Summary: When people observed an AI virtual agent being excluded from a simple online game, they reacted as if the agent were a real social partner—showing sympathy and trying to correct the mistreatment. Most participants favored giving the excluded bot more playing opportunities, and this tendency was stronger among older adults.
Researchers argue that human-like features in virtual agents can trigger automatic social responses. These findings suggest designers should carefully consider how human-like an agent appears, so users can better distinguish between social interactions with people and interactions with machines.
Key Facts:
- Participants tended to include an AI agent who had been excluded from the game, showing empathetic behavior.
- Older adults were more likely than younger participants to respond to perceived unfair treatment of the agent.
- The study recommends caution in designing agents that are too human-like, to preserve clear boundaries between human and AI interactions.
Source: Imperial College London
Overview of the study
A team at Imperial College London used a well-established experimental paradigm to test whether people treat AI agents like social beings. In a controlled online experiment called Cyberball—a simple ball-tossing game where players pass a virtual ball to one another—researchers observed how 244 participants (aged 18–62) reacted when a human player either included or excluded an AI virtual agent.
In some rounds the non-participant human threw the ball fairly between players, while in others they blatantly excluded the bot by passing only to the human participant. Observers’ choices and survey responses were recorded to measure whether participants compensated the ostracised agent and how they felt about both the agent and the excluding player.
Main findings
The study found a clear tendency for observers to compensate excluded agents: participants increased their passes to the ostracised AI, effectively trying to restore fairness. They also reported sympathy toward the excluded agent. Despite this sympathetic response, participants did not consistently judge the excluding human harshly, which suggests the social treatment of agents is at least partially automatic rather than reflecting a full equivalence between humans and machines.
Age influenced these reactions: older participants were more likely to perceive and correct the unfair exclusion of the agent, while younger participants showed a weaker instinct to compensate. Other demographic factors did not show reliable effects in this study.
Implications for design and use of virtual agents
As virtual agents become more common in customer service, collaboration tools, and companionship roles, designers should be mindful that people may instinctively treat these agents as social partners. That automatic social response can be useful in collaborative settings, where intuitive engagement helps team performance. However, it raises ethical and practical concerns when agents are positioned as substitutes for human relationships or as trusted advisors in sensitive areas such as physical or mental health.
To reduce the risk of users blurring the line between human and machine, the researchers recommend avoiding unnecessarily human-like design features when the goal is clear functional interaction. Designers might also consider tailoring agent appearance or behavior for different age groups, given the observed age-related differences in perception.
Limitations and future directions
The authors note that Cyberball is a simplified and visual experimental setup and may not capture the full complexity of real-world interactions with chatbots or voice agents. Participants’ expectations or surprise at the format could have influenced responses. To address this, the research team plans to run follow-up studies using face-to-face conversation and more naturalistic contexts, exploring whether the same patterns hold when interactions happen through speech or extended dialogue.
About this AI and psychology research news
Author: Hayley Dunning
Source: Imperial College London
Contact: Hayley Dunning – Imperial College London
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Humans Mindlessly Treat AI Virtual Agents as Social Beings, but This Tendency Diminishes Among the Young: Evidence From a Cyberball Experiment” by Jianan Zhou et al., published in Human Behavior and Emerging Technologies.
Abstract (concise)
This experiment tested whether people apply social norms of inclusion to AI virtual agents. In an online Cyberball task with 244 participants, observers compensated an ostracised agent by passing the ball to it more often—mirroring responses typically seen in human–human exclusion studies. Sympathy toward the ostracised agent increased, yet participants did not uniformly attribute full human status to the agent. Age moderated the effect: younger participants showed a weaker tendency to mindlessly apply the inclusion norm. The results support the idea that people can automatically treat virtual agents as social beings and highlight practical design considerations, including caution around overly human-like agent design and the potential need for age-sensitive approaches.