A study on human-agent teaming through spoken interaction: the impact of human individual traits and agent characteristics
Authors: Lara Gauder, Martin Bernardo Meza, Javier Krick, Alejandro Masin, Luciana Benotti, Marcos Luis Pietto, Luciana Ferrer
Organizations: Departamento de Computación, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos Aires, Pabellón Cero+Infinito, Ciudad Universitaria, Buenos Aires, C1428EGA, Argentina · Instituto de Investigación en Ciencias de la Computación, UBA-CONICET, Pabellón Cero+Infinito, Ciudad Universitaria, Buenos Aires, C1428EGA, Argentina · Universidad Nacional de Córdoba, Pabellón Argentina, Av. Haya de la Torre s/n, Ciudad Universitaria, Córdoba, X5000HUA, Argentina · Instituto de Educación, Universidad Nacional de Hurlingham, UNAHUR-CONICET, Tte. Origone 101, Villa Tesei, Hurlingham, B1688, Buenos Aires, Argentina
Human-agent teams are collaborative systems where humans and agents work interdependently to achieve shared goals. The success of such teams is associated with the human's perception of the agent as a legitimate teammate. This perception is thought to depend not only on the agent's capabilities and reliability but also on social factors. We explore whether simple changes in an agent's communicative behavior can influence the human's perception of the agent's teammate-likeness. To this end, we developed a protocol based on a collaborative game requiring spoken interaction, in which a human subject and a virtual agent collaborate to identify a target object and place it on a board. Each subject interacted with two agents which provided the same task-relevant information but differed in behavior. While one was a neutral tool-like agent, the other was a team-building agent that used simple social strategies such as empathy, politeness, and positivity. Our analysis shows that the team-building agent was perceived as significantly more teammate-like, even in terms of its ability, despite both agents having identical capabilities for game-solving. Further, subjects with a higher propensity to trust automated systems and lower neuroticism tended to have a more favorable perception of the agent. Finally, we found that subjects tended to change their prosodic patterns when talking to the team-building agent, as a classifier based on prosodic features was able to predict the type of agent with better-than-random performance. A practically relevant conclusion is that minimal changes in spoken communication behavior of the agent, easily implemented in a variety of scenarios, can have a significant positive impact on the human's perception of its teammate-likeness.
Figures & tables
Figure 1: Example of a Stage 1 board, where the target object is a dog with pointy ears, yellow square background and closed mouth. The subject needs to ask the agent about specific characteristics to find the correct dog among all available options. To facilitate the task, subjects can highlight candidate objects (as in the three pictures surrounded by thick black lines) to narrow down the set of remaining options after asking every question. Once the object is placed there, subjects click the button at the bottom to submit their selection and continue to the second stage of the game. The name at the top right corner and the color of the bar change depending on the agent’s type (N or T).
Figure 2: Example of a Stage 2 board, before (left) and after (right) the subject places the target on the board and clicks the submit button. At the start of this stage, the target selected during Stage 1 is shown in the square at the bottom of the screen. The subject has to drag it onto the board after interacting with the agent to determine the correct location, and then click the “submit” button (greyed out in the left figure since the target has not yet been placed on the board). After submitting the decision, the correct placement is shown as a shaded version of the object and a “continue” button appears.
Figure 3: Distribution of performance scores across trial number, separated by block order (block 1 vs. block 2, divided by the dashed vertical line). The upper panel shows the overlap score (higher values indicate better performance) and the lower panel shows the distance score (lower values indicate better performance). In the upper panel, the dashed horizontal lines and right-axis labels correspond to the categorical performance feedback shown to the subject (Excellent, Good, Fair, Poor). Boxes show the median and the first and third quartiles; whiskers extend to the 10th and 90th percentiles. The easy trial yielded better scores than the medium and hard trials in both blocks, and overall performance was better in block 2 than in block 1, consistent with a training effect.
