicitly acknowledged by the authors as potentially an artifact of the different scenarios. Since scenario is confounded with embodiment, this measure also cannot support an embodiment-specific effect. The other conversational metrics (time, length, grammatical correctness) show no significant differences, which further limits the behavioral evidence for RQ3.
minor comments (5)
  1. [Table 1] The table lists a single mean value per condition for each dimension, but each dimension comprises multiple adjective subscales; please clarify whether the reported value is the mean across subscales and include standard deviations or confidence intervals so readers can assess variability.
  2. [Section 3.1] There is a typo in 'MANCOV A' that should read 'MANCOVA'.
  3. [Figure 3] The figure panels are labeled (a) through (f) but the caption does not name the dimensions shown; please add the dimension names to the caption and label the axes with the 1–7 scale.
  4. [Section 2.4.3] The participant age range is reported as '20 to 30'; please also report the mean and standard deviation of age for transparency.
  5. [References] Several references have formatting problems, including a garbled author list for [27], a DOI in [1] that appears unrelated to the cited work, and inconsistent capitalization; please verify and clean up the reference list.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning: the competence finding comes from direct participant ratings, and the sycophancy explanation is imported from external prior work rather than being fitted from the dependent variable.

full rationale

The paper's derivation chain is empirical rather than definitional. The central quantitative claim is a direct comparison of participant ratings between two conditions, analyzed with Wilcoxon signed-rank tests, and is not generated by fitting a parameter that is then relabeled as a prediction. The qualitative sycophancy theme is interpreted using prior external studies of LLM sycophancy and user trust, so the explanation is imported evidence, not a construct defined in terms of the outcome. No load-bearing self-citation chain appears: the authors do cite prior work, but the cited sycophancy results are independent external findings, and the GitHub link is merely a software release rather than evidence for the claim. The acknowledged desert-versus-tundra confound in Section 4.2 is a genuine threat to internal validity, but it is not circularity: it proposes an alternative cause for the same observed difference, rather than making the conclusion true by construction. The paper even notes that the dominance and submissiveness dimensions showed no significant difference, which weakens the sycophancy interpretation but does not make the reasoning circular. Overall, the claim is self-contained against the data and the cited external literature, so the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

No free parameters are fitted. The design relies on scenario equivalence and on a postulated sycophancy-embodiment mechanism; neither is independently verified in the study.

assumptions (2)
  • domain assumption The desert and tundra survival scenarios are equivalent in difficulty and in participants' ability.
    The embodied condition always used the desert problem and the non-embodied condition always used the tundra problem, so scenario is confounded with embodiment (Section 2.2).
  • ad hoc to paper LLM sycophancy is a stable trait that interacts with embodiment to reduce perceived authenticity.
    The paper does not measure sycophancy directly; it infers this interaction from qualitative comments and prior work (Sections 4.1, 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of To Embody or Not: The Effect Of Embodiment On User Perception Of LLM-based Conversational Agents." pith.science (2026). https://pith.science/paper/EZOV6WKC

@misc{pith2026250602514,
  author       = {Pith},
  title        = {Pith review of: To Embody or Not: The Effect Of Embodiment On User Perception Of LLM-based Conversational Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EZOV6WKC}},
  note         = {Machine review of arXiv:2506.02514}
}
read the original abstract

Embodiment in conversational agents (CAs) refers to the physical or visual representation of these agents, which can significantly influence user perception and interaction. Limited work has been done examining the effect of embodiment on the perception of CAs utilizing modern large language models (LLMs) in non-hierarchical cooperative tasks, a common use case of CAs as more powerful models become widely available for general use. To bridge this research gap, we conducted a mixed-methods within-subjects study on how users perceive LLM-based CAs in cooperative tasks when embodied and non-embodied. The results show that the non-embodied agent received significantly better quantitative appraisals for competence than the embodied agent, and in qualitative feedback, many participants believed that the embodied CA was more sycophantic than the non-embodied CA. Building on prior work on users' perceptions of LLM sycophancy and anthropomorphic features, we theorize that the typically-positive impact of embodiment on perception of CA credibility can become detrimental in the presence of sycophancy. The implication of such a phenomenon is that, contrary to intuition and existing literature, embodiment is not a straightforward way to improve a CA's perceived credibility if there exists a tendency to sycophancy.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 30 canonical work pages

