Human preferences for cognitive and emotional capabilities in robots across different application domains

In navigating social situations, humans strongly rely on the perception or attribution of mental states—such as beliefs, desires, intentions and feelings—to others, based on inferences drawn from observable behavior. A substantial body of human-robot interaction (HRI) research indicates that such folk-psychological interpretations are not limited to human and animal behavior but are also highly prevalent in people's interpretations of the behavior of robotic systems. This has been argued to be due to the fact that the human mind tries to understand and explain robot behavior based on the same conceptual framework normally used to explain human behavior.

While both causes and effects of mental state attribution to robots are well studied, relatively little is known concerning what preferences people hold regarding mental capabilities in robots. A pioneering study on such robot mind preferences by Malle and Magar (2017) found that people generally wanted logical robots without emotions, but still able to have empathy; that study found no differences in preferences across domains (domestic, nursing, military). Thellman et al. (2023) conducted a survey exploring people's preferences for higher-order mental states in artificial agents, such as robots' beliefs about human intentions, and found that preferences vary by context and function. Building onto this work, the present study adopts a more expansive approach, exploring a broader range of robot application domains, investigating human factors influencing these preferences, and collecting qualitative data regarding participants' reasoning, using an existing two-dimensional model of mind (experience and agency).

As robots become increasingly integrated into various domains of society, understanding these preferences is important for facilitating successful human-robot interactions across different contexts. People might, for example, want their own social companion robot to have certain emotional capabilities, and they might want automated vehicles to understand the goals and intentions of vulnerable road users, but they might not be willing to let robotic sales staff read their mind and influence their shopping decisions. Hence, the objective of this study is to investigate if preferences for mental capacities in robots vary across different application domains and identify potential factors influencing these preferences. While previous research mainly has focused on mind perception/attribution in interactions with robots, this study shifts the focus to individuals' preferences, aiming to fill a knowledge gap in the HRI literature. The following research questions guided the investigation: RQ1: Do participants' preferences for capabilities of agency and experience in robots vary across different application domains? RQ2: Do participants' demographics (age, gender, educational background) affect preferences for capabilities of agency and experience in robots?

Prior HRI research shows people routinely attribute mental states to robots, interpreting robot behavior through a human-centered folk-psychological lens. Malle and Magar (2017) reported a general desire for robots that are logical and unemotional, yet empathetic, with no domain differences observed (domestic, nursing, military). Thellman et al. (2023) found that preferences for higher-order mental states in artificial agents vary by agent context and function. The study builds on the two-dimensional model of mind—agency (action-related, cognitive) and experience (emotion-related)—originally proposed by Gray et al. (2007) and refined by McMurtrie (2023). While most prior work has focused on mind perception/attribution, the present study targets desired mental capabilities (preferences), expanding the range of application domains, considering demographic moderators, and adding qualitative insights to understand underlying reasons.

Design: Between-subjects survey study with six application domains (healthcare-surgical robot, defense-military robot, household-cleaning robot, social-companion robot, education-educational robot, customer service-service robot). Participants were randomly assigned to one domain.

Participants: N = 271; women 115 (42.4%), men 155 (57.2%), one identified as other (excluded from gender analyses). Age 19–76 years (M = 26.3, SD = 10.2); 76.8% aged 18–25. Education: 6.3% no university/college; 59.0% Bachelor’s (ongoing/completed); 34.7% Master’s (ongoing/completed). Computer science education: 28.4% (51 men, 26 women). Recruitment via convenience sampling at Linköping University campuses (Sweden) and online; inclusion criterion: age > 18. Sample culturally homogeneous Swedish-speaking.

Measures: Dependent variable was preference for mental capabilities, grounded in the two-dimensional mind model (experience and agency). From McMurtrie’s refined 22-capability model, 12 capabilities were selected (six per dimension) using criteria to balance dimensions, exclude bodily sensations, reduce redundancy and vagueness, and prioritize items with strong factor loadings. Experience: feel happy, feel pleasure, feel pain, feel panic, love specific people, have intense urges. Agency: plan for the future, understand a person’s goals, explain their decisions, praise moral actions, disapprove of immoral actions, reason logically.

Procedure: LimeSurvey questionnaire (20 items). Demographics collected first (gender, age, education level, computer science background). Participants were introduced to one robot type in a specified domain (with a purpose description) and asked to imagine its appearance and operation, then briefly describe the imagined robot (free text). Preferences for each of the 12 capabilities were rated on a 0–100 visual analog (slider) scale, items presented in randomized order. After ratings, participants could add free-text comments explaining their reasoning. Data assumptions (normality of residuals, homogeneity of variance) were checked and met for the reported analyses. All data available at https://osf.io/renvt/.

Analysis: For RQ1, One-Way ANOVAs tested domain effects on mean agency and experience preference scores (averaged across their six items), with Tukey post hoc tests. For RQ2, Pearson correlation tested age effects; independent t-tests tested gender and computer science education effects; One-Way ANOVA with Tukey post hoc tested education level effects. Qualitative free-text responses (n=84 after cleansing) were thematically analyzed manually by the first author following Braun and Clarke’s approach.

Overall pattern: Participants generally preferred high agency capabilities and low experience capabilities across domains.

