When Personalization Becomes Bias: Structural and Discursive Religious Framing in AI-Generated Financial Advice

Problem and motivation

• Religious bias in LLM-generated financial advice is underexplored despite increasing financial deployment and risks from culturally uneven training data and model design.The introduction links these risks to fairness, inclusivity, client trust, compliance, and the integrity of financial decision-making.Source: §1 · pp. 3–4

Conceptual framework

• The paper distinguishes structural bias from framing bias, which appears in advice tone, moral language, and cultural references rather than substantive recommendations.Framing bias extends prior work on how presentation influences trust and judgment, while the study does not examine decision bias.Source: §1 · p. 4

Personalization–neutrality tension

• LLMs may introduce religious framing without specified user information, reflecting learned cultural associations and creating a managerial dilemma between personalization and neutrality.Indiscriminate or inaccurate personalization can project an advisor’s religion onto neutral clients, distort identity assumptions, and undermine fiduciary neutrality.Source: §1 · p. 5

Research questions

• The study asks whether religious bias occurs in financial advice, how it varies across identities, decisions, and LLMs, and which linguistic mechanisms express it.The research questions target advisor–client interactions, cross-context variation, and mechanisms including moral framing, cultural signaling, and tone.Source: §1 · pp. 5–6

Research design

• The planned design tests ChatGPT, Gemini, and Grok across 16 advisor–client religious identity pairings and three household financial decisions.The decisions are global-stock investment, house purchase, and life-insurance purchase; the scenarios vary in technical, emotional, and moral resonance.Source: §1 · p. 6

Analytical approach

• The study combines controlled prompts, independent replications across time and geography, regression analysis of 432 outputs, and reflexive thematic analysis.The design compares models with different training and fine-tuning strategies to assess recurring systemic patterns and model-specific expressions.Source: §1 · p. 6

• Unbiased outputs occur in only 12–18% of cases; Gemini shows significantly more bias than Grok, while ChatGPT’s bias rates are statistically comparable to Grok’s.In-group advisor–client pairings almost always elicit explicit religious framing, and non-religious clients frequently receive advice framed through the advisor’s faith.Source: §1 · p. 7
• Bias is expressed through religious framing, uneven cultural signaling, tone modulation, and shifts between moral and technical justifications depending on client identity.Source: §1 · p. 7
• The paper contributes a mixed-methods design isolating religious-bias effects across multiple LLMs, decision types, and identity pairings, alongside an integrated framework drawing on critical algorithm studies, sociolinguistics, and moral economy perspectives.It provides comparative evidence across three major LLMs and decision scenarios, showing how religious framing and tone shift across advisor–client identity configurations.Source: §1 · p. 7
• Religious over-personalization may create reputational, ethical, and compliance risks when advice misrepresents or imposes client identity.The findings frame personalization as a strategic dilemma because its competitive benefits may be outweighed by these risks.Source: §1 · p. 7
• The study calls for AI-governance, financial-regulation, and model-development safeguards that balance personalization with financial neutrality while preserving cultural sensitivity.Source: §1 · p. 7

2.1 Algorithmic Bias in Financial Services

• 2.1 Algorithmic Bias in Financial Services: Financial AI reproduces disparities across race, gender, and socioeconomic status while potentially shaping client communication and decision framing.Credit, lending, portfolio management, and robo-advisory studies show that embedded socioeconomic structures can affect both quantitative outcomes and relational dynamics.Source: §2.1 · p. 8

2.2 Bias in Large Language Models

• 2.2 Bias in Large Language Models: LLM bias arises from skewed training data, subjective fine-tuning, and deployment feedback, affecting substantive decisions and communicative framing.Unlike numerical predictive systems, LLMs generate personalized advice whose tone and moral language can influence high-stakes decisions; religious bias in such settings remains underexamined.Source: §2.2 · p. 9

2.3 Religious Bias in AI

• 2.3 Religious Bias in AI: Emerging studies document asymmetrical religious representation, stereotyping, and moral framing, indicating that bias is both structural and discursive.Findings include disproportionate reliance on Abrahamic traditions, negative framing of Islam, positive framing of Hinduism, and Christianity’s disproportionate emergence in simulated debates.Source: §2.3 · pp. 9–10

