HOW PERCEIVED ANTHROPOMORPHISM SHAPES TRUST AND ADOPTION IN AI FINANCIAL ADVISORY — A USER TYPOLOGY AND DESIGN FRAMEWORK

УДК 004

Abstract

AI-powered financial advisors are increasingly deployed in high-stakes consumer contexts, yet user adoption remains uneven and the psychological mechanisms governing trust formation in these systems are poorly understood. This study investigates how perceived anthropomorphism shapes trust and adoption intention in financial AI, and for whom these effects are strongest. A four-condition between-subjects experiment (N = 210) compared user responses to financial AI interfaces framed as an algorithm, chatbot, basic AI, or humanlike AI. Perceived anthropomorphism, interaction quality, trust, and adoption intention were measured alongside demographic moderators. OLS path analysis revealed a robust mediation chain: perceived anthropomorphism predicted trust (B = 0.59, p < .001), which in turn predicted adoption intention (B = 0.65, p < .001), together explaining 57% of trust variance and 36% of adoption variance. Condition framing had negligible influence on perceived anthropomorphism (R² = .006), indicating a fundamental decoupling between design intent and user perception. Cross-tabulating age and financial knowledge identified four user segments with distinct trust-adoption profiles. Building on these findings, a five-principle Earned Anthropomorphism Design Framework is proposed. Results carry implications for financial AI interface design, trust calibration, and consumer protection.

Аннотация

ИИ-консультанты в сфере финансов получают всё более широкое распространение, однако уровень их принятия пользователями остаётся неоднородным, а психологические механизмы формирования доверия к таким системам изучены недостаточно. В настоящем исследовании анализируется, каким образом воспринимаемый антропоморфизм формирует доверие и намерение использовать финансовые ИИ-системы, а также для каких пользователей эти эффекты наиболее выражены. В рамках межгруппового эксперимента с четырьмя условиями (N = 210) сравнивались реакции пользователей на интерфейсы финансового ИИ, представленные как алгоритм, чат-бот, базовый ИИ или человекоподобный ИИ. Измерялись воспринимаемый антропоморфизм, качество взаимодействия, доверие и намерение использовать систему, а также демографические модераторы. Регрессионный анализ путей выявил устойчивую цепочку медиации: воспринимаемый антропоморфизм предсказывал доверие (B = 0,59; p < 0,001), которое в свою очередь предсказывало намерение использовать систему (B = 0,65; p < 0,001); совокупно эти переменные объясняют 57% дисперсии доверия и 36% дисперсии намерения. Условная формулировка интерфейса оказала пренебрежимо малое влияние на воспринимаемый антропоморфизм (R² = 0,006), что свидетельствует о фундаментальном разрыве между замыслом дизайнера и восприятием пользователя. Перекрёстный анализ возраста и уровня финансовой грамотности позволил выделить четыре пользовательских сегмента с качественно различными профилями доверия и принятия. На основе полученных данных предложена модель дизайна «заслуженного антропоморфизма», включающая пять принципов. Результаты имеют значение для проектирования интерфейсов финансовых ИИ-систем, калибровки доверия и защиты потребителей.

Keywords: financial AI advisory, perceived anthropomorphism, trust calibration, adoption intention, algorithm aversion, user segmentation, interface design framework, human-computer interaction.

Ключевые слова: финансовые ИИ-консультанты, воспринимаемый антропоморфизм, калибровка доверия, намерение использовать, неприятие алгоритмов, сегментация пользователей, модель дизайна интерфейса, взаимодействие человека и компьютера.

Introduction

Artificial intelligence is reshaping the financial services landscape at an unprecedented pace. Robo-advisors, conversational savings coaches, and algorithmic portfolio managers now handle interactions that were once exclusively the preserve of human financial professionals [1]. Despite this proliferation, adoption patterns remain heterogeneous and often counterintuitive: technically superior systems are rejected while simpler ones gain traction, and identical interfaces produce vastly different behavioural responses across user populations [2].

The dominant theoretical explanation draws on anthropomorphism research - the idea that users who perceive an AI as human-like are more likely to extend trust and act on its recommendations [3]. A substantial literature links anthropomorphic design cues such as conversational language, avatar presence, and named personas to improved user attitudes [4, 5, 6]. In financial contexts where decisions carry genuine economic consequences, the perceived nature of the advisor may be particularly consequential for trust formation and compliance [7].

