While typically used as a working appliance, the dynamic capabilities of AI allow it to be considered a potential teammate. As Rix (2022) summarizes, factors such as a team setting and collaborative actions are “drivers of the formation of impactful human-machine teams” (p. 398). While there are many forms of collaboration (e.g., Parasuraman et al., 2000) and concepts of interaction, like “human centered assistance applications” (Schmidtler et al., 2015, p. 85) and human-AI teaming (Berretta et al., 2023), there is no agreed-on differentiation of AI roles within a work group.

Researchers acknowledge the existence of “various modes of cooperation between humans and AI” (Li et al., 2022, p. 1) — here referred to as AI roles — with an ongoing debate on whether technological agents can, should and may be considered team members or mere tools (Rix, 2022).

Perceiving AI as a teammate goes along with certain behaviors like proactivity (co-creation) and relationship-building (Rix, 2022), and the active and distinct role of AI (Li et al., 2022). The role of AI as a supportive tool (e.g., Fan et al., 2022) is described as helping with tasks while limited in its function and decision spectrum, not going beyond what told to do and not reaching into social spheres. However, more nuanced taxonomies are missing, and research on the impact of different forms of implementation of AI is urgently needed.

Previous research has shown that humans prefer to work with other humans rather than with AI because they perceive the other human as a teammate (Sadeghian & Hassenzahl, 2022). Following this insight, team cohesion can be influenced by implementing AI as a team partner or a tool, by its role’s perception and by the trust in the AI system (see Figure 1).

Regarding technology and AI use, gender is a potentially influential variable affecting attitudes and perception (see e.g., Cai et al., 2017). The findings are ambivalent: Ray and colleagues (1999) showed that men and women are equally comfortable with technology at work. In fact, women even have a more positive attitude towards it. Nevertheless, gender is related to how much time people spend with technology (Cai et al., 2017).

Thus, the variables gender and experience might have an influence on how people perceive AI within teams and the collaboration with it. Because research does present an ambiguous picture of how these influences look like, experience and gender are used as control variables within this study.

The study was developed during a university seminar. As this paper serves as an overview of the research findings, the hypotheses deviate from those specified in the preregistration (see Tausch et al., 2024), which were individual student’s hypotheses for the seminar. This manuscript is an exploratory investigation of the results with exploratory hypotheses.

The experimental online study tested the influence of different types of AI application on team cohesion, moderated by trust in the AI and controlled for gender and AI experiences. The data collection via SoSci Survey took place from March 12 to June 23, 2024. The study was classified as ethically agreeable by the ethics committee (Application 919) of Ruhr-University Bochum.

Using a vignette study, participants were assigned a “job”, either in recruiting or in the scientific field, and were asked to imagine an exemplary workday scenario with typical tasks, supported by an AI system. They then evaluated the depicted situation. The corresponding vignettes can be viewed in full length at Hoffmann et al. (2025). They vary in their central characteristics regarding the role of the AI system, which results in different experimental conditions (Rungtusanatham et al., 2011). The dispersion of participants was partially randomized, as a software guaranteed balanced distribution of participants across the conditions within fully finished questionnaires. Data collection was anonymous and not incentivized.

For measuring team cohesion, the RoBoCo-scale was selected (Tausch & Kluge, 2026). This scale captures cohesion between robots and humans but can also be used in human–AI teams according to the authors. Thus, it was deemed more suitable than general cohesion measures for human-only teams. The participants rated statements like “I have a good relationship with the AI” on a scale from 1 (Completely agree) to 7 (Strongly disagree).

Trust in the AI was measured using the Human-Computer Trust Scale (HCTM), which is specifically designed to analyze human trust in different technological systems and comprehensive enough to capture trust in the presented AI in a differentiated way (Gulati et al., 2019). On a five-point Likert scale, participants rated statements such as “I believe that the AI is acting in my best interests” from strongly disagree to completely agree. The German-language items were provided by the original authors.

Fourteen self-designed items separated into two scales were used to record participants’ subjective rating of whether they perceived AI as a tool or as a teammate (seven items each). On a five-point Likert scale, the participants rated statements such as “The AI and I together form a social unit” (Teammate) or “The AI is a useful tool” (Tool) from 1 (Strongly disagree) to 5 (Completely agree). Additionally, a continuum was included, on which the participants were asked to specify from 1 to 100 if the AI was perceived more as a tool (1) or as a teammate (100), because it remains unclear if the roles are distinct constructs or extremes on a continuum.

Four hundred and eighty-four participants were acquired using the students’ personal networks and the platform SurveyCircle. The sample was filtered according to the exclusion criteria defined in the preregistration (see Tausch et al., 2024), decreasing it to 217 participants. A detailed description of the exclusion can be viewed at Hoffmann et al. (2025). For the results, a sub-sample of 150 participants was relevant, excluding those receiving the vignettes of collaboration with a human assistant. The following data relates to the sub-sample of people being confronted with an AI vignette. The average age was M = 31.40 years (SD = 13.30), with a minimum age of 18 and a maximum age of 73. Educational qualifications were also surveyed. 36.87% stated that their highest educational qualification was a university degree, 8.29% a university of applied sciences degree, 46.08% a high school diploma, 2.76% an intermediate school leaving certificate, 0.92% a lower secondary school leaving certificate and 5.07% a doctorate or habilitation. In general, 81 participants (33.33% male) received the vignette in which AI was presented as a tool, whereas 65 participants (30.77% male) received the vignette where AI was introduced as a teammate.

