- The paper finds that 89% of participants had encountered an AI agent they considered socially intelligent, basing judgments mainly on conversational fluency, context awareness, emotion recognition, empathy, and helpfulness rather than internal agency.
- The paper identifies a persistent support-adoption gap: participants supported Social-AI services for others more than they would personally use them by 0.20 points on average, even though higher AI literacy increased overall acceptability.
- The paper shows strong preferences for privacy-preserving deployment, including active rather than continuous sensing, speech rather than video, local rather than cloud storage, and autonomous rather than teleoperated home robots.
Overview and motivation
This paper reports a mixed-methods survey of 200 adults in the United States examining how laypeople perceive social intelligence in AI agents, what abilities signal that construct to them, how acceptable they find Social-AI deployments across contexts, and what concerns they hold. The work is motivated by a participatory-AI framing: Social-AI agents are increasingly deployed in healthcare, education, hospitality, and service settings for layperson users, yet design and deployment decisions are made by a small set of researchers and developers. The authors treat social intelligence as a perceived construct rather than purely a technical property, building on the "computers are social actors" tradition, and argue that deployment and governance decisions should be informed by public perception alongside technical capability.
The survey addresses four research questions: (RQ1) the extent to which the public believes current AI agents have social intelligence; (RQ2) which abilities convince the public an agent is socially intelligent; (RQ3) how contextual factors—setting, stakes, agent role, embodiment—influence acceptability; and (RQ4) what risks or harms concern the public.
Methodology
Participants were recruited via Prolific (N=200; 53% men, 45% women; ages 18–78, M=40.9), with a mean completion time of 19.53 minutes and compensation at $12/hour. The instrument comprised five blocks: prior interactions with agents perceived as socially intelligent (Block A); ratings of current AI capabilities on an eight-item Perceived AI Social Intelligence Measure (Block B); acceptability ratings of 12 scenario conditions (Block C); open-ended concerns plus structured data-preference questions (Block D); and demographics, AI experience, and a seven-item AI Literacy Scale (Block E).
The scenario design crossed four contexts—defined by setting (home vs. hospital), relative stakes (lower vs. higher), and a concrete agent role (work support, fall alert, wayfinding, distress alert)—with three agent types (chatbot, autonomous robot, teleoperated robot). For each scenario, participants rated three statements: willingness to seek the service for themselves, willingness to avoid it, and gladness that it existed for others. This triad was designed to separate personal adoption from broader support, a design choice that proved central to the paper's key finding. Participants were instructed to assume human-level competence, isolating acceptability judgments from capability skepticism. Notably, no non-Social-AI control condition was included, so the study does not address whether Social-AI is preferred over alternative service providers.
Both multi-item scales showed strong internal consistency (Perceived AI Social Intelligence: Cronbach's $\alpha$ = .896, McDonald's $\omega=.901;AILiteracy:\alpha$ = .801). Quantitative analysis used linear mixed-effects models with participant random intercepts, Wilcoxon signed-rank tests, Spearman correlations, and McNemar tests. Open-ended responses were coded with inductively developed multi-label codebooks; human-human agreement was high (Krippendorff's $\alpha=0.87–0.89),andcodebook−constrained<ahref="https://www.emergentmind.com/topics/llm−assisted−coding"title=""rel="nofollow"data−turbo="false"class="assistant−link"x−datax−tooltip.raw="">LLM−assistedcoding</a>(claude−opus−4−7)achievedstrongagreementwithhumans(\alpha=0.82–0.84).Thewithin−participantscenariodesignyields2,400acceptabilityratings,poweringdetectionofstandardizedmeandifferencesaroundd_z \approx 0.20.</p><h2class=′paper−heading′id=′perceived−social−intelligence−in−current−ai−agents′>PerceivedsocialintelligenceincurrentAIagents</h2><p>Astrikingheadlineresultisthat<strong>89\rho=.367,versus\rho$ = .636 for emotion recognition).
The authors draw a consequential implication: while "social intelligence in AI" is often framed as a long-term research goal, laypeople already attribute the construct to deployed systems based on observable interaction behaviors rather than beliefs about internal agency or intent. Given prior evidence that sociable AI shapes mental-capacity attributions and can induce overtrust, this suggests governance challenges will arise before systems satisfy robust scientific definitions of social intelligence.
Qualitative coding of RQ2 responses showed participants reason about social intelligence as multi-faceted: the most prevalent codes were navigating conversations (36.0%), humanlike communication (31.5%), current context awareness (30.0%), recognizing emotions (29.5%), expressing empathy (27.0%), proactive helpfulness (25.0%), and task competence (23.5%). Only two participants rejected the premise entirely, grounding their rejection in the view that social intelligence requires genuine inner experience—a minority position indicating most judgments rest on behavior rather than phenomenology. The authors propose that future Social-AI evaluation should separately measure (1) behaviors users perceive during interaction and (2) abilities users attribute as a result, since benchmark performance and perceived social intelligence are not interchangeable.
