Power asymmetries in experimental studies of robot-human interaction: learning from artificial intelligence

The development of artificial intelligence and its impact upon the practice of scientific disciplines and on the employment of people in a wide range of jobs and professions, continues to generate debate and concern across the community. In particular there has been criticism of the quality of the data used to generate large language models which have been voiced in many ways. We extend this concern about the quality of evidence to examine critically the methods and the outputs of experimental studies which are used to evaluate the ways in which humans and robotic systems may interact in future jobs involving the augmentation and perhaps replacement of humans in organizations and systems. A common method is the “Wizard of Oz” technique, a metaphor borrowed from the book of that name in which a society appears to be controlled by an eminent and wise person, but who is in fact a wizened and frightened old man. We examine the extent to which this method has been used to study “interactions”, which appear to be autonomous and symmetrical in power, and which may determine ways in which robotic systems replace humans, are in fact asymmetric and biased. We further extend this analysis to include studies of human-human interaction which also produce widespread biased and error-ridden data. We conclude by presenting possible methods and procedures to mitigate the errors which can result from the study of how humans and robots may interact in the future.