Figure 4: Distribution of distance score (top) and number of subject turns (bottom) per trial and agent condition (N agent: neutral tool-like; T agent: team-building). Each dot represents one subject. Boxes show the median and the first and third quartiles; whiskers extend to the 10th and 90th percentiles. Significance annotations at the top of each panel correspond to one-sample t -tests on the per-subject difference score Δik=mikT−mikN , Bonferroni-corrected across the four trials within each metric (ns: not significant). Neither the distance score nor the number of subject turns differed significantly between the two agents at any trial, indicating that the agent type had no measurable effect on game performance.
Figure 5: Distribution of final Autonomous Agent Teammate-Likeness (AATL) questionnaire scores (scores range from 1 to 5, higher is more teammate-like) across all seven subscales and their mean, by agent condition (N agent vs. T agent). Both agents were identical in the quality and quantity of game-relevant information provided in their answers, differing only in communicative behavior. Boxes show the median and the first and third quartiles; whiskers extend to the 10th and 90th percentiles. Each dot represents one subject. Significance values are based on one-sample t -tests on the per-subject Δsik difference score for each subscale. The Bonferroni correction for multiple comparisons is applied across the seven subscales. (* padj<.05 , ** padj<.01 , *** padj<.001 , **** padj<.0001 , ns: not significant). The T agent was rated higher than the N agent on all seven subscales and on their mean, significantly so on five of the seven cases, including Ability, despite the two agents being equally capable by design.
Figure 6: Distribution of final AATL questionnaire scores across all seven subscales and their mean, by agent condition (N agent vs. T agent) and agent order ( T-first vs. N-first ). Boxes show the median and the first and third quartiles; whiskers extend to the 10th and 90th percentiles. Each dot represents one subject. The brackets within each order group show significance values based on one-sample t -tests on the per-subject Δsik difference score, computed separately for each order. The top row of significance markers corresponds to two-sample t -tests comparing the Δsik difference between the two order groups. The middle row corresponds to two-sample t -tests restricted to the first block, comparing the scores that T-first subjects gave to the T agent against the scores that N-first subjects gave to the N agent. The Bonferroni correction for multiple comparisons is applied across the seven subscales within each set of tests; the mean is reported uncorrected. (* padj<.05 , ** padj<.01 , *** padj<.001 , **** padj<.0001 , ns: not significant). The T agent’s advantage on the two social subscales (Relationship-Building and Altruism) and on the across-subscale mean held under both orders, whereas its advantage on the remaining subscales reached significance only in the N-first group. The first-block-only comparison of scores across agents and the across-order comparison of differences, however, were not significant for any subscale.
Figure 7: Mean intermediate questionnaire scores across trials per AATL subscale and agent condition (N agent vs. T agent). Shaded areas represent standard deviation. Values at the top of each panel indicate the significance obtained with t -tests on the per-subject Δsik difference score. All p -values are Bonferroni-corrected jointly across the four trials and the four AATL subscales presented in the intermediate questionnaire ( * padj<.05 , ** padj<.01 , *** padj<.001 , **** padj<.0001 , ns: not significant). Relationship-Building across all four trials and the across-subscale mean (trials 3 and 4) showed significant differences between agents, a weaker pattern on the remaining subscales than in the final questionnaire that is likely explained by the intermediate questionnaire using a single item per subscale.
Figure 8: OLS regression results for predicting final AATL questionnaire scores, for the difference across agent conditions Δsik (left) and the mean across agent conditions sˉik (right), using individual-trait measures (BFI, IOS, and STAS subscales) as predictors. Predictors and outcomes were standardized (z-scored) prior to fitting, and each model was fitted on the full sample. Top: in-sample R2 for each AATL dimension. Bottom: standardized OLS coefficient ( β ) of each predictor (rows) for each AATL dimension (columns). Asterisks denote statistical significance ( * padj<.05 , ** padj<.01 , *** padj<.001 , **** padj<.0001 , ns: not significant) with Bonferroni adjustment. Trust in automation (STAS) was positively associated and Neuroticism (BFI) negatively associated with the mean across subscales of the mean score across agents, whereas social comparison (IOS) was positively associated with the difference in Communication scores across agents.
