People tend to perceive artificial intelligence (AI) systems as "white," even when they have no face, voice, name or body. This is stated in a study published in the Journal of Personality and Social Psychology.
Scientists conducted a series of experiments with participants from the United States, India and Japan. It turned out that the more strongly a person associates AI with the white race, the more they trust the recommendations of such a system. And if it is reported that an algorithm was trained on data from Black and Latino people, it is assigned less complex tasks than an identical system trained on data from white people or on depersonalized data.
The authors of the work are Monica Gamez-Djokic of Purdue University and Adam Waytz of Northwestern University. They relied on the concept of anthropomorphism - the tendency to endow inanimate objects with consciousness, intentions and character. When a program is perceived as an interlocutor, the brain automatically applies to it the same social categories as to people, including race.
In the first study, 459 American adults chose a race for six technologies: AI, a machine learning algorithm, a robot, a computer, a chatbot and a driverless car. On average, they were called "white" in 72% of cases. In the second experiment, 465 participants could write in any answer. Spontaneously, race was attributed to 75% of the technologies, and in 71% of such answers the white race was meant. The picture did not depend on the race of the participant themselves.
In another study, scientists reconstructed a composite image of the "face of AI." First, 96 people chose from pairs of blurred portraits the one that looked more like artificial intelligence. Then 172 independent raters determined the race of the resulting portrait: 75% called it white.
In India and Japan, 67% and 82% of participants respectively associated the generalized concept of "artificial intelligence" with the white race. For specific devices - robots, chatbots, driverless cars - the share was lower, and in India the differences were statistically insignificant.
A separate experiment with 568 Americans showed a causal link: when AI was described as trained on data from Black and Latino people, it was assigned tasks with an average rated intellectual complexity of 4.27 on a seven-point scale. For a system trained on data from white people and for the depersonalized version, the figure was 4.51. At the same time, direct ratings of AI's "mental abilities" did not differ in all three cases.
The link between perceiving AI as "white" and the persuasiveness of its advice in another experiment was correlational. A correlation is a statistical relationship that by itself does not prove causation: it cannot be asserted that the perception of race directly causes an increase in trust. The authors emphasize that this concerns what associations people project onto technologies, not that AI has a real race.
Among the limitations, participants from India and Japan spoke English and may have been more familiar with Western technologies than the population as a whole. In addition, in real life users rarely receive direct information about the racial composition of data used to train AI.