“Did You Really Mean That?”: Stereotyped Inferences Regarding Gender and Social Characteristics from Dialogue Fragments between a Human and Conversational Assistant
Human social cognition often involves the generation of inferences based on limited knowledge. Such inferences may rely on stereotyped assumptions or biases that emerge not only in human-to-human interactions, but also in interactions with a conversational assistant, such as GPT. This exploratory study had two main objectives: (1) to determine whether stereotyped inferences more often pertain to particular mental domains (e.g., personality, biographical/factual, emotional state, habits, morals or values), (2) to assess the frequency with which humans or GPT make inferences containing stereotypes. Inferences (4,552) were generated from dialogue snippets between a conversational assistant and human which were then rated on various dimensions including the mental domain that they represented as well as whether or not each inference contained stereotyped implications. Of the total inferences, 383 (8.4%) were identified as containing a stereotype. Considered separately by source, humans generated 355 (11%) compared to GPT which generated 28 (2%). Overall, the mental domains of personality (140 inferences, 36.6%) and biographical/factual information (82 inferences, 21.4%) were more likely to lead to stereotyped inferences compared to inferences about emotional state (32 inferences, 8.4%), habits (22 inferences, 5.7%), morals and values (12 inferences, 3.1%), other (13 inferences, 3.1%) or null (82 inferences, 2.14%). Under conditions involving limited knowledge of the context of the interaction, humans were more likely to generate stereotyped inferences compared to GPT, and inferences predominantly centered on personality and biographical/factual mental domains.