To test the connection between the perception of AI and their influence on team cohesion, moderated by trust, the statistic software R 4.3.1 was used (R Core Team, 2023). Data was merged for the two different jobs presented, as distinctions were not relevant for this analysis. Scale analyses for reliability information and scale improvements were performed. Then, the data was analyzed descriptively.

To test the hypotheses for significance, a path analysis (see Figure 1) was performed. A distinction was made between the direct mediation of the perception of AI (scales and continuum, M_1-3) from the various vignettes (X) to team cohesion (Y) and a moderation effect through trust (W). Three different types of mediation were distinguished: M₁) via the evaluation of AI as a teammate, M₂) via participants’ rating of AI as a tool, and M₃) via rating on a continuum from tool to teammate. This analysis was performed considering the covariates gender and experience with AI using the lavaan package (Rosseel, 2012).

All descriptive values can be found in Table 1.

All hypotheses were tested in a path model depicted in Figure 2.

It is important to stress that the hypotheses and the path model explored in this study are exploratory because the preregistration happened at the beginning of the university seminar and included our methods but also the individual students’ hypotheses. After the seminar, it became clear that the manipulation had little to no direct effect on the different dependent variables mentioned in the preregistration (e.g., team cohesion), so this paper aimed to explore if role perceptions play a mediating role.

Additionally, as mentioned above, vignettes as stimuli come with limited external validity (Groß & Börensen, 2009). Maybe vignettes simply don’t produce role perceptions involving cognitive and affective components to the same extent as experiencing the actual situation would (Collett & Childs, 2011) and therefore participants base their rating on prior experience. Statements based on the effects of the vignettes should therefore be made carefully. Although all participants who stated that they were unable to imagine the given scenario were excluded, this is only a subjective assessment. In a follow-up study, we will address the issue of vignette immersion as well as an even clearer, more nuanced differentiation of AI roles by, e.g., using video snippets of an interaction with AI. This is supposed to increase the influence of the experimentally manipulated AI roles on people’s perception and lead to clearer results regarding factors influencing cohesion in human–AI teams. Furthermore, no substantial difference was found between the two AI perception types, which may indicate that this cannot be regarded as dichotomous. It may be possible to use AI as a tool while at the same time considering it as a teammate.

It should also be noted that team cohesion and teammate perception correlate substantially with each other (.80). This might indicate possible equality of the two constructs. Looking into the definitions of both as stated by Wynne and Lyons (2018), it becomes clear that an AI system that is perceived as a “communicative and synchronized agentic teammate” (p.355) might as well evoke motivation to work together. The differentiation between the two is thus unclear. Nevertheless, important differences between the scales should not be neglected. While team cohesion asks to what extent the participants are willing to correct mistakes in the team, team partner perception asks whether the AI is a helpful team partner.

In the following, we will look deeper into how perception of AI might influence felt team cohesion. Our study supports the findings of Rix (2022) that in order to work together on common goals, one has to perceive the AI as an “other”, while working with AI as a tool is a unidirectional interaction. Our results show that the AI here is probably not intensely perceived as a social partner, even if it is presented as such in the scenario using framing and showing, e.g., proactive behavior. It remains to be investigated whether this is due to insufficient operationalization of teammate-likeness according to the model of Wynne and Lyons (2018), due to a lack of real interaction and team development, or simply due to the fact that it is hard to imagine current systems to be perceived as partners. Further studies should more systematically follow the definitions by Wynne and Lyons (2018) and realize longitudinal designs with longer-term interactions with existing or simulated AI systems in different roles. Simultaneously, theory is needed to systematize roles of AI within a team.

Trust in AI can be a good starting point for people to strengthen team cohesion under certain circumstances. The moderating effect of trust shows potential for companies to support teaming by actively promoting trust in AI teammates. Therefore, companies could offer targeted training programs or discourse formats to promote adequate use of AI and exchange best practices. Employees can be actively involved in the process of establishing AI in the company, also giving them the opportunity to express concerns (Lee & See, 2004). This can reduce doubts and foster trust, which in turn helps experiencing cohesion with it as a collaborator, which is then related to positive outcomes (Grossman et al., 2022). However, it is important to note that the moderating effect also means that trusting AI tools correlates with lower team cohesion. A possible explanation might be that learning that AI is reliable will strengthen trust but also the impression that AI doesn’t act on its own but on behalf of the human worker which weakens feelings of team cohesion (based on the definitions of trust and team cohesion, see above). Working with AI as a tool, team cohesion might nevertheless not be as favorable as in situations with AI teammates and promoting trust might not be beneficial. Future research could aim to identify more favorable outcome variables in companies working with AI tools.

As team cohesion is a variable that has a crucial influence on performance, organizations working with AI as a team member should aim to increase it. In this context, it is beneficial if AI is seen as a teammate rather than a tool. Furthermore, in human–AI–teams, trust can be helpful to strengthen team cohesion but not if AI is perceived as a tool. Still, team cohesion is not the only relevant outcome variable for modern companies and, depending on a company’s needs, AI tools might be more suitable. Further research could be conducted on the direct effect of role perceptions on performance or work satisfaction of workers that use, or work together with, AI. As this study has shown, interpersonal factors normally only researched in human teams, such as team cohesion and trust, are also relevant when working with technological systems and need to be properly addressed in theory, research, and organizations.