Perception of an AI Teammate in an Embodied Control Task Affects Team Performance, Reflected in Human Teammates' Behaviors and Physiological Responses

TL;DR

This study investigates how a human-like AI teammate affects performance in an embodied VR sensorimotor task. Using the ADCT in triads, with and without a Wizard-of-Oz AI agent, it collects multi-modal data including performance metrics, speech, pupil size, blink rate, EEG, and subjective ratings. The key finding is that human-only teams outperform human-AI teams, especially as task difficulty increases, and that AI presence disrupts coordination, elevates arousal, and reduces communication, even as trust in the AI grows over time. The results underscore the need for human-centered AI design that supports shared mental models and adaptive, better-integrated AI teammates to maintain performance in demanding collaborative settings.

Abstract

The integration of artificial intelligence (AI) into human teams is widely expected to enhance performance and collaboration. However, our study reveals a striking and counterintuitive result: human-AI teams performed worse than human-only teams, especially when task difficulty increased. Using a virtual reality-based sensorimotor task, we observed that the inclusion of an active human-like AI teammate disrupted team dynamics, leading to elevated arousal, reduced engagement, and diminished communication intensity among human participants. These effects persisted even as the human teammates' perception of the AI teammate improved over time. These findings challenge prevailing assumptions about the benefits of AI in team settings and highlight the critical need for human-centered AI design to mitigate adverse physiological and behavioral impacts, ensuring more effective human-AI collaboration.

Figures (8)

• Figure 1: Apollo Distributed Control Task (ADCT) overview and performance analysis. a, Experiment setup for human-only and human-AI teams. The left panel shows the different team configurations and participant roles. In human-AI teams, the AI agent always acts as the ThrustPilot. The middle panel illustrates the virtual environment where participants collaborate to control a spacecraft, navigating through red rings and returning to Earth. The right panel depicts the relative positioning and partial views for each role. b, Multi-modal data collection for each role. Behavioral data were recorded for the ThrustPilot, while both physiological and behavioral data were recorded for the YawPilot and PitchPilot in both human-only and human-AI teams. c, Three task conditions in this experiment. From top to bottom, our experiment includes three difficulty levels (Easy, Intermediate, Hard), three experimental sessions (Session 1, 2, 3), and three communication protocols (Incommunicado, Command Word, Free Speech) d-g, Team performance under different task conditions, measured by the number of rings passed. Bars represent performance as mean ± s.e.m. Individual dots represent the number of rings passed by each team (human-only: n=10 teams; human-AI: n=12 teams). d, Team performance across all task conditions. e, Team performance within each task difficulty level. The color key for the team is used for f and g. f, Team performance within each experimental session. g, Team performance within each communication protocol. $\cdot~P< 0.1, *~P < 0.05, **~P < 0.01, ***~P < 0.001$ by One-way analysis of variance (ANOVA) test with Bonferroni correction.
• Figure 2: Behavioral data comparison between human-only and human-AI teams. a, Illustration of an analyzed epoch centered around the moment a ring is passed. Each epoch includes two seconds before and after the ring. b-e, Number of remote controller actions in an epoch for each role under different task conditions (human-only: n=10 teams; human-AI: n=12 teams). b, Number of remote controller actions across all conditions. The color key for the team is used for c-e c, the Number of remote controller actions under different task difficulty levels. d, Actions under different experimental sessions. e, Actions under different communication protocols. f-j, Comparison of communication frequency and duration under different task conditions. f, Speech frequency and duration across all roles and task conditions. g, Speech frequency and duration for each role. h, Speech frequency and duration under different task difficulty levels. i, Speech frequency and duration under different experimental sessions. j, Speech frequency and duration under different communication protocols. One-way ANOVA with Bonferroni correction: ns. not significant, $*~P < 0.05, **~P < 0.01, ***~P < 0.001.$
• Figure 3: Pupil size changes of participants with different roles in human-only and human-AI teams (human-only: n=10 teams; human-AI: n=12 teams). a, Percent of pupil size change for participants with different roles across all task conditions. The color key for the team is used for b-d. The gray area indicates $p < 0.05$ by Welch’s t-test. b-d, Pupil size changes for different roles under various task conditions. b, Changes by difficulty level. c, Changes by experimental session. d, Changes by communication protocol.
• Figure 4: Blink rate and inter-brain synchrony between participants in human-only and human-AI teams. a-d, Blink rate of participants under different task conditions (human-only: n=10 teams; human-AI: n=12 teams). Black dots represent the means. a, Blink rate across all conditions. b, Blink rate of each role under different task difficulty levels. The color key for the team is used for c-d. c, Blink rate of each role within each experimental session. d, Blink rate of each role under different communication protocols. e, Inter-brain synchrony of different frequency bands (human-only: n=10 teams; human-AI: n=12 teams). In each team, inter-brain synchrony is measured using Total Interdependence (TI) between YawPilot and PitchPilot. One-way ANOVA with Bonferroni correction: ns. not significant, $\cdot~P < 0.1, *~P < 0.05, **~P < 0.01, ***~P < 0.001.$
• Figure 5: Subjective rating of ThrustPilot's helpfulness, leadership, and trust of the AI agent (human-only: n=10 teams; human-AI: n=12 teams). One-way ANOVA with Bonferroni correction and repeated measures ANOVA for AI trust, ns., not significant; $*~P < 0.05$.