How Far Do Capability Cues Travel? Anthropomorphism and Differentiated Trust in a Platform-Embedded AI Assistant

1 Introduction

• The study distinguishes perceived multifunctionality, perceived intelligence, and anthropomorphism, and treats ability, integrity, and benevolence as separate trust judgments.The trust dimensions concern competence, principled conduct, and attention to users’ interests.Source: §1 · p. 2
• A randomized 2 × 2 experiment with 270 U.S.-based Reddit users varied whether an embedded assistant displayed one or three functions and whether it provided a brief rationale.The post, comments, and engagement indicators were held constant across conditions.Source: §1 · p. 2
• Displaying three functions increased perceived multifunctionality, but no other randomized main effect survived correction across the focal outcomes, and rationale did not reliably increase perceived intelligence.Exploratory analysis indicated stronger uptake of the functional display among participants with higher objective AI literacy.Source: §1 · p. 2
• The paper argues that visible functionality and rationales have limits as capability signals, and that capability perceptions and social attribution carry different trust profiles.It recommends evaluating the specific judgment a feature is intended to change and the trust dimension at stake.Source: §1 · pp. 2–3
• Perceived intelligence was associated with anthropomorphism and all three trust dimensions, while anthropomorphism was additionally associated with benevolence, but not reliably with integrity or ability.These relationships are associational, not randomized effects.Source: §1 · p. 2

2.1 Reading Capability from Interface Cues

• Visible interface cues can inform capability judgments, but displayed functions and users’ perceived functional breadth are distinct; explanation availability may not equal explanation usefulness.The setting concerns a bounded presentation rather than learning through repeated interaction or performance changes over time.Source: §2.1 · p. 3
• Prior work links transparency and explainability with cognitive evaluations, while the direction and location of differences in cue interpretation by AI knowledge remain unsettled.Studies also suggest capability displays may not be taken up equally by all users.Source: §2.1 · p. 3

2.2 Perceived Intelligence and Anthropomorphism

• Perceived intelligence concerns cognitive capability, whereas anthropomorphism concerns personlike characteristics, motivations, or intentions; prior studies report positive associations between them.The literature treats these as distinct but potentially related judgments.Source: §2.2 · p. 4

2.3 Trust as a Di!erentiated Judgment

• Trust can be multidimensional, and prior findings do not establish a uniform mapping from cognitive and social perceptions to ability, integrity, and benevolence.Some social-cue effects have been concentrated in benevolence without corresponding changes in ability or integrity.Source: §2.3 · p. 4

2.4 AI Assistants Embedded in Platform Contexts

• Research on embedded AI has examined interpersonal messages, repeated community participation, and downstream opinion formation more than users’ immediate judgments of an embedded assistant.The paper focuses on the immediate evaluative context within a platform encounter.Source: §2.4 · p. 5

3 Hypothesis Development and Research "estion

• The framework progresses from experimentally assigned interface cues through capability perceptions and anthropomorphism to differentiated trust, reserving exploratory status for supplementary analyses.Directional hypotheses are used when prior theory supports a prediction, and a research question when it does not support a particular pattern.Source: §3 · p. 5

3.1 Multifunctionality as a Capability Cue

• The multifunctionality section distinguishes displayed breadth from users’ perceived breadth, predicts that three displayed functions will raise perceived multifunctionality, and predicts a positive association with intelligence.The assigned-versus-perceived distinction reflects the possibility that users interpret the same display differently.Source: §3.1 · p. 5

3.2 Explanation as a Capability Cue

• The explanation hypothesis predicts higher perceived intelligence when a rationale is available, on the premise that visible reasoning can provide evidence of cognitive capability.The rationale’s value may depend on its content and audience.Source: §3.2 · pp. 5–6

3.3 Perceived Intelligence and Anthropomorphism

• Perceived intelligence and anthropomorphism are treated as distinct judgments, with H4 predicting a positive association between cognitive competence and personlike attribution.The cited rationale is that observable evidence of cognitive capability may contribute to humanlike interpretations.Source: §3.3 · p. 6

3.4 Perceived Intelligence, Anthropomorphism, and Di!erentiated Trust

• Because prior findings do not establish a consistent dimension-specific trust profile, RQ1 asks how intelligence and anthropomorphism are differentially associated with ability, integrity, and benevolence when estimated together.The three trust dimensions are retained as separate outcomes.Source: §3.4 · p. 6