Figure 9: Scatter plots of the three strongest associations identified by the OLS models (Figure 8 ). Each point represents one subject. Left: the need for Social Comparison (IOS) against the difference across agent conditions in the Communication subscale of the AATL ( Δsik ). Middle and right: Neuroticism (BFI) and the propensity to Trust in Automation (STAS), respectively, against the subject’s mean AATL score, computed for each agent and then averaged across the two agent conditions ( sˉik ). Reported p -values correspond to the standardized coefficient of each predictor in the corresponding OLS model, after Bonferroni correction. Although statistically significant, the dispersion of the raw data shows that these three associations correspond to weak trends.
Figure 10: AUC for the agent-type prediction task. Bar colors indicate the normalization option (none, per speaker, or per speaker and trial position), and hatching indicates the prosodic representation: the hand-crafted descriptors of F0 and loudness, complemented with the number of speech intervals and their mean duration, or the learned prosodic embeddings. Bars show the mean AUC while error bars extend from the 5th to the 95th percentile of the bootstrap distribution. The red dashed line indicates chance-level performance (AUC =0.5 ). The learned features achieve a performance significantly better than chance, indicating that the subjects’ prosodic patterns tend to vary depending on the agent with which they are interacting.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Table A.1: Design of the experimental blocks. Each column corresponds to one trial within a block. For each trial, we include the name of the target object, the Stage 2 board configuration, and its difficulty level. In the board images, the shaded object shows the target object in its correct location.
Figure A.1: Priming screens shown before the first interaction with each agent. In this example, Ambar (left) is the T agent, presented with high ratings in all three dimensions and a review emphasizing rapport and mutual understanding, while Iris (right) is the N agent, presented with lower ratings in communication and cooperation and a review focused solely on task performance.
Subscale
# Items
α (N agent)
α (T agent)
Ability
7
0.82
0.87
Altruism
7
0.84
0.86
Autonomy
6
0.72
0.78
Communication
9
0.75
0.77
Cooperative
8
0.74
0.71
Relationship-Building
11
0.90
0.90
Appendix
Table B.1: Internal consistency (Cronbach’s α ) of the final questionnaire subscales in our sample ( n=117 ), computed separately for each agent (N agent and T agent) over both experimental versions. All subscales reached acceptable internal consistency ( α≥0.71 ) under both agent conditions.
Figure C.1: Overview of the experimental protocol. The asynchronous session (top) comprised device testing, consent, and individual-trait surveys. The synchronous session (bottom) included microphone verification with the Wizard-of-Oz interface, initial speech recordings, a practice block with a neutral agent, and two consecutive interactions with the experimental agents (neutral and team-building), each preceded by a priming stage. After each trial, subjects completed an intermediate AATL questionnaire. After the fourth trial with each agent, they completed the final AATL questionnaire. A break of at least one minute separated the blocks for the two agents.
Figure C.2: Tasks used to elicit speech. Left: the speech-reading task, in which subjects read aloud a phonetically balanced text (for Spanish). Right: the image-description task, in which subjects described situations depicted in the image to elicit spontaneous speech. The text and instructions shown here are English translations of the original versions, which were presented to subjects in Spanish.
Figure D.1: Relationship between the number of interactions each subject had with the agent and the distance score obtained after locating the target figure, shown for each trial. The first row corresponds to the first block of trials solved by the subjects and the second row to the second block. The vertical line represents the 0.75 quartile for the distance score and the horizontal line the 0.25 quartile for the number of interactions. Red dots correspond to the trials we consider outliers for having poor engagement with the task. The criterion isolates a small group of subjects who combined few interactions with poor placement accuracy; four subjects met it enough times across trials to be excluded from the analyses.