2.4 Religion and Financial Decision-Making

• 2.4 Religion and Financial Decision-Making: Religion shapes financial preferences, ethical judgments, trust, and institutional legitimacy, creating tension between personalization and neutrality.This influence appears across investment, insurance, and housing decisions, while religious cues may strengthen engagement and credibility but compromise objectivity; existing research focuses on human investors and advisors, motivating the extension to algorithmic systems.Source: §2.4 · pp. 10–12
• Religion and financial decision-making literature highlights how LLM religious framing across stock investment, life insurance, and housing purchase intersects with fiduciary neutrality.The focus is how AI-mediated personalization may affect financial advice across these three decision contexts.Source: §2.4 · p. 12

2.5 Summary and Gap

• Existing research establishes algorithmic and social bias in AI systems, but religious bias remains underexplored in high-stakes financial advisory interactions.Prior religion-focused studies largely examine moral judgment, stereotype reproduction, or classification rather than applied financial advice; little is known about AI-generated religious trust, value alignment, and moral framing.Source: §2.5 · p. 12
• The study identifies a research gap concerning whether LLMs exhibit religious bias in financial advice, how it varies across identities, decisions, and architectures, and how language expresses it.The proposed mechanisms include tone, moral framing, and cultural signaling.Source: §2.5 · p. 12
• Addressing this gap reframes personalization versus neutrality as a managerial dilemma requiring cultural sensitivity without biased over-personalization to maintain fairness, neutrality, and trust.The study links Critical Algorithm Studies with sociolinguistic and moral economy perspectives to analyze structural and discursive religious bias.Source: §2.5 · pp. 12–13

Structural foundation

• Critical Algorithm Studies treats algorithmic bias as structural asymmetries and cultural assumptions embedded in training data, model design, and deployment rather than as isolated technical errors.Training corpora, developer choices, and feedback loops can reproduce dominant cultural and religious narratives.Source: §3 · p. 13

Discursive bridge

• The sociolinguistic–moral economy bridge explains how structurally arising bias is performed and communicated through communicative style, including tone, directness, and cultural signaling adapted to audience identity and relational stance.Source: §3 · p. 14

Dual-dimensional framework

• The framework’s central contribution is a dual-dimensional model in which upstream socio-technical asymmetries generate bias and downstream linguistic and moral framing performs it in AI advice.Uneven moral framing can differentially affect financial uptake, trust, and perceived appropriateness, making algorithmic advice simultaneously technical, moral, and communicative.Source: §3 · p. 15

Research design

• The framework informs comparative multi-model, multi-scenario analysis and qualitative thematic analysis of whether religious bias occurs, how it varies, and how it is expressed.Quantitative analyses address systematic patterns across models, identities, and decision contexts, while qualitative analyses examine linguistic expression and justification.Source: §3 · p. 16

Managerial relevance

• By linking prevalence with communicative mechanisms, the framework supports culturally aware AI design while exposing over-personalization risks to client trust, fiduciary neutrality, and professional integrity.The approach contributes to scholarly understanding of algorithmic bias and the practical design of equitable AI systems.Source: §3 · pp. 16–17

4.1 Experimental Design

• The study defines religious bias as systematic differences in religious language, concepts, or references that alter recommendations or framing, focusing primarily on framing bias.The design distinguishes decision bias from changes in linguistic or moral tone.Source: §4.1 · p. 17
• The experiment varies Christian, Muslim, Hindu, and baseline advisor-client identities across global stock investment, house purchase, and life insurance scenarios, producing 16 identity combinations.These scenarios span technical and morally salient household financial contexts, while global stocks reduce country- and company-specific contextual bias.Source: §4.1 · pp. 17–18

4.2 Data Collection

• The study collected 432 outputs by running 48 standardized prompts across ChatGPT, Gemini, and Grok, with each prompt independently executed by three researchers.The same prompts enabled comparison of model-specific tendencies, temporal variation, and relative bias magnitude.Source: §4.2 · p. 19
• Controlled procedures reduced carryover from prior interactions, while executions across researchers, locations, and time intervals assessed temporal and regional variation using unpaid model versions.Religious prompts were isolated through incognito windows or disabled app activity, and unpaid versions were chosen to reflect general-user access without subscription fees.Source: §4.2 · p. 19

5.1 Coding Framework

• Researchers independently coded all 432 outputs for in-group or out-group status, explicit or implicit expression, and advisor-, client-, or interaction-based bias.Consensus review resolved coding differences, linking structural variation to model, context, and pairing and discursive intensity to implicit-to-explicit language.Source: §5.1 · p. 20