3.5 Supplementary Exploratory Analysis of Randomized E!ects

• The supplementary exploratory analysis estimates assigned multifunctionality and explanation main effects, plus their interaction, across all six latent outcomes.Effects corresponding to H1 and H3 are interpreted in relation to those hypotheses; the remaining randomized paths and interactions are exploratory.Source: §3.5 · p. 6

4.1 Experiment Design

• The study used a randomized 2 × 2 between-participants design varying AI functionality (single-function vs. multifunctional, with three functions displayed in the latter) and whether an explanation was present or absent.No live model generation occurred during study sessions; the extension rendered the corresponding pre-generated outputs locally according to each participant’s randomized condition.Source: §4.1 · pp. 6–7

4.2 Experiment Procedure

• Participants reviewed the same designated Reddit discussion, rendered for their assigned condition, then completed a post-survey measuring perceptions, trust, AI literacy, and manipulation checks.The stimulus concerned an article about updated COVID-19 vaccines and severe health outcomes; the discussion content was simulated and prepared before data collection.Source: §4.2 · pp. 8–9

4.3 Ethical Considerations

• With informed consent, participants were temporarily not told the hypotheses or that the Reddit encounter was simulated, and were debriefed afterward.The study system kept activity off Reddit and did not collect login credentials; Lehigh University’s Institutional Review Board approved the procedures.Source: §4.3 · p. 9

4.4 Recruitment and Participants

• The final analytic sample comprised 270 participants from 283 after 11 attention-check and 2 rendering exclusions, exceeding the 210 participants required for 95% power; no significant differences across four conditions were detected in age, gender, or education.Retained condition cell sizes ranged from 64 to 74.Source: §4.4 · pp. 9–10

4.5 Experimental Manipulations

• Assigned multifunctionality had two levels: the interface displayed one distinct function or multiple distinct functions on the Reddit post.The post-survey included a perceived-multifunctionality scale and a factual check asking participants how many distinct functions the assistant performed.Source: §4.5.1 · p. 10

4.5.1 Assigned Multifunctionality.

• The assigned-functionality manipulation contrasted one displayed function with multiple functions, and participants were asked how many the assistant performed.The survey assessed both perceived multifunctionality and factual recognition of the manipulation.Source: §4.5.1 · p. 10
• 90.0% of participants correctly identified the displayed functional breadth, and analyses retained randomized assignment regardless of recognition responses.Recognition was strongly associated with assigned condition (Cramér’s V = .806, p < .001).Source: §4.5.1 · p. 10

4.5.2 Assigned Explanation.

• 82.6% correctly identified whether a rationale was displayed, and analyses retained randomized assignment regardless of recognition responses.Recognition was strongly associated with assigned condition (Cramér’s V = .661, p < .001), but was imperfect.Source: §4.5.2 · p. 10

4.6.1 Perceived Multifunctionality.

• Perceived multifunctionality used four adapted items from Lee and Shin on a seven-point agreement scale, modeled as one latent construct.Items asked whether the assistant had multiple functions and performed multiple tasks.Source: §4.6.1 · p. 10

4.6.2 Perceived Intelligence.

• Perceived intelligence used the five-item Godspeed dimension [2], retaining its semantic-differential pairs on five-point scales as indicators of a latent factor.The referent was adapted to the AI assistant.Source: §4.6.2 · p. 11

4.6.3 Anthropomorphism.

• Anthropomorphism adapted four items from Chen et al. and Choi et al., used seven-point responses, and was modeled as a latent factor.Example items concerned whether the assistant felt like a person and had its own personality.Source: §4.6.3 · p. 11

4.6.4 Trust.

• Trust items adapted from Wang and Benbasat measured benevolence, integrity, and ability as separate latent factors with seven-point responses.The three dimensions used three, three, and five items, respectively; residual covariances among the factors were freely estimated.Source: §4.6.4 · p. 11

4.6.5 Objective AI Literacy.

• Objective AI literacy used a 17-item knowledge test adapted from Tully et al., scored from 0 to 17, with observed scores ranging from 6 to 17.Items assessed AI applications, machine learning, data interpretation, sensing, programmability, and ethical issues.Source: §4.6.5 · p. 11