Two critical gaps remain in this literature. First, most studies treat anthropomorphism as a design condition administered to users rather than as a perceptual state that users actually form during interaction. An earlier study by the authors found that anthropomorphic interface design affects user trust and adoption intention but left open the precise psychological mechanism linking interface framing to behavioural intent [24]. Chi and Vu [8] similarly found that anthropomorphic features do not generate trust unless they first produce perceptions of communication quality, suggesting the active ingredient is interaction experience rather than the framing label. Second, the user population is typically treated as homogeneous, obscuring important variation in who is most responsive to anthropomorphic cues.

This study addresses both gaps. The aim is to specify the psychological mechanism through which perceived anthropomorphism shapes trust and adoption intention in financial AI, to identify which users are most susceptible to this mechanism, and to translate these empirical patterns into actionable interface design guidance. To achieve this, we test a chained mediation model in a four-condition between-subjects experiment (N = 210), construct a data-driven user typology by cross-tabulating age and financial knowledge, and derive a five-principle design framework from the combined quantitative and qualitative findings.­

Materials and Methods

Theoretical Background and Hypotheses. Anthropomorphism emerges from three psychological antecedents [3]: accessible human-agent knowledge schemas, effectance motivation (the drive to predict an agent's behaviour), and sociality motivation (the desire for social connection). Conversational AI systems engage all three simultaneously [9]. A review of 74 empirical studies confirmed trust-enhancing effects of anthropomorphic cues while noting undertheorised risks including overtrust and perceived manipulation [11]. Within financial services, naming a robo-advisor meaningfully increased reliance on its recommendations [12], and anthropomorphic appeals were found to reduce algorithm aversion and increase investment intent [13].

Whether effects operate through the label or the experience is a key unresolved question. Riedl et al. [14] showed that behavioural cues exerted far stronger effects on perceived human-likeness than descriptive labels alone. A complementary body of research examines algorithm aversion — the tendency to prefer human over algorithmic advice even when algorithms demonstrably outperform humans [17]. Both excess trust and insufficient trust produce suboptimal financial outcomes [16], motivating the study of trust calibration. Wang et al. [1] found that human-likeness as a technological feature positively predicted AI financial advisor adoption but left the psychological mechanism unspecified. Chi and Vu [8] filled part of this gap by showing that anthropomorphism must first generate communication quality before it produces trust — motivating the inclusion of interaction quality as a parallel mediator. Overtrust risk was demonstrated by Cabitza et al. [21] and Scharowski et al. [19], carrying particularly acute implications for financial AI [23].

Based on the above, we advance five hypotheses. H1: Perceived anthropomorphism will positively predict trust. H2: Interaction quality will positively and independently predict trust. H3: Trust will positively predict adoption intention. H4: Condition framing will positively predict perceived anthropomorphism (exploratory; competing hypothesis: labels do not generate the perceptions they name). H5: The trust-to-intention relationship will be stronger among older users and those with lower financial knowledge.

Participants and Design. A between-subjects experiment assigned participants to one of four AI framing conditions. The analytical sample after excluding incomplete records comprised N = 210. Participants were predominantly aged 18-24 (74.8%), with 25-34 (15.2%), 35-44 (8.1%), and 45 or older (1.9%). Gender distribution was 52.4% male and 47.6% female. Financial knowledge was self-reported as Low (28.6%), Moderate (52.4%), High (14.8%), or at the extremes (4.3% combined). Condition membership was determined empirically by identifying which block of Likert items a participant completed, since the self-reported random routing token proved unreliable across data collection contexts.

To assess comparability of experimental groups on the key moderator variable, Table 1 presents the distribution of self-reported financial knowledge across conditions. The proportions are broadly similar across groups, with Moderate knowledge participants comprising the majority in all conditions. The Basic AI condition shows a somewhat higher proportion of Low-knowledge participants (45.8%) compared to other conditions, which reflects the unequal and non-random condition assignment method described above rather than a systematic design difference. This distributional variation is acknowledged as a limitation when interpreting the subgroup moderation results.