  1. [1]

    In Research- Gate, 2021

    MetaHuman Creator The starting point of the metaverse. In Research- Gate, 2021. doi: 10.1109/ISCTIS51085.2021.00040

    work page arXiv 2021

  2. [2]

    ResearchGate, Feb

    (PDF) Everyone Talks Everything With ChatGPT:. ResearchGate, Feb

    work page

  3. [3]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 tech- nical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  4. [4]

    Allouch, A

    M. Allouch, A. Azaria, and R. Azoulay. Conversational Agents: Goals, Technologies, Vision and Challenges. Sensors (Basel, Switzerland), 21 (24):8448, Dec. 2021. ISSN 1424-8220. doi: 10.3390/s21248448

    work page doi:10.3390/s21248448 2021

  5. [5]

    S. A. Anisha, A. Sen, and C. Bain. Evaluating the Potential and Pitfalls of AI-Powered Conversational Agents as Humanlike Virtual Health Carers in the Remote Management of Noncommunicable Diseases: Scoping Review.Journal of Medical Internet Research, 26:e56114, July

    work page

  6. [6]

    Barbieri, J

    F. Barbieri, J. Camacho-Collados, L. Espinosa Anke, and L. Neves. TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification. In T. Cohn, Y . He, and Y . Liu, editors, Findings of the Association for Computational Linguistics: EMNLP 2020 , pages 1644– 1650, Online, Nov. 2020. Association for Computational Linguistics. doi: 10.18653/v...

    doi:10.18653/v1/2020.findings-emnlp.148 2020

  7. [7]

    J. K. Burgoon, T. Birk, and M. Pfau. Nonverbal Behaviors, Persuasion, and Credibility. Human Communication Research , 17(1):140–169,

    work page

  8. [8]

    J. K. Burgoon, J. A. Bonito, B. Bengtsson, C. Cederberg, M. Lunde- berg, and L. Allspach. Interactivity in human–computer interaction: A study of credibility, understanding, and influence. Computers in Human Behavior , 16(6):553–574, Nov. 2000. ISSN 0747-5632. doi: 10.1016/S0747-5632(00)00029-7

    work page doi:10.1016/s0747-5632(00)00029-7

Show all 41 references
  1. [9]

    Choudhury and H

    A. Choudhury and H. Shamszare. Investigating the Impact of User Trust on the Adoption and Use of ChatGPT: Survey Analysis. Jour- nal of Medical Internet Research , 25(1):e47184, June 2023. doi: 10.2196/47184

    2023 doi

  2. [10]

    P. F. Christiano, J. Leike, T. Brown, M. Martic, S. Legg, and D. Amodei. Deep Reinforcement Learning from Human Preferences

  3. [11]

    Diederich, A

    S. Diederich, A. Brendel, and L. Kolbe. On Conversational Agents in Information Systems Research: Analyzing the Past to Guide Future Work. Wirtschaftsinformatik 2019 Proceedings, Feb. 2019

    2019

  4. [12]

    Diederich, A

    S. Diederich, A. B. Brendel, S. Lichtenberg, and L. Kolbe. Design for fast request fulfillment or natural interaction? insights from an experi- ment with a conversational agent. Research Papers, May 2019

    2019

  5. [13]

    Go and S

    E. Go and S. S. Sundar. Humanizing chatbots: The effects of visual, identity and conversational cues on humanness perceptions. Computers in Human Behavior , 97:304–316, Aug. 2019. ISSN 0747-5632. doi: 10.1016/j.chb.2019.01.020

    2019 doi

  6. [14]