4.7 Data Analysis Methods

• Analyses estimated a six-factor measurement model and a latent-outcome factorial model, adjusting randomized main effects and interactions across separate test families.The structural SEM used MLR with FIML; literacy moderation was exploratory, WLSMV was a robustness analysis, and only randomized effects permit causal interpretation.Source: §4.7 · p. 11

5.1 Measurement Model and Discriminant Validity

• The six-factor model fit well; all 15 construct pairs passed the Fornell–Larcker criterion and remained below the .90 HTMT threshold.The largest HTMT value was .676, for ability and integrity.Source: §5.1 · p. 12

5.2 Experimental E!ects on Latent Outcomes

• 𝑈= .402: displaying multiple functions increased perceived multifunctionality, the only randomized main effect surviving correction; explanation did not reliably increase perceived intelligence.No multifunctionality-by-explanation interaction was significant.Source: §5.2 · p. 12

5.3 Structural Associations among Capability, Anthropomorphism, and Trust

• Perceived multifunctionality was positively associated with intelligence; intelligence was associated with anthropomorphism and all three trust dimensions, while anthropomorphism added an association only with benevolence.Paths among post-exposure perceptions reflect concurrent associations, not temporal or causal order.Source: §5.3 · p. 13

5.4 Exploratory Moderation by Objective AI Literacy

• Objective AI literacy strengthened the association between assigned multifunctionality and perceived multifunctionality; this same-sample exploratory finding needs confirmation in an independent sample.The moderation was the only one of 10 screened interactions to remain significant after correction.Source: §5.4 · p. 15

5.5 Ordinal-Indicator Robustness Check

• WLSMV preserved the principal profile: intelligence remained associated with all trust dimensions, while anthropomorphism’s additional association remained specific to benevolence.The rationale–intelligence association remained nonsignificant, and the intelligence–anthropomorphism association remained positive.Source: §5.5 · p. 15

6.1 Recognizing functional breadth and judging intelligence

• Three rather than one displayed function increased perceived multifunctionality (𝑈= .402, 𝑇< .001), the only randomized main effect surviving correction; its intelligence effect was smaller (𝑈= .128, 𝑄= .043, 𝑇= .172).90.0% correctly recognized assigned breadth, and the effect on perceived multifunctionality increased with objective AI literacy (𝑈= .264, 𝑇< .001).Source: §6.1 · p. 16

6.2 When a rationale serves as evidence of intelligence

• Although 75.4% reported seeing the rationale, its estimated effect on perceived intelligence remained near zero in factorial and structural analyses.The study measured rationale availability, not its perceived usefulness or diagnostic value.Source: §6.2 · p. 17

6.3 Capability judgments as material for personlike a#ribution

• Perceived intelligence was positively associated with anthropomorphism under MLR (𝑈= .393) and WLSMV (𝑈= .547), but concurrent measurement does not establish causation.Anthropomorphic presentation was held constant across conditions, so personlike construal covaried with perceived cognitive capability.Source: §6.3 · p. 17

6.4 Personlike a#ribution concerns whose interests the assistant serves

• Perceived intelligence was associated most strongly with ability-based trust (𝑈= .651), followed by integrity (𝑈= .493) and benevolence (𝑈= .364).Wald comparisons supported that ordering, including integrity exceeding benevolence (𝑇= .034).Source: §6.4 · p. 17
• After accounting for perceived intelligence, anthropomorphism was associated with benevolence (𝑈= .335), but not reliably with ability or integrity.Its associations with ability and integrity were individually nonsignificant (both 𝑇< .001).Source: §6.4 · pp. 17–18

6.5 Evaluating AI as an interpreter of platform content

• In the controlled Reddit setting, surrounding posts and community signals stayed stable while the assistant’s display and rationale varied; evaluation should distinguish competence from apparent orientation toward users’ interests.The authors frame the assistant as an interpreter of platform content, warranting assessment of both its ability to interpret a post and its apparent concern for readers.Source: §6.5 · p. 18
• Future work could test whether agreement with an assistant’s interpretation bears on confidence in its competence, while judgments about whose interests it serves bear on benevolence.These proposed tests connect embedded assistants to social evidence in platform discussions and extend the differentiated trust pattern.Source: §6.5 · p. 18