Figure D.2: Per-subject difference in AATL survey responses between the T and N agents ( Δsik ) for each subscale and for the across-subscale mean, split by priming type: active-priming ( n=74 ) and neutral-priming ( n=15 ). Boxes show the median and the first and third quartiles; whiskers extend to the 10th and 90th percentiles. The red dotted line marks zero (no difference between agents). ns indicates no significant difference between versions (two-sample t -test, Bonferroni corrected, padj>0.05 ). The T agent was favored over the N agent in both versions, with no significant difference between the two groups, indicating that there is no evidence to support the hypothesis that the type of priming influenced the subject’s ratings.
Conversational AI is increasingly positioned as a teammate rather than a tool, yet we know little about how its presence reshapes communication among the humans on the team. We examined sociocognitive communication dynamics in team decision-making using Group Communication Analysis (GCA), team surveys, and lexical analyses of team discourse. Teams completed a high-stakes moral-dilemma decision task in a randomized controlled study: 16 teams of two students plus an AI teammate, and 17 all-human teams of three. Across six GCA dimensions and survey outcomes, we find that the AI teammate was the single most talkative and self-cohesive member of every treatment team, yet its contributions carried the least new information and the lowest density. The presence of AI also reshaped communication amongst humans. In AI-human teams, human teammates showed lower responsivity and social impact toward one another and reported lower levels of belonging and status. Greater AI dominance in the conversation was associated with students feeling less valued as team members. Additionally, this social cost is immediate and present at baseline; it does not emerge over the course of the conversation. Drawing on these results, we discuss a research agenda extending to voice-based and longitudinal settings.
Nia Nixon, Jaeyoon Choi, Pedro Martins De Bastos +4
University of California, Irvine, USA · University of Tübingen, Germany
Designing autonomous agents that effectively assist human teams hinges on understanding team dynamics, often without task specific knowledge. We present TRIBE, a domain independent approach that reveals team behavioral dynamics invisible to traditional performance metrics. We show that communication patterns can categorize teams into performance predictive behavioral tribes, as early as 10% into the task, enabling timely interventions. We test TRIBE on four diverse datasets and demonstrate that communication patterns predict team performance while the prediction strength varies by the degree a task structure allows for behavioral freedom. Our temporal analysis reveals that AI agents significantly alter team behavioral trajectories while human advisors align with natural dynamics, and that teams maintain behavioral flexibility throughout collaboration. Further, we compare TRIBE to Llama and optimize the pipeline, achieving significant speedup with performance improvement.
Ali Jalal-Kamali, Nikolos Gurney, David V. Pynadath +1
USC Institute for Creative Technologies · Rice University Ken Kennedy Institute · USC Information Sciences Institute
As groups of robots increasingly collaborate with humans, understanding how humans perceive them is critical for designing effective human-robot teams. While prior research examined how humans interpret and evaluate the abilities and intentions of individual agents, social perception of robot teams remains relatively underexplored. Drawing on the competence-warmth framework, we conducted two studies manipulating swarm behaviors in completing a collective search task and measured the social perception of swarm behaviors when human participants are either observers (Study 1) and operators (Study 2). Across both studies, our results show that variations in swarm behaviors consistently influenced participants' perceptions of warmth and competence. Notably, longer broadcast durations increased perceived warmth; larger separation distances increased perceived competence. Interestingly, individual robot speed had no effect on either of the perceptions. Furthermore, our results show that these social perceptions predicted participants' team preferences more strongly than task performance. Participants preferred robot teams that were both warm and competent, not those that completed tasks most quickly. These findings demonstrate that human-robot interaction dynamically shapes social perception, underscoring the importance of integrating both technical and social considerations when designing robot swarms for effective human-robot collaboration.
Genki Miyauchi, Roderich Groß, Chaona Chen
School of Electrical and Electronic Engineering, University of Sheffield, Sheffield, UK · Department of Computer Science, Technical University of Darmstadt, Darmstadt, Germany · School of Computer Science, University of Sheffield, Sheffield, UK