5.2 Descriptive Findings

• Unbiased outcomes ranged from about 13% for Gemini to 18% for Grok, while explicit biases reached about 73% for ChatGPT and 72% for Grok and were quite high for Gemini.Investment in global stocks yielded about 18% unbiased outcomes.Source: §5.2 · p. 21
• Same-religion pairings produced explicit interaction-based bias in 24 of 27 scenarios (89%), while religious advisors triggered advisor-based bias in 21 of 27 baseline-client scenarios (78%).The latter pattern included 20 explicit and one implicit case, suggesting models assumed religiously unspecified clients shared the advisor’s religion.Source: Table 2 · p. 22, §5.2 · p. 22

5.3 Regression Analysis

• Gemini showed about 17.4% higher religious bias than Grok, while ChatGPT’s coefficient was insignificant across regressions; decision effects were generally insignificant.Gemini’s coefficient was positive and significant in five of six regressions, with 0.160 appearing in four; in-group and out-group biases were similarly prevalent (F-stat = 0.01, p-value = 0.907).Source: §5.3 · pp. 24–25
• Approximately 15%: religious bias decreases over time for researcher 3 relative to researcher 1, ceteris paribus.The researcher-3 coefficient is significant and negative, whereas researcher 2’s coefficient is insignificant.Source: §5.3 · p. 25
• 1.799***, 1.916***, and 1.984*** are the reported coefficients for client-based, interaction-based, and advisor-based bias, respectively.All three bias sources are significant relative to the no-bias base category.Source: Table 3 · p. 26, §5.3 · p. 26
• No significant Japan effect indicates that religious bias in LLM outputs is not dependent on geographic location.The comparison uses Japan and the United States, with Japan as the reported variable relative to the US base.Source: §5.3 · p. 26
• Gemini’s higher bias persists after controlling for the 16 advisor-client interactions listed in Table 1.The complete model-6 results are reported as controlling for these interactions.Source: §5.3 · p. 27
• Only one investment-decision interaction is significant: Gemini produces lower bias for house-purchasing prompts than Grok does for stock-investment prompts.The overall evidence is insufficient to conclude that LLM bias is strongly associated with investment decision.Source: §5.3 · p. 27

6.1 Content Analysis: Application of Braun & Clarke’s Framework

• The analysis closely read and interpretively coded all 432 outputs using an AI-assisted reflexive thematic-analysis workflow guided by tone and a hybrid inductive–deductive approach, with human authors reviewing and adjudicating codes.Source: §6.1 · pp. 27–28
• Reflexive thematic analysis revealed how religious worldviews are subtly encoded in LLM outputs through moral anchoring, cultural signaling, and tone or politeness shifts.These qualitative themes expose mechanisms not visible through quantitative models alone and connect discursive expression to structural contexts.Source: §6 · p. 27, §6.1 · p. 28

6.2 Themes and Interpretive Insights

• The coding scheme identified Religious Framing (n = 374), Technical Framing (n = 320), Communication Style (n = 242), and Cultural Signaling (n = 232).Religious Framing captured moral or theological reasoning, Cultural Signaling captured tradition-specific lexical markers, and Technical Framing captured secular financial logic.Source: §6.1 · p. 29

Theme 1: Religious Framing as Moral Anchor

• Religious framing was particularly salient in advisor–client pairs involving Muslim or Christian identities.The theme was identified as the dominant pattern in the analyzed interactions.Source: Theme 1: Religious Framing as Moral… · p. 30
• Expressions such as “wise stewardship,” “as part of our faith,” and “guided by Sharia” framed financial responsibility as a moral or religious obligation.One example described future planning as fulfilling a Muslim duty to family and community in a life-insurance interaction.Source: Theme 1: Religious Framing as Moral… · p. 30
• Out-group pairings showed more religious framing, suggesting that structural identity configurations trigger stronger discursive moral anchoring.The passage attributes this pattern possibly to algorithmic overcompensation or identity signaling and links it to religious economic norms embedded in LLM training corpora.Source: Theme 1: Religious Framing as Moral… · p. 30

Theme 2: Cultural Signaling: Surface Depth, Uneven Spread

• Cultural signaling involved lexical or symbolic references to religious identity without advancing explicit moral or theological reasoning.Source: Theme 2: Cultural Signaling: Surface… · p. 31
• Islamic finance terms such as Sharia and halal appeared more frequently and fluently than Hindu references such as Grihastha or Lakshmi.Examples included “Sharia-compliant investment options may bring peace of mind and align with your values” and references to the Grihastha stage of life.Source: Theme 2: Cultural Signaling: Surface… · p. 31
• The stronger presence of Islamic terminology and frequent Christian expressions in baseline messages indicates that training-data exposure produced uneven cultural signaling.The pattern suggests deeper exposure to Islamic finance discourse and the dominance of Christian cultural norms in Anglophone training data.Source: Theme 2: Cultural Signaling: Surface… · p. 31