6.6 Limitations and future work

• The experiment captured initial evaluations of fixed, pre-generated outputs with limited interaction; repeated tasks, behavioral reliance, and continued use could show how judgments develop.Participants had limited opportunity to test capabilities through interaction.Source: §6.6 · p. 19
• Because focal perceptions were measured in one post-exposure survey, shared method variance and uncertain temporal order limit causal interpretation of structural paths.The authors suggest independently manipulating capability and anthropomorphic presentation and separating measurements over time.Source: §6.6 · p. 19
• The multifunctional condition bundled a function-count label with extra content, so its effect concerns that display package rather than functional breadth alone.A follow-up could hold content volume constant while varying breadth; the explanation result concerns one brief rationale, with perceived explanation quality unmeasured.Source: §6.6 · p. 19
• A single COVID-19 vaccination post in Old Reddit and a U.S. Prolific sample limit transfer to other topics and contemporary layouts; anthropomorphism was relatively low (N = 2.61 on a seven-point scale).Prior topic attitudes may matter.Source: §6.6 · p. 19
• Because literacy moderation was identified and estimated in the same comparatively high-scoring sample, replication across broader AI knowledge levels is needed.Source: §6.6 · p. 19

7 Conclusion

• Displayed additional functions reliably increased perceived functional breadth, while other randomized changes were limited after correction.Source: §7 · p. 19
• Perceived intelligence was associated most strongly with ability-based trust and also with integrity and benevolence.Source: §7 · p. 19
• Anthropomorphism had an additional association concentrated in benevolence.Source: §7 · p. 19
• For platform-embedded AI, users judge both the assistant’s capacity and its apparent orientation toward their interests.Source: §7 · p. 19
• Future work can test whether community feedback shapes judgments of ability, integrity, and benevolence differently.Source: §7 · p. 19

A Appendix for the measures used in the Study

• The appendix includes item wording for the six constructs in the retained SEM.Source: A Appendix for the measures used in… · p. 23
• It also includes the wording for objective AI literacy.Source: A Appendix for the measures used in… · p. 23
• The reproduced wording reflects the items presented to participants.Source: A Appendix for the measures used in… · p. 23

A.1 Perceived Multifunctionality

• Perceived multifunctionality was assessed with four statements about the assistant’s number of functions, task range, and functional needs.Source: §A.1 · p. 23

A.2 Perceived Intelligence

• Perceived intelligence was rated on five adjective pairs, including incompetent–competent and unintelligent–intelligent.Source: §A.2 · p. 23

A.3 Anthropomorphism

• Anthropomorphism items asked whether the assistant felt like a person and had a personality or intentions.Source: §A.3 · p. 23

A.4 Trust Dimensions

• Trust dimensions were assessed with agreement statements and treated as benevolence, integrity, and ability item sets.Source: §A.4 · p. 23, §A.4.1 · p. 23, §A.4.2 · p. 23

A.4.1 Benevolence.

• Benevolence items address the assistant’s concern for users’ best interests and helpfulness.Source: §A.4.1 · p. 23

A.4.2 Integrity.

• Integrity items ask whether the assistant provides unbiased information, is honest, and has integrity.Source: §A.4.2 · pp. 23–24

A.4.3 Ability.

• Ability items assess the assistant’s expertise, capacity to identify important information, and helpfulness about the Reddit post.Source: §A.4.3 · p. 24

A.5 Objective AI Literacy

• Objective AI literacy was automatically scored in Qualtrics, with one point per correct response and a total score from 0 to 17.Source: §A.5 · p. 24

• All pairs passed both checks: each absolute latent correlation was below the smaller square root of AVE, and each HTMT value was below .90.Source: Table 4 · p. 25, B Full discriminant-validity results… · p. 25

• Each interaction was screened in a separate OLS regression using standardized composite scores and HC3 robust standard errors, with Benjamini–Hochberg adjustment across 10 tests.Source: C Exploratory Objective AI Literacy… · p. 25, Table 5 · p. 26
• Only assigned multifunctionality × objective AI literacy predicting perceived multifunctionality remained significant after correction; it was then re-estimated in the latent model.Source: C Exploratory Objective AI Literacy… · p. 25, Table 5 · p. 26

• Only assigned multifunctionality and assigned explanation varied; post content, engagement indicators, and discussion materials were held constant across conditions.Source: D Full Experimental Interface · p. 26
• The three-function, explanation-absent interface displayed topic tagging, a summary, and a fact-check, with study materials presented locally.Source: D Full Experimental Interface · p. 26, Figure 4 · p. 27