Acceptability and the support-adoption gap
All 12 scenarios were rated above the neutral midpoint, but acceptability varied more across scenario contexts (mean range 0.15) than across agent types (mean range 0.05). The highest-rated condition was the high-stakes home fall alert (M = 0.74 averaged over agent types; chatbot version M = 0.77); the lowest was the low-stakes home work support (M = 0.58; teleoperated robot version M = 0.54). Because setting, stakes, and role were bundled within each context, the paper cannot isolate which contextual feature drove this effect—an acknowledged limitation.
The central discovery is a support-adoption gap: across all 12 scenarios, participants supported the existence of Social-AI services for others substantially more than they would personally adopt them. Mean gaps ranged from 0.14 to 0.23 on the normalized scale; a mixed-effects model confirmed support ratings exceeded adoption ratings by 0.20 points (p < .001), with all scenario-level Wilcoxon tests surviving Holm correction. This finding has direct methodological implications: public-opinion surveys and impact assessments that use general support as a proxy for personal adoption willingness will systematically overstate acceptance.
Higher AI literacy was associated with higher overall acceptability (Spearman ρ = 0.395, FDR-adjusted p < .001), with mean acceptability rising from 0.57 in the lowest literacy quartile to 0.79 in the highest. However, literacy reduced but did not eliminate the gap (0.17 vs. 0.24 between highest and lowest quartiles), and the participant-level literacy–gap association was not significant (ρ = −0.087, p = .223). The authors conclude that personal reluctance toward Social-AI is not merely a knowledge deficit correctable through education. Similarly, participants who rated current AI as more socially intelligent showed higher acceptability (ω0 = 0.477) and smaller gaps (ω1 = −0.148), linking perception and acceptance.
Public concerns and data preferences
Open-ended concern coding revealed that the most frequent substantive concerns were reduced human connection (16.5%), data privacy (16.5%), overreliance (14.0%), data misuse (11.5%), accountability/governance (10.0%), manipulation (10.0%), and hallucination/incorrect responses (10.0%). A substantial 36.5% explicitly reported no concerns; because the survey required a written statement rather than allowing blanks, these are treated as deliberate signals. The no-concerns group had a distinct profile—higher acceptability (0.74 vs. 0.61), smaller gaps (0.15 vs. 0.23), and higher perceived AI social intelligence (3.73 vs. 3.38)—supporting the interpretation that these responses reflect favorable orientation rather than disengagement.
Structured privacy questions about in-home embodied agents produced clear preferences:
| Condition |
Speech continuous |
Speech active |
Video continuous |
Video active |
| Autonomous / local |
23.0% |
76.5% |
19.0% |
63.0% |
| Autonomous / cloud |
9.0% |
68.5% |
6.0% |
46.0% |
| Teleoperated / local |
13.5% |
66.0% |
10.0% |
52.0% |
| Teleoperated / cloud |
6.5% |
61.0% |
4.5% |
42.0% |
Active-interaction sensing was strongly preferred over continuous sensing across all conditions (all McNemar tests significant after Holm correction), speech was preferred over video, local storage over cloud storage, and autonomous robots over teleoperated ones. The least acceptable configuration—continuous video sensing by a teleoperated robot with cloud storage—was selected by only 4.5% of participants. These results carry concrete engineering implications: home-deployed Social-AI should offer user-controllable sensing modalities and storage locations, motivating on-device social perception algorithms that reduce reliance on cloud processing of sensitive interaction data. The aversion to teleoperation implies governance frameworks should distinguish teleoperated from autonomous agents rather than treating physical robots as a single category.
Limitations
The sample consists of English-speaking US adults recruited through Prolific at a single point in time, limiting generalizability to national public opinion and to future populations as deployment patterns evolve. Definitional scaffolding provided before the survey may have shaped how participants interpreted social intelligence; the authors suggest pre/post-scaffolding comparisons as future work. Because scenario contexts bundled setting, stakes, and agent role, the study cannot isolate independent contextual effects. Additionally, participants' prior Social-AI exposure was overwhelmingly chatbot-based (72.5% had never used a physical robot), so embodiment-related findings reflect stated preferences rather than experienced interaction. Low-prevalence qualitative codes (n < 10) are exploratory.
Conclusion
This survey establishes that social intelligence is already a construct laypeople attribute to AI agents—grounded primarily in observable conversational and affective behaviors rather than attributed agency—and that public acceptability of Social-AI depends more on deployment context than on embodiment. Its most transferable contributions are the demonstration of a persistent support-adoption gap that survives differences in AI literacy, and quantified privacy preferences showing strong resistance to continuous audio/video sensing and cloud storage in home robots. Together, these findings argue for governance processes and evaluation practices that distinguish public support from personal adoption, and for Social-AI systems whose data-collection behavior matches the narrow, user-controlled sensing envelope the public is willing to accept.