Domain effects (RQ1):

• Agency preferences: Significant effect of application domain, F(5,265) = 3.50, p = 0.005, partial η² = 0.081. Tukey tests showed lower agency preference for the household cleaning robot compared to surgical (mean difference = 15.8, t = 3.65, p = 0.004), military (16.8, t = 3.84, p = 0.002), companion (14.5, t = 3.45, p = 0.008), and educational robots (17.0, t = 4.04, p = 0.001). Service vs cleaning was not significant. Mean agency scores across types ranged 68.9–75.6, with cleaning robot an outlier at 58.5.
• Experience preferences: Significant effect of application domain, F(5,265) = 3.73, p = 0.004, partial η² = 0.068. Tukey tests showed higher experience preference for the social companion robot relative to service (mean difference = −16.5, t = −3.76, p = 0.003), cleaning (−16.7, t = −3.68, p = 0.004), and educational robots (−13.3, t = −2.99, p = 0.046); surgical vs companion was not significant. Mean experience scores across types ranged 24.9–28.6, with the companion robot an outlier at 41.6.

Demographic effects (RQ2):

• Age: No significant correlations with agency or experience preferences.
• Gender: Women showed a slightly stronger experience preference than men (women M = 33.3, SD = 21.7; men M = 27.0, SD = 22.3), t(268) = −2.33, p = 0.021. No gender difference for agency (women M = 71.9, SD = 20.0; men M = 70.5, SD = 21.0).
• Education level: Small effect on experience preference, F(2,268) = 3.30, p = 0.046, partial η² = 0.021. Participants with no university/college education had higher experience preference (M = 42.0) than those with Bachelor’s (M = 28.7), Tukey mean difference = −13.4, t = −2.37, p = 0.048. Master’s level M = 28.8. No effect on agency.
• Computer science education: Individuals with CS education showed lower experience preference (M = 24.8, SD = 22.1) than those without (M = 31.4, SD = 22.1), t(269) = −2.24, p = 0.026. No agency difference (CS M = 74.4, SD = 20.1; non-CS M = 69.6, SD = 20.8).

Qualitative themes (n = 84 analyzed responses):

• Preference for absence of feelings (reported by ~54.8% of commenters): desire for unemotional, objective, logical robots functioning as tools; emotions seen as unnecessary or unsettling.
• Preference for empathy without own feelings (~8.3%): desire for robots to understand and reason about human emotional/mental states (e.g., pain) without themselves feeling.
• Sympathy for the robot (~8.3%): some participants expressed concern for robots with emotions, implying they would merit respectful treatment; if emotions are present, preference to shield from negative emotions.
• Moral concerns (~19.0%): doubts about robots’ ability to navigate subjective morality; questions about whose moral framework to adopt and issues of accountability for moral decisions.

Findings show a robust preference for agency-related capabilities (e.g., planning, reasoning, goal understanding, moral evaluation) and low preference for experience-related capabilities (e.g., feeling happiness, pain) across domains. Qualitative data echoed this, emphasizing an ideal of objective, logical robots without emotions, yet capable of empathic understanding of humans. This aligns with prior work (Malle and Magar, 2017) while extending it by demonstrating domain-specific effects: cleaning robots evoked weaker agency preferences, and social companion robots evoked stronger experience preferences. The latter may reflect anthropomorphism of a benign, pet-like companion and a desire to avoid causing harm, consistent with links between perceived experience and moral patiency. Lower agency preference for cleaning robots may reflect perceived low responsibility or low-stakes consequences relative to surgical or military robots, consistent with prior associations between agency and responsibility attributions.

Demographic influences were small but notable: women and those without higher education showed higher experience preferences; individuals with computer science education showed lower experience preferences. Age showed no effect, though the age distribution was skewed toward younger adults, limiting sensitivity. Ethical reflections from participants highlight discomfort with robots that would possess or display emotions or consciousness, concerns about moral decision-making frameworks, and the need for accountability mechanisms. Overall, results suggest designers should emphasize agency capabilities broadly while calibrating the expression of emotional/experience-like features to the application domain and user sensitivities, prioritizing empathic understanding over genuine affect display.

This study contributes evidence that desired mental capabilities in robots are not uniform: people generally prefer high-agency, low-experience robots, with meaningful domain-specific deviations (lower agency for household cleaning, higher experience for social companionship). It also identifies user factors (gender, educational background, computer science education) associated with small differences in preferences, and provides qualitative insights explaining why users favor unemotional but empathically understanding robots and raising ethical concerns about morality and accountability.

Implications for design include prioritizing robust agency (planning, reasoning, goal understanding, explainability) across domains, carefully limiting the expression of emotional states, and tailoring capability expression to domain functions and perceived stakes. Future research should: (1) model empathic understanding as a potentially distinct capability beyond the agency/experience dichotomy; (2) test whether perceived responsibility and risk drive agency preferences across domains; (3) broaden sampling beyond a culturally homogeneous population; (4) examine longitudinal changes in preferences; and (5) explore how concrete robot embodiments and interaction modalities influence desired mental capabilities.

• Sampling and generalizability: Convenience sample primarily of Swedish-speaking university-affiliated participants; culturally homogeneous; age distribution skewed toward 18–25, limiting generalizability and power to detect age effects.
• Domain scope: Only six application domains were studied; other domains may elicit different preferences.
• Measurement reduction: The 22-item mind model was reduced to 12 capabilities, which may omit nuances and affect dimensional coverage (e.g., empathy as a distinct construct).
• Self-report and hypothetical scenarios: Preferences were elicited via imagined robots and slider scales, which may differ from in-situ behavior with real robots.
• Qualitative analysis: Thematic analysis was conducted by a single coder (first author), introducing potential interpretive bias.
• Between-subjects design: Each participant rated only one domain, limiting within-person comparisons across domains.