Table 1. Distribution of financial knowledge levels across experimental conditions

Procedure. All participants read a personalised financial scenario and then received a description of their assigned advisor. Condition 1 — Algorithm (n = 46): a data-driven system learning from financial history to produce personalised recommendations. Condition 2 — Chatbot (n = 51): a natural language system understanding the user's financial history and communication style. Condition 3 — Basic AI (n = 24): an AI understanding human preferences, adapting over time to specific needs. Condition 4 — Humanlike AI (n = 89): an AI sensing, reasoning, and adapting similarly to a human, with full personalisation.

Measures. Perceived anthropomorphism (6 items, alpha = .817): personality attribution, understanding of intentions, emotion perception, need comprehension, emotional attunement, and care for financial wellbeing (1 = Strongly Disagree to 5 = Strongly Agree). Interaction quality (3 items, alpha = .797): clarity, accuracy, and smoothness of the interaction. Trust (3 items, alpha = .817): perceived benevolence, comfort with future use, and overall confidence in the advisor. Adoption intention (1 item): likelihood of accepting and acting on the advice (1 = Very Unlikely to 5 = Very Likely). Open-ended responses were thematically analysed using inductive coding.

Analysis Strategy. Kruskal-Wallis tests examined condition differences given unequal group sizes. Mann-Whitney post-hoc tests with Bonferroni correction followed significant omnibus effects. OLS path analysis estimated the mediation model; indirect effects were assessed via 500-iteration bootstrap confidence intervals. Pearson correlations characterized subgroup relationships. For the user typology, participants were cross-tabulated on age band and self-rated financial knowledge (High vs. Low/Moderate), producing four segments whose trust-adoption profiles were compared. Thematic analysis of open-ended responses followed an inductive approach with iterative code refinement.

Results and Discussion

Descriptive Statistics and Scale Reliability. Table 2 shows descriptive statistics and Cronbach's alpha for all multi-item scales. Each scale exceeded the alpha = .70 adequacy threshold [22], supporting composite score use. Variable means were near the 5-point scale midpoint, indicating substantive variance for analysis.

Table 2. Descriptive statistics and scale reliability

PA = Perceived Anthropomorphism; IQ = Interaction Quality. Adoption intention is a single item.

Condition Effects (H4). Table 3 presents condition means and Kruskal-Wallis statistics. H4 was not supported: condition framing had no significant effect on perceived anthropomorphism (H = 1.06, p = .787), trust (H = 2.09, p = .554), or adoption intention (H = 4.62, p = .202). The only significant omnibus effect emerged for interaction quality (H = 9.33, p = .025). Post-hoc Mann-Whitney tests showed that the Basic AI condition (M = 4.04) was rated significantly higher than Chatbot (M = 3.56, p = .012), and Humanlike AI (M = 3.87) was similarly higher than Chatbot (p = .012). The Basic AI versus Algorithm comparison did not reach significance (p = .103).

This null result is theoretically consequential. Participants assigned to the Humanlike AI condition perceived the AI as no more human-like than those assigned to the Algorithm condition (M = 3.22 vs. M = 3.18). Describing an AI as human-like does not make users feel it is — a fundamental decoupling between design intent and user perception, consistent with Riedl et al. [14] who found that behavioural cues rather than descriptive labels drive perceived human-likeness.

Table 3. Means by condition and Kruskal-Wallis test statistics

PA = Perceived Anthropomorphism; IQ = Interaction Quality. + p < .05 vs. Chatbot (Mann-Whitney post-hoc)

Mediation Analysis (H1-H3). Table 4 shows OLS path coefficients. H1 was strongly supported: perceived anthropomorphism significantly predicted trust (B = 0.59, SE = 0.07, t = 9.13, p < .001). H2 was also supported: interaction quality independently predicted trust (B = 0.38, SE = 0.06, t = 5.85, p < .001), together explaining 57% of trust variance (R2 = .570). H3 was supported: trust strongly predicted adoption intention (B = 0.65, SE = 0.06, t = 10.88, p < .001, R2 = .363).

