Do human-agent teams develop shared mental models as consistently as human-human teams?

What a shared mental model actually is in these studies

A shared mental model is the overlapping picture teammates carry of the task, the team, and what each member is doing. It is what lets people coordinate without spelling everything out. The studies here measured it with standard team-cognition questionnaires after teams played NeoCITIES, an emergency-response simulation in which players manage simulated crises together [1][3].

The key comparison is between all-human teams and mixed teams in which one or two members are AI agents. The consistent finding is that the process of building this shared picture looks familiar in both settings: it develops iteratively over rounds, and communication speeds it along [1]. What differs is the consistency of the end result, not the basic mechanism.

Why agent teammates produce a less even shared picture

In the direct comparison, human-agent teams showed significantly inconsistent levels of team mental model similarity relative to all-human teams, and members reported perceiving less team cognition when working with agents than with humans [1]. Trust followed a similar pattern: teams trusted agent teammates less when the team contained only agents and no other humans [1]. That combination matters, because a mental model is partly a guess about what your teammate knows and will do, and lower trust or a fuzzier read on an agent can leave different humans holding different pictures.

The communication recipe also shifts. In human-agent teams, action-related communication and explicitly stated shared goals were specifically beneficial for developing team cognition [1]. In other words, the shared picture does not come together the same way it does among humans; it needs more explicit scaffolding. That is consistent with the idea that agents do not pick up implicit cues the way human teammates do, though the study characterizes the pattern rather than proving the cause.

Sometimes the mixed team comes out ahead

Consistency is not the same as performance, and here the picture complicates the simple story. In the spatial-information study, all-human teams did not perform at the same level as human-agent teams, and teams with multiple agents performed best [3]. So a less uniform shared mental model did not translate into worse output in that simulation.

Spatial awareness was the lever. Human-agent teams trended toward better mental model similarity when spatial information was high rather than low, while all-human teams showed the reverse trend; the interaction was significant but the follow-up comparisons were inconclusive, so this is a trend rather than a settled effect [3]. Qualitatively, high spatial awareness appeared to jump-start team cognition, support more accurate shared mental models, make the agent easier to understand, and help cognition develop iteratively over time [3]. A plausible reading is that showing everyone the same spatial picture gives humans a concrete anchor for modeling an agent, but that is an interpretation, not an established mechanism.

What this means if you are putting humans and agents on one team

The practical takeaway is to treat shared understanding with an agent as something you have to make explicit rather than assume. Since action-related talk and clearly stated goals were the communication moves that helped human-agent teams build team cognition [1], teams should say out loud what they are doing and what they are aiming for, rather than relying on the tacit coordination that works among humans.

There is also a performance angle worth knowing. Robots that build and maintain shared mental models with their human partners have been associated with better overall team effectiveness than a baseline with no shared mental models, and adding proactive robot behavior improved task efficiency and performance over merely reactive behavior in high cognitive-load settings [2]. That work supports the value of the shared-model idea for agent design, but it does not test how evenly or stably those models are held across a team, so it should not be read as evidence that human-agent teams match human teams on consistency.

How knowledge is managed in conversation may also differ. Work on multiparty dialogue suggests that speakers' strategies for building common ground and modeling their audience differ between human and human-agent conversations, likely reflecting how speakers mentally model their interlocutors [5]. That points to a human-side adjustment when an agent is in the room, though the consistency of the resulting shared models was not directly evaluated. Overall, the honest answer is that human-agent teams can get to a workable shared understanding, sometimes with better performance, but they do not reach it as uniformly as all-human teams, and the conditions that narrow the gap are still being mapped.

Sources used in this answer

Related research questions