Implications of AI Bias in HRI: Risks (and Opportunities) when Interacting with a Biased Robot

Social robotic behavior is commonly designed using AI algorithms which are trained on human behavioral data. This training process may result in robotic behaviors that echo human biases and stereotypes. In this work, we evaluated whether an interaction with a biased robotic object can increase participants' stereotypical thinking. In the study, a gender-biased robot moderated debates between two participants (man and woman) in three conditions: (1) The robot's behavior matched gender stereotypes (Pro-Man); (2) The robot's behavior countered gender stereotypes (Pro-Woman); (3) The robot's behavior did not reflect gender stereotypes and did not counter them (No-Preference). Quantitative and qualitative measures indicated that the interaction with the robot in the Pro-Man condition increased participants' stereotypical thinking. In the No-Preference condition, stereotypical thinking was also observed but to a lesser extent. In contrast, when the robot displayed counter-biased behavior in the Pro-Woman condition, stereotypical thinking was eliminated. Our findings suggest that HRI designers must be conscious of AI algorithmic biases, as interactions with biased robots can reinforce implicit stereotypical thinking and exacerbate existing biases in society. On the other hand, counter-biased robotic behavior can be leveraged to support present efforts to address the negative impact of stereotypical thinking.