Bootstrap mediation (500 iterations, Humanlike AI vs. Algorithm contrast, n = 135) yielded non-significant indirect effects via perceived anthropomorphism (0.015 [95% CI: -0.092, 0.114]) and interaction quality (0.034 [-0.014, 0.106]). Condition assignment does not reliably activate the mediation chain because it does not alter the perception that initiates it. The full model explained R2 = .426 of adoption intention variance, with trust (B = 0.41, p < .001) and interaction quality (B = 0.34, p < .001) as dominant predictors.

Table 4. OLS path analysis results

Reference category: Algorithm. PA = Perceived Anthropomorphism. * p < .05; *** p < .001

Subgroup Moderation (H5). Table 5 presents within-subgroup Pearson correlations. H5 was supported. Older users (25+) showed a trust-to-intention correlation of r = .70 compared to r = .53 among younger users (18-24). Users with low or moderate financial knowledge showed r = .63 versus r = .43 for high-knowledge users. Users who cannot independently evaluate advice quality rely more heavily on affective trust as a proxy for output quality [19]. Older users in the Algorithm condition gave a mean adoption intention of 2.0 compared to 3.4 among younger users in the same condition — a 1.4-point gap indicating strong algorithm aversion consistent with the literature [2]. This aversion attenuated substantially in the Humanlike AI condition (M = 3.16 for 25+ vs. 3.63 for 18-24), suggesting that anthropomorphic framing partially offsets age-related algorithm aversion.

Table 5. Pearson correlations within subgroups

PA = Perceived Anthropomorphism. All correlations significant at p < .05 or lower unless noted.

Qualitative Themes. Inductive coding of 37 open-ended responses identified five themes. Calibrated scepticism was most prevalent: "you need to have a head on your shoulders and not completely trust AI." This reflects appropriate trust calibration [16] and appeared across all conditions. Affective dependency risk was the counterpoint: "trusting in AI is becoming my main thing in my usual life, and sometimes I think I'm useless without that," echoing Cabitza et al. [21] on systematic overreliance risk. Generic advice as credibility failure was prominent: "the answer seemed too vague — general truth but not easily applicable," pointing to a mechanism through which interaction quality mediates the anthropomorphism-trust pathway. Transparency and error intolerance: "the typos and references to the app which was not named clearly made me lose confidence," showing that minor presentation errors disproportionately damage trust in anthropomorphic conditions. Human-AI collaboration framing: "AI is responsible for processing data; human consultants for providing emotional support and bearing responsibility", well-calibrated trust stance that interface designers should support rather than undermine.

User Typology: Four Segments of Financial AI Adopters. Cross-tabulating age and financial knowledge produces four segments with qualitatively distinct adoption profiles (Table 6).

Table 6. Pearson correlations within subgroups

PA = Perceived Anthropomorphism. All correlations significant at p < .05 or lower unless noted.

Segment 1: Affective Adopters (Young, Low/Moderate Knowledge, n = 136) shows a moderate trust-to-intention link (r = .56) and responds most positively to richer AI personas — Basic AI and Humanlike AI yield the highest adoption intentions (M = 3.69 and 3.65). These users are driven by affective cues but retain some critical distance and represent the primary market baseline.

Segment 2: Analytical Evaluators (Young, High Knowledge, n = 21) shows a non-significant trust-to-intention link (r = .30, p = .193), indicating that affective trust does not reliably convert to adoption intent. These users prefer the Chatbot condition (M = 3.67) over Humanlike AI (M = 3.50) and apply additional deliberative evaluation before committing to adoption.

Segment 3: Sceptical Matchers (Older, Low/Moderate Knowledge, n = 39) shows the strongest trust-to-intention link (r = .75) and the most dramatic algorithm aversion — mean adoption intention of just 1.75 in the Algorithm condition. The Chatbot condition outperformed all others (M = 3.64). This segment constitutes the group most at risk of both undertrust under algorithmic framing and overtrust given high affective trust combined with low financial knowledge.

Segment 4: Cautious Experts (Older, High Knowledge, n = 14) shows the highest anthropomorphism-to-trust correlation in the study (r = .83) but a moderate non-significant trust-to-intention link (r = .51, p = .061). High knowledge appears to decouple trust from adoption even when trust is formed strongly through perceived human-likeness.