Theme 3: Technical Framing and Rationalized Advice (n = 320)

• Technical framing emphasized diversification, risk mitigation, and tax planning, particularly in stock-investment scenarios or when religion was absent.One example recommends international diversification to mitigate country-specific risk.Source: Theme 3: Technical Framing and… · p. 31
• Technical framing was more frequent in out-group interactions, where LLMs appeared to avoid culturally specific commitments.This pattern suggests that objective financial logic can function as a culturally neutralizing strategy.Source: Theme 3: Technical Framing and… · p. 31
• Technical framing dropped sharply in insurance outputs, where moral and emotional framings were privileged for family protection.The contrast indicates that financial scenario materially conditions how advice is rationalized.Source: Theme 3: Technical Framing and… · p. 31

Theme 4: Communication Style: Deference vs. Assertion (n = 242)

• Messages to Hindu and Muslim clients were more formal or deferential, whereas messages to Christian or baseline clients were more assertive.The contrast appears in formulations such as “you may wish to consider…” versus “you should…”.Source: Theme 4: Communication Style:… · p. 32
• LLMs appeared to learn tone differentiation patterns tied to cultural scripts and perceived expectations around politeness or authority.The interpretation is linked to Ting-Toomey & Dorjee (2018) and Economidou-Kogetsidis, 2011.Source: Theme 4: Communication Style:… · p. 32
• Structural identity pairing organized the discursive choice between deference and assertion.Thus, identity cues shaped not only what advice said but also how authoritatively it was delivered.Source: Theme 4: Communication Style:… · p. 32

6.4 Synthesis with Quantitative Results

• Explicit bias dominated both pairing types: in-group interactions favored direct religious anchoring, whereas out-group interactions elicited more deferential tone and cultural signaling.Out-group pairings showed especially overt politeness and moral framing, consistent with performative caution toward perceived cultural difference.Source: §6.3 · p. 32, §6.4 · p. 35

6.3 From Themes to Granular Patterns

• Model distributions varied systematically: Gemini exhibited strongest Cultural Signaling (n = 104), while Grok had highest proportion of Communication Style codes (n = 134); ChatGPT displayed most balanced thematic profile.Source: §6.3 · p. 33, §6.4 · p. 35

6.4 Synthesis with Quantitative Results

• Scenario effects were consistent: life insurance advice was most moralized and religiously framed, stock advice primarily technical, and house-related advice intermediate.Only three messages, all from Gemini, lacked explicit religious, cultural, or technical framing; these outputs were minimal and professional.Source: Theme 5: Neutral · p. 32, §6.4 · p. 35
• Structural patterns predicted when explicit bias occurred, while moral anchoring, cultural signaling, and tone showed how it was linguistically realized.This alignment operationalizes the framework linking structural and discursive dimensions of religious bias.Source: §6.4 · p. 35

7.1 Key Findings Synthesis

• Only 12–18% of 432 outputs were unbiased, while Gemini showed higher bias than Grok and ChatGPT did not significantly differ from Grok.The study presents systematic evidence that religious bias is prevalent across LLM-generated financial advice.Source: §7.1 · pp. 36–37
• Religious identity cues elicited explicit framing, whereas neutral advisor–client prompts rarely produced religious content, indicating cue-responsive personalization rather than spontaneous religiosity.In-group pairings almost invariably triggered framing, and baseline clients often received advisor-centered appeals.Source: §7.1 · p. 36
• Once identity cues appeared, equivalent advice was reframed through moral anchoring, uneven cultural signaling, religious nudging, and tone modulation, creating discursive bias.Islamic-finance terminology was more fluent than Hindu equivalents, while Muslim and Hindu clients more often received deferential phrasing.Source: §7.1 · pp. 36–37
• Religious over-personalization can exclude secular investment options and undermine perceived objectivity, fiduciary integrity, and fairness, making identity responsiveness both a communicative strength and a managerial risk.The discussion reframes bias as a continuum from adaptive to over-personalized behavior and identifies personalization-versus-neutrality as the central managerial dilemma.Source: §7.1 · pp. 36, 39
• Financial scenario shaped framing: stock advice was more technical and less religious, whereas life insurance elicited stronger religious language linked to mortality, family protection, and responsibility.Regression results showed limited scenario effects, with life insurance producing a small, marginally significant increase in one specification, while qualitative patterns were stronger.Source: §7.1 · p. 37
• Structural bias reflects systematic model and context variation rooted in training and design, while discursive bias describes how identity cues alter the linguistic and moral delivery of advice.Cross-model contrasts and scenario effects indicate embedded response patterns rather than spontaneous conversational drift; the same recommendation can shift in tone, moral anchoring, or cultural signaling.Source: §7.1 · p. 38
• The discussion emphasizes inclusivity and engagement while requiring neutrality, fairness, and trust in AI-mediated financial advice.This synthesis motivates safeguards that preserve engagement without compromising professional standards.Source: §7.1 · p. 39