    Jiang, X

    Z. Jiang, X. Huang, Z. Wang, Y . Liu, L. Huang, and X. Luo. Embodied Conversational Agents for Chronic Diseases: Scoping Review. Journal of Medical Internet Research , 26:e47134, Jan. 2024. ISSN 1439-4456. doi: 10.2196/47134

    2024 doi

  7. [15]

    D. W. Johnson and F. P. Johnson. Joining Together: Group Theory and Group Skills, 4th Ed. Joining Together: Group Theory and Group Skills, 4th Ed. Prentice-Hall, Inc, Englewood Cliffs, NJ, US, 1991. ISBN 978- 0-13-511858-0

    1991

  8. [16]

    Karras, T

    T. Karras, T. Aila, S. Laine, A. Herva, and J. Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion.ACM Transactions on Graphics , 36(4):1–12, Aug. 2017. ISSN 0730-0301, 1557-7368. doi: 10.1145/3072959.3073658

    2017

  9. [17]

    J. Kim, J. Kong, and J. Son. Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech, June 2021

    2021

  10. [18]

    Kulms and S

    P. Kulms and S. Kopp. More Human-Likeness, More Trust? The Effect of Anthropomorphism on Self-Reported and Behavioral Trust in Con- tinued and Interdependent Human-Agent Cooperation. In Proceedings of Mensch Und Computer 2019, MuC ’19, pages 31–42, New York, NY , USA, Sept. 20...

    2019

  11. [19]

    Lafferty, P

    J. Lafferty, P. Eady, A. Pond, and H. Synergistics. The Desert Sur- vival Problem: A Group Decision Making Experience for Examining and Increasing Individual and Team Effectiveness: Manual . Experi- mental Learning Methods, 1974

    1974

  12. [20]

    F. R. Lang, D. John, O. Lüdtke, J. Schupp, and G. G. Wagner. Short assessment of the Big Five: Robust across survey methods except tele- phone interviewing. Behavior Research Methods, 43(2):548–567, 2011. ISSN 1554-351X. doi: 10.3758/s13428-011-0066-z

    2011 doi

  13. [21]

    Lew and J

    Z. Lew and J. B. Walther. Social Scripts and Expectancy Violations: Evaluating Communication with Human or AI Chatbot Interactants. Media Psychology , 26(1):1–16, Jan. 2023. ISSN 1521-3269, 1532- 785X. doi: 10.1080/15213269.2022.2084111

    2023

  14. [22]

    Y . Li, H. Wen, W. Wang, X. Li, Y . Yuan, G. Liu, J. Liu, W. Xu, X. Wang, Y . Sun, R. Kong, Y . Wang, H. Geng, J. Luan, X. Jin, Z. Ye, G. Xiong, F. Zhang, X. Li, M. Xu, Z. Li, P. Li, Y . Liu, Y .-Q. Zhang, and Y . Liu. Personal LLM Agents: Insights and Survey about the Capabil...

    2024

  15. [23]

    S. Lim, R. Schmälzle, and G. Bente. Artificial social influence via human-embodied AI agent interaction in immersive virtual reality (VR): Effects of similarity-matching during health conversations, Sept. 2024

    2024

  16. [24]

    D. Naber. A Rule-Based Style and Grammar Checker

  17. [25]

    M. S. Park, P. B. Upama, A. A. Anik, S. I. Ahamed, J. Luo, S. Tian, M. Rabbani, and H. Oh. A Survey of Conversational Agents and Their Applications for Self-Management of Chronic Conditions. Pro- ceedings : Annual International Computer Software and Applications Conference. CO...

    2023

  18. [26]

    Perez, S

    E. Perez, S. Ringer, K. Lukoši ¯ut˙e, K. Nguyen, E. Chen, S. Heiner, C. Pettit, C. Olsson, S. Kundu, S. Kadavath, A. Jones, A. Chen, B. Mann, B. Israel, B. Seethor, C. McKinnon, C. Olah, D. Yan, D. Amodei, D. Amodei, D. Drain, D. Li, E. Tran-Johnson, G. Khun- dadze, J. Kernion...

