Humans perceive warmth and competence in artificial intelligence

Abstract

Artificial intelligence (A.I.) increasingly suffuses everyday life. However, people are frequently reluctant to interact with A.I. systems. This challenges both the deployment of beneficial A.I. technology and the development of deep learning systems that depend on humans for oversight, direction, and regulation. Nine studies (N = 3,300) demonstrate that social-cognitive processes guide human interactions across a diverse range of real-world A.I. systems. Across studies, perceived warmth and competence emerge prominently in participants' impressions of A.I. systems. Judgments of warmth and competence systematically depend on human-A.I. interdependence and autonomy. In particular, participants perceive systems that optimize interests aligned with human interests as warmer and systems that operate independently from human direction as more competent. Finally, a prisoner's dilemma game shows that warmth and competence judgments predict participants' willingness to cooperate with a deep-learning system. These results underscore the generality of intent detection to perceptions of a broad array of algorithmic actors.

Keywords: Artificial intelligence; Human-computer interaction; Psychology.

Graphical abstract

Figure 1

People perceive A.I. systems as social actors Error bars reflect 95% confidence intervals. (A) People attribute significantly greater intentionality to A.I. systems than to tools with similar uses. See also Figure S1. (B) People believe that A.I. systems possess a mind of their own to a significantly greater degree than do tools with similar uses. See also Figure S2. (C) Despite feeling similar levels of gratitude toward A.I. systems and tools for their use, people report being significantly more likely to follow politeness norms when interacting with A.I. systems. See also Figures S3 and S4. (D) People endorse others’ application of politeness norms to A.I. systems to a significantly greater degree than to tools. See also Figure S5.

Figure 2

Warmth and competence emerge as prominent dimensions in impressions of real-world A.I. systems Error bars indicate 95% confidence intervals. (A) On average, impressions of the A.I. systems contained significantly more competence-related content than warmth-related content. See also Figure S6. (B) Warmth and competence-related content appears in impressions at significantly higher levels relative to other common perceptual dimensions. Morality and sociability content constitutes the warmth dimension; ability and assertiveness content compose the competence dimension. See also Tables S1–S3 and Figure S7. (C) An A.I. system’s role significantly predicted warmth coverage. See also Table S4. (D) Similarly, an A.I. system’s role significantly predicted competence coverage. See also Table S5.

Figure 3

Impressions of real-world A.I. systems vary systematically by perceived warmth and competence (Study 6) Circles and font color indicate a priori identified AI system roles. See also Figure S8.

Figure 4

Warmth and competence judgments of A.I. systems vary systematically by the system’s role and correlate with perceived covariation of interests, status, and autonomy Error bars and bands represent 95% confidence intervals. (A) System role significantly affected warmth judgments. See also Table S9. (B) Similarly, system role significantly influenced competence judgments. See also Table S10. (C) Warmth evaluations positively correlated with perceived covariation of interests. (D) Competence evaluations positively associated with perceived status. (E) Competence evaluations also exhibited a positive association with perceived autonomy. For panels (C)–(E), see also Figure S9.

Figure 5

Interdependence and autonomy drive warmth and competence judgments of A.I. systems Error bars and bands depict 95% confidence intervals. (A) Providing information about an A.I. system’s reward scheme significantly influenced the covariation of interests perceived by participants. See also Figure S10 and Table S11. (B) Perceived covariation of interests exhibited a significant positive association with warmth evaluations. See also Figures S11 and S12. (C) Providing information about an A.I. system’s ability to initiate actions by itself significantly affected perceived autonomy. See also Figure S13 and Table S12. (D) Perceived autonomy in turn demonstrated a significant positive correlation with competence judgments. See also Figures S14 and S15.

Figure 6

Interdependence scaffolds impressions of A.I. co-players in incentivized interactions Error bars and bands represent 95% confidence intervals. (A) On average, participants judged A.I. co-players as significantly more competent than warm. (B) The degree of alignment between the A.I. co-player’s reward and participant score significantly altered perceived warmth (see also Figure S17). This effect also appeared in post-interaction impressions (see Figure S19). (C) The autonomy of the A.I. co-player significantly influenced perceived competence (see also Figure S18). This effect also emerged in post-interaction impressions (see Figure S20). (D) Initial judgments of warmth significantly predicted participants’ in-game choice of how many tokens to transfer to their A.I. co-player in the first round. A similar relationship emerged in the second round (see Figure S21). (E) Initial judgments of competence did not significantly correlate with initial transfer choices. However, perceived competence demonstrated a significant relationship with participant choices in the second round (see Figure S22). The y axis in panels (D) and (E) is re-scaled to depict the range of participant actions (transfer zero through 10 tokens).