Humans Mindlessly Treat AI Virtual Agents as Social Beings, but This Tendency Diminishes Among the Young: Evidence From a Cyberball Experiment

Source: Jianan Zhou, Talya Porat, Nejra Van Zalk; Human Behavior and Emerging Technologies, journal article (open access) · 2024

The “social being” perspective has largely influenced the design and research of AI virtual agents. Do humans really treat these agents as social beings? To test this, we conducted a 2 between (Cyberball condition: exclusion vs. fair play) × 2 within (coplayer type: AGENT vs. HUMAN) online experiment employing the Cyberball paradigm; we investigated how participants (N = 244) responded when they observed an AI virtual agent being ostracised or treated fairly by another human in Cyberball, and we compared our results with those from human–human Cyberball research. We found that participants mindlessly applied the social norm of inclusion, compensating the ostracised agent by tossing the ball to them more frequently, just as people would to an ostracised human. This finding suggests that individuals tend to mindlessly treat AI virtual agents as social beings, supporting the media equation theory; however, age (no other user characteristics) influenced this tendency, with younger participants less likely to mindlessly apply the inclusion norm. We also found that participants showed increased sympathy towards the ostracised agent, but they did not devalue the human player for their ostracising behaviour; this indicates that participants did not mindfully perceive AI virtual agents as comparable to humans. Furthermore, we uncovered two other exploratory findings: the association between frequency of agent usage and sympathy, and the carryover effect of positive usage experience. Our study advances the theoretical understanding of the human side of human–agent interaction. Practically, it provides implications for the design of AI virtual agents, including the consideration of social norms, caution in human-like design, and age-specific targeting.

Technological advances have brought artificially intelligent (AI) virtual agents—software systems that intelligently perform tasks on behalf of and at the request of humans—into people’s daily lives. Consumer-level examples like Amazon Alexa, Apple Siri, Google Assistant, Microsoft Cortana, and OpenAI ChatGPT are now commonly used for tasks such as playing music, sending messages, retrieving information, and controlling household appliances. Their utility also reaches the workspace, providing assistance with professional tasks. As these agents are further integrated into people’s personal and professional spheres, their roles are evolving from mere task helpers to collaborative partners. The growing interdependence between humans and AI virtual agents necessitates better human–agent interaction, which hinges not only on enhancing the agents’ technical capabilities but also on capturing the variations of human responses in interaction.

Intentionally or not, AI virtual agents are being developed to possess increasingly human-like features, with the accompanying assumption that people would attribute them to human-like minds and treat them as social beings. This assumption echoes the basic ideas of the media equation theory, which proposes that when computerised agents present human-like cues, such as interactivity and human roles, individuals respond to these technologies as they would to other humans by applying social norms from human–human interaction. The “social being” perspective has made significant impacts on the design of AI virtual agents and the direction of relevant research (e.g.,); specific study tools have been developed to assess humans’ beliefs in agents as social beings. The validity of this perspective, however, has been questioned by a body of empirical research. For example, Clark et al. found that interviewed participants consistently considered agents as user-controlled tools rather than social equals. Corti and Gillespie found that the mere awareness of the interactant being an “agent” prevented participants from investing in intersubjective effort common in human–human conversation. It has also been argued that individuals apply specific machine heuristics to human–agent interaction. Therefore, the current study is aimed at investigating humans’ tendency to treat AI virtual agents as social beings.

Reeves and Nass highlighted the mindlessness of humans’ social responses to technologies, with mindlessness referring to passive information processing, reliant on pre-established categories/distinctions and scripts without conscious adaptation, as opposed to mindfulness characterised by conscious adaptation to the new context and content of information in the current situation. Nass and Moon showed that participants mindlessly followed social norms to interact with computers and demonstrated behaviours like those in response to other humans, but they mindfully denied how they had behaved during the debriefing. Xu et al. confirmed that mindlessness had greater explanatory power than mindfulness in social perception of technologies. Accordingly, individuals appear to respond to technologies differently depending on whether they mindlessly or mindfully process relevant information.

We therefore connect this idea with a theoretical model of human cognition (i.e., dual-processing model;). This model identifies Type 1 and Type 2 cognitive processing, corresponding to how the brain processes information in mindlessness or mindfulness, respectively. Specifically, Type 1 processing relies on innately specified processing modules or experiential associations to generate automatic, intuitive, and associative responses (i.e., mindless responses); whereas Type 2 processing requires working memory to generate controlled, reflective, and analytic responses (i.e., mindful responses).

A few studies have applied the dual-processing model to investigate individuals’ perceptions of robots. Złotowski et al. identified two forms of anthropomorphism towards robots, resulting from Type 1 and Type 2 processing, respectively, though their methods lacked clarity in differentiation between these two forms. Jahn and Nissen found that participants reflectively distinguished between humans and all types of robots, perceiving robots as less human-like, trustworthy, and warm, but they could only automatically categorise humans and less anthropomorphically designed robots. Nissen and Jahn also explored how participants’ perceived trustworthiness varied depending on their cognitive processing. While these studies have focused on specific perceptual responses, their results all suggest the necessity to consider the involvement of dual cognitive processing behind the apparent human responses.

Therefore, we argue that it is important to differentiate between mindless and mindful human responses when investigating humans’ tendency to treat AI virtual agents as social beings. Moreover, compared to mindful responses, mindless responses generated by Type 1 processing are resistant to change through conscious, intentional efforts and thus can provide insight into the fundamental ways individuals respond during human–agent interaction. Given the difficulty in directly identifying these two types of human responses during natural interactions, we engage participants in specific cognitive processing via experimental approaches to control their working memory load and measure their responses accordingly. As suggested by Lombard and Xu, we incorporate a combination of objective and subjective, as well as real-time and retrospective methods, when measuring mindless and mindful responses.

Our investigation of humans’ tendency to treat AI virtual agents as social beings centres on ostracism/inclusion scenarios, as these originally human–human scenarios have emerged in human–agent interaction as well. Existing human–agent studies typically examine ostracism scenarios in which individuals reject/accept AI virtual agents (e.g.,) or are themselves ostracised by agents (e.g.,). Less is known about human–agent ostracism from a third-person perspective, where individuals observe agents being ostracised by humans. Therefore, in this study, we examine whether individuals in the observing ostracism situation respond to AI virtual agents as they would to other humans.

To establish study scenarios, we employ the Cyberball paradigm, a widely used experimental paradigm that simulates ostracism in virtual, experimentally controlled conditions via an online ball-tossing game. Participants are led to believe that they play a ball-tossing game with other online participants who are actually pre-programmed, fake players. By setting up a situation where participants observe one player ostracised by other players, research has used this pseudo-interactive game to investigate individuals’ responses to observing ostracism during human–human interaction.