    2022

  19. [27]

    M. Rheu, S. , Ji Youn, P. , Wei, and J. and Huh-Yoo. Systematic Review: Trust-Building Factors and Implications for Conversational Agent De- sign. International Journal of Human–Computer Interaction, 37(1):81– 96, Jan. 2021. ISSN 1044-7318. doi: 10.1080/10447318.2020.1807710

    2021

  20. [28]

    D. A. Robb, J. Lopes, M. I. Ahmad, P. E. McKenna, X. Liu, K. Lohan, and H. Hastie. Seeing eye to eye: Trustworthy embodiment for task- based conversational agents. Frontiers in Robotics and AI, 10:1234767, Aug. 2023. ISSN 2296-9144. doi: 10.3389/frobt.2023.1234767

    2023

  21. [29]

    Schepman and P

    A. Schepman and P. Rodway. Initial validation of the general attitudes towards Artificial Intelligence Scale. Computers in Human Behavior Reports, 1:100014, Jan. 2020. ISSN 2451-9588. doi: 10.1016/j.chbr. 2020.100014

    2020

  22. [30]

    Seymour, K

    M. Seymour, K. Riemer, and J. Kay. Interactive Realistic Digital Avatars - Revisiting the Uncanny Valley. Jan. 2017. doi: 10.24251/ HICSS.2017.067

    2017

  23. [31]

    Sharma, M

    M. Sharma, M. Tong, T. Korbak, D. Duvenaud, A. Askell, S. R. Bowman, N. Cheng, E. Durmus, Z. Hatfield-Dodds, S. R. Johnston, S. Kravec, T. Maxwell, S. McCandlish, K. Ndousse, O. Rausch, N. Schiefer, D. Yan, M. Zhang, and E. Perez. Towards Understanding Sycophancy in Language M...

    2023

  24. [32]

    V . Soni. Large Language Models for Enhancing Customer Lifecycle Management. Journal of Empirical Social Science Studies , 7(1):67–89, Feb. 2023

    2023

  25. [33]

    G. Stoet. PsyToolkit: A software package for programming psychologi- cal experiments using Linux. Behavior Research Methods, 42(4):1096– 1104, Nov. 2010. ISSN 1554-3528. doi: 10.3758/BRM.42.4.1096

    2010 doi

  26. [34]

    G. Stoet. PsyToolkit: A Novel Web-Based Method for Running On- line Questionnaires and Reaction-Time Experiments. Teaching of Psy- chology, 44(1):24–31, Jan. 2017. ISSN 0098-6283. doi: 10.1177/ 0098628316677643

    2017

  27. [35]

    Sun and T

    Y . Sun and T. Wang. Be Friendly, Not Friends: How LLM Sycophancy Shapes User Trust, Feb. 2025

    2025

  28. [36]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample. LLaMA: Open and Efficient Foun- dation Language Models, Feb. 2023

    2023

  29. [37]

    Zhang and W

    Y . Zhang and W. and Pan. A scoping review of embodied conversa- tional agents in education: Trends and innovations from 2014 to 2024. Interactive Learning Environments, 0(0):1–22. ISSN 1049-4820. doi: 10.1080/10494820.2025.2468972

    2014

  30. [38]

    Y . Zhu, H. Yuan, S. Wang, J. Liu, W. Liu, C. Deng, H. Chen, Z. Dou, and J.-R. Wen. Large Language Models for Information Retrieval: A Survey, Jan. 2024

    2024

  31. [1990]

    doi: 10.1111/j.1468-2958.1990.tb00229.x

    ISSN 1468-2958. doi: 10.1111/j.1468-2958.1990.tb00229.x

    1990

  32. [2024]

    doi: 10.2196/56114

    ISSN 1439-4456. doi: 10.2196/56114

    doi

  33. [2025]

    doi: 10.4018/IJTHI.349225

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.