When your boss is a robot: workers are more spiteful to robot supervisors that seem more human

Kai Chi Yam, E-Yang Goh, Ryan Fehr, Randy Lee, Harold Soh, Kurt Gray

TL;DR

While anthropomorphism is often touted as the "secret sauce" for successful human-robot interaction, this paper reveals a significant dark side: making a robot supervisor look and act human backfires when it delivers negative feedback. Because human-like features signal "agency" (the capacity for intent), workers perceive a "human" robot’s criticism as a deliberate personal attack, leading to higher rates of retaliation compared to a purely mechanistic robot.

What Problem Does This Paper Address?

As AI and robotics move from the factory floor to managerial roles—already seen in "algorithmic management" at companies like Amazon—organizations face a dilemma. The standard design advice has been to make robots more human-like to foster trust and comfort. However, the authors argue this "panacea" ignores how we perceive the "mind" of the robot.

Drawing on Mind Perception Theory, the researchers explore whether imbuing a robot with a "mind" makes its unpopular decisions seem more malicious. Specifically, they look at the delivery of negative performance feedback. If a machine is just a machine, its "bad news" is just data. But if a machine seems like a person, does that bad news become "abuse"?

Methods

The researchers conducted two controlled laboratory experiments using a physical robot (a Fetch Robotics mobile manipulator).

  • The Manipulation: Participants were randomly assigned to either an anthropomorphized condition (the robot had a digital face, a human name "Paul," and a natural American accent) or a mechanistic condition (blank screen, a serial number "Robot 17,541," and a flat, synthesized voice).
  • The Task: Participants performed ambiguous tasks (a football knowledge test and a simulation). Regardless of their actual performance, the robot delivered harsh, scripted negative feedback (e.g., "Your score was significantly below average... This is very bad").
  • The Measure of Retaliation: In a clever behavioral twist, participants were told the robot used significant electricity and they could "divert processor power" away from it using a physical knob. Turning the knob up would slow the robot's thinking and eventually put it into "stasis." This served as a proxy for supervisor-directed retaliation.

Datasets

As this is an experimental social psychology paper rather than a machine learning paper, "datasets" refers to the participant samples:

  • Study 1: 179 undergraduate students from a large university in Asia. This study focused on the first-person experience of receiving feedback.
  • Study 2: 164 undergraduate students from the same population. This study examined third-party observers who watched the robot deliver negative feedback to a "partner" (a confederate).

Key Insights and Ideas

  • The Agency Trap: Human-like features specifically increase perceptions of Agency (the ability to plan, intend, and act) rather than Experience (the ability to feel). When a robot has high perceived agency, people attribute a "guilty mind" (mens rea) to its negative actions.
  • Disambiguating the Hurt: Negative feedback is inherently painful. We use the "humanness" of the source to decide why it hurts. If the robot is mechanistic, we see the feedback as a cold statement of fact. If the robot is human-like, we perceive the feedback as a choice to be "rude" or "abusive."
  • The Observer Effect: The "backfire" isn't limited to the victim. Third-party observers who see a human-like robot "bully" a colleague are just as likely to retaliate against the robot as the person receiving the feedback.
  • The "Panacea" Myth: The study challenges the industry assumption that anthropomorphism is always better. In "high-warmth" roles (caregiving, teaching), it helps; in "low-warmth/high-authority" roles (managerial feedback), it hurts.

Key Results

  • Perceived Agency: In Study 1, the anthropomorphized robot was rated significantly higher in agency ($M = 5.41$) than the mechanistic robot ($M = 4.92$).
  • The Chain Reaction: The data supported a "serial mediation" model: Anthropomorphism $\rightarrow$ Higher Agency $\rightarrow$ Perception of Abuse $\rightarrow$ Higher Retaliation.
  • Retaliation Levels: Participants were significantly more likely to "power down" the human-like robot after receiving criticism.
  • Comparison to Humans: A pilot study (noted in the footnotes) found that while human bosses are still seen as more "agentic" and "abusive" than any robot, the anthropomorphized robot bridges that gap, making it feel much more like a human abuser than a tool.

Limitations and Future Work

  • Cultural Context: All participants were from an Eastern (Asian) context. The authors note that because robots are more culturally integrated in parts of Asia, Western participants (who might find robot bosses more "novel" or "creepy") might react differently.
  • Stakes: The "retaliation" (slowing a processor) is relatively low-stakes compared to real-world workplace sabotage or HR complaints.
  • The "Mode" Switch: Future work could investigate if a robot should "switch modes"—acting human-like when praising an employee but switching to a mechanistic, data-driven persona when delivering a performance PIP (Performance Improvement Plan).
  • Feature Isolation: The study manipulated face, name, and voice simultaneously. It remains unclear which specific feature (e.g., just the voice or just the face) triggers the agency perception most strongly.

Summary updated

Abstract

Robots are transforming organizations, with pundits forecasting that robots will increasingly perform managerial tasks. One such key managerial task is the evaluation and delivery of feedback regarding an employee’s performance, including negative feedback. However, within this context of delivering negative feedback, we suggest that anthropomorphism—a factor most practitioners and researchers consider as a panacea for overcoming difficulties in human-robot interaction—can backfire. Drawing upon the theory of mind perception, we find that an anthropomorphised robot supervisor delivering negative feedback is more likely than a non-anthropomorphised robot to be perceived as possessing agency. This perceived agency causes perceptions of abuse to arise, which in turn leads to higher supervisor-directed retaliation (operationalized as powering down the robot supervisor; Study 1). These findings even extend to third-person observers witnessing the delivery of negative feedback, again culminating in supervisor-directed retaliation (Study 2). We conclude by discussing the theoretical and practical implications of these findings.
BibTeX
@article{yam2022when,
  author = {Kai Chi Yam and E-Yang Goh and Ryan Fehr and Randy Lee and Harold Soh and Kurt Gray},
  title = {{When your boss is a robot: workers are more spiteful to robot supervisors that seem more human}},
  journal = {Journal of Experimental Social Psychology},
  year = {2022},
  doi = {10.1016/j.jesp.2022.104360},
}

Loading related content...

Figure