Earned Anthropomorphism Design Framework. The central empirical finding suggests a principle we term earned anthropomorphism: perceived human-likeness must be demonstrated through actual interaction behaviour rather than conveyed through descriptive framing. Building on this and the four segment profiles, we propose a five-principle framework (Table 7).

Table 7. Earned Anthropomorphism Design Framework

Principles are derived from empirical findings reported in this study; segment labels follow Table 6.

The framework moves toward segment-adaptive design, a direction consistent with emerging calls for personalised AI interfaces in high-stakes domains [14, 23]. Principle 5 deserves particular emphasis: among the 62 users (29.5% of the sample) combining high trust with low financial knowledge, mean adoption intention reached 3.87, the highest in the dataset. Interface-level safeguards such as embedded calibration prompts represent a design response that preserves user autonomy while protecting against the worst-case outcomes of overtrust. The overlap between the vagueness as failure qualitative theme and the interaction quality statistical result is remarkable: users and statistics converge on the same insight, that specificity and interaction fluency are the actual engines of AI credibility in financial contexts.

Conclusion

This study examined the psychological pathway through which anthropomorphic AI design translates into financial advisory adoption and developed a user typology and design framework grounded in empirical patterns. The mediation chain of perceived anthropomorphism, trust, adoption intention is well-supported and explains meaningful variance in behaviour. The finding that condition labels do not shift perceived anthropomorphism (R2 = .006) constitutes a theoretical contribution distinguishing design-level manipulation from psychological reality, with practical implications for how financial AI systems are positioned to users.

Four user segments show markedly different trust-adoption profiles and condition preferences, calling for segment-adaptive interface strategies. The five-principle Earned Anthropomorphism Design Framework translates these patterns into design guidance. For designers, the core implication is that human-like qualities must be earned through interaction rather than claimed through description. For regulators, the findings highlight the need for transparency standards ensuring calibrated trust in AI financial advisors, particularly for older and less financially literate populations. Limitations include unequal condition sizes, a predominantly young student sample, and the single-item adoption intention measure. The Basic AI condition contained a proportionally higher share of low-financial-knowledge participants (45.8% vs. 21.7–35.3% in other conditions), which may have influenced condition-level means for trust and adoption intention independently of the anthropomorphism manipulation. Future research should replicate with more balanced designs, community samples spanning wider age ranges, and longitudinal measures of actual financial behavior.

References:

1. Wang J., Zhang R., Luo Y. et al. An empirical study of AI financial advisor adoption through technology vulnerabilities in the financial context. Journal of Product Innovation Management, 2026. https://doi.org/10.1111/jpim.12795
2. Mahmud H., Islam A.K.M.N., Ahmed S.I., Smolander K. What influences algorithmic decision-making? A systematic literature review on algorithm aversion. Technological Forecasting and Social Change, 2022, vol. 175, 121390. https://doi.org/10.1016/j.techfore.2021.121390
3. Epley N., Waytz A., Cacioppo J.T. On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 2007, vol. 114, no. 4, pp. 864-886. https://doi.org/10.1037/0033-295X.114.4.864
4. Borau S., Otterbring T., Laporte S., Fosso Wamba S. AI robo-advisor anthropomorphism: The impact of anthropomorphic appeals and regulatory focus on investment behaviors. ResearchGate Working Paper, 2023. https://www.researchgate.net/publication/373583709
5. Yang Y., Liu Y., Lv X., Ai J., Li Y. Anthropomorphism and customers willingness to use artificial intelligence service agents. Journal of Hospitality Marketing and Management, 2022, vol. 31, no. 1, pp. 1-23. https://doi.org/10.1080/19368623.2021.1926037
6. Li Q., Luximon Y., Zhang J. The influence of anthropomorphic cues on patients perceived anthropomorphism, social presence, trust building, and acceptance of health care conversational agents. Journal of Medical Internet Research, 2023, vol. 25, e44479. https://doi.org/10.2196/44479
7. Choung H., David P., Ross A. Trust in AI and its role in the acceptance of AI technologies. International Journal of Human-Computer Interaction, 2023, vol. 39, no. 9, pp. 1727-1739. https://doi.org/10.1080/10447318.2022.2050543
8. Chi N.T.K., Hoang Vu N. Investigating the customer trust in artificial intelligence: The role of anthropomorphism, empathy response, and interaction. CAAI Transactions on Intelligence Technology, 2023, vol. 8, no. 1, pp. 260-273. https://doi.org/10.1049/cit2.12133
9. Waytz A., Cacioppo J.T., Epley N. Who sees human? The stability and importance of individual differences in anthropomorphism. Perspectives on Psychological Science, 2010, vol. 5, no. 3, pp. 219-232. https://doi.org/10.1177/1745691610369336
10. Cheng X., Zhang X., Cohen J., Mou J. Human vs. AI: Understanding the impact of anthropomorphism on consumer response to chatbots from the perspective of trust and relationship norms. Information Processing and Management, 2022, vol. 59, no. 3, 102940. https://doi.org/10.1016/j.ipm.2022.102940
11. Glikson E., Woolley A.W. Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals, 2020, vol. 14, no. 2, pp. 627-660. https://doi.org/10.5465/annals.2018.0057
12. Hodge F.D., Mendoza K.I., Sinha R.K. The effect of humanizing robo-advisors on investor judgments. Contemporary Accounting Research, 2021, vol. 38, no. 4, pp. 2755-2789. https://doi.org/10.1111/1911-3846.12694
13. Liu-Thompkins Y., Okazaki S., Li H. Artificial empathy in marketing interactions: Bridging the human-AI gap in affective and social customer experience. Journal of the Academy of Marketing Science, 2022, vol. 50, no. 6, pp. 1198-1218. https://doi.org/10.1007/s11747-022-00892-5
14. Riedl R., Kindermann H., Auinger A., Javor A. To trust or not to trust a human(-like) AI: A scoping review and conjoint analyses on factors influencing anthropomorphism and trust. Zeitschrift fur Arbeitswissenschaft, 2025. https://doi.org/10.1007/s41449-025-00481-6
15. Murtaza G., Yousaf M., Iqbal M. et al. Unveiling trust in AI: The interplay of antecedents, consequences, and cultural dynamics. AI and Society, 2025. https://doi.org/10.1007/s00146-025-02477-6
16. Gsenger R., Strle T. Trust, automation bias and aversion: Algorithmic decision-making in the context of credit scoring. Interdisciplinary Description of Complex Systems, 2021, vol. 19, no. 4, pp. 540-558.
17. Dietvorst B.J., Simmons J.P., Massey C. Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 2015, vol. 144, no. 1, pp. 114-126. https://doi.org/10.1037/xge0000033
18. Schaap G., van den Broek T.C., Janssen M. The ABC of algorithmic aversion: Not agent, but benefits and control determine the acceptance of automated decision-making. AI and Society, 2023. https://doi.org/10.1007/s00146-023-01649-6
19. Scharowski N., Perrig S.A.C., Svaton M. et al. Trust and reliance on AI: An experimental study on the extent and costs of overreliance on AI. Computers in Human Behavior, 2024, vol. 158, 108309. https://doi.org/10.1016/j.chb.2024.108309
20. Horowitz M.C., Kreps S.E., Fuhrmann-Rieger M. Bending the automation bias curve: A study of human and AI-based decision making in national security contexts. International Studies Quarterly, 2024, vol. 68, no. 2, sqae020. https://doi.org/10.1093/isq/sqae020
21. Cabitza F., Campagner A., Bassi M. The need to consider the risks of overreliance on AI systems for high-stakes decision-making. HHAI 2023: Augmenting Human Intellect. IOS Press, 2023. https://doi.org/10.3233/FAIA230080
22. Hair J.F., Black W.C., Babin B.J., Anderson R.E. Multivariate Data Analysis (7th ed.). Pearson, 2010.
23. Floridi L., Cowls J., Beltrametti M. et al. An ethical framework for a good AI society: Opportunities, risks, principles, and recommendations. Minds and Machines, 2018, vol. 28, no. 4, pp. 689-707. https://doi.org/10.1007/s11023-018-9482-5
24. Akhmetova M., Gutoreva A. Anthropomorphized interfaces and user perception in financial AI assistants. 2026 18th International Conference on Electronics, Computer, and Computation (ICECCO). IEEE, Kaskelen, Kazakhstan, 10-11 April 2026. https://doi.org/10.1109/ICECCO67619.2026.11488763