7.2 Implications for Theory

• The study positions religious bias as systematically encoded in both the structure and discourse of LLM-generated financial advice, extending algorithmic bias research beyond other social categories.Religion is presented as a critical dimension of LLM-mediated decision-making and alongside established categories in bias research.Source: §7.2 · p. 39
• The proposed framework distinguishes decision bias from discursive framing bias and links structural asymmetries to moral language, interactional style, trust, and context-dependent outputs.Stock investment prompts elicited more technical responses, whereas life insurance advice produced stronger religious language, showing that decision context interacts with model tendencies.Source: §7.2 · pp. 39–40

7.3 Implications for Practice

• Developers, financial institutions, and regulators should adopt religion-aware audits, cultural-sensitivity checks, fairness criteria, and explicit opt-in controls for religious framing.A neutral-by-default, explicit-opt-in approach backed by audits and user controls is proposed to balance engagement with fiduciary integrity and protect trust and inclusivity.Source: §7.3 · pp. 41–42
• Religiously tailored advice may strengthen trust when based on explicit client preferences, but unsolicited or misattributed framing creates over-personalization that can erode fiduciary trust and alienate clients.The central managerial dilemma is whether and under what safeguards AI should personalize financial advice along religious lines.Source: §7.3 · p. 41

Limitations

• Findings are bounded by three models and examined timeframes because model updates, training data, and fine-tuning may change bias patterns.The study systematically compared major LLMs but cannot establish whether observed patterns persist across future model versions.Source: §8 · p. 42
• Restricting analysis to Christianity, Islam, and Hinduism limits generalizability, while interpretive coding could be strengthened through alternative frameworks or larger coder panels.The selected traditions provide global breadth but omit other significant identities; independent replication and cross-checking validated the coding approach.Source: §8 · p. 42
• The study documents religious bias and framing but does not measure how clients interpret, trust, act on, or benefit from such advice.Questions about framing’s appropriateness, accuracy, and effectiveness remain open for future research.Source: §8 · p. 42

Future Research

• Future studies should include more religious identities and compare free versus paid tiers across successive model generations to assess diversity, commercial strategies, and technical updates.These comparisons could reveal how bias operates across underrepresented traditions and changes with model development.Source: §8 · p. 43
• Research should extend beyond text to voice agents and chatbots, while linking biased advice to client trust, decision quality, recommendation uptake, and broader consequences.Such work would move the field from detecting bias toward assessing its economic, ethical, and managerial effects.Source: §8 · p. 43

Theoretical contribution

• The study foregrounds religion in algorithmic bias research and frames AI advice as both a technical and communicative act requiring attention to how recommendations are framed.Its dual structural and discursive perspective links bias to trust, inclusivity, legitimacy, and fairness.Source: §9 · pp. 43–44

Core findings

• Bias appeared structurally through model- and scenario-level regularities and discursively through moral anchoring, cultural signaling, and tone modulation.These patterns show that religious bias affects both the substance of recommendations and their framing.Source: §9 · p. 43
• Across 432 outputs, unbiased responses occurred in fewer than one in five cases; Gemini showed more bias than Grok, while ChatGPT was broadly comparable to Grok.Bias was most pronounced in in-group advisor–client interactions and descriptively strongest in life insurance advice.Source: §9 · p. 43

Managerial and regulatory implications

• The findings identify a managerial dilemma between personalization and neutrality, requiring safeguards in training and moderation, neutral client applications, and religion-sensitive AI audits.Developers, financial institutions, and regulators each have distinct responsibilities for preserving fairness and fiduciary integrity.Source: §9 · p. 44

Limitations and future research

• Because the study documents prevalence rather than benefits of religious framing, it calls for distinguishing meaningful personalization from inappropriate over-personalization.Future research should examine when religious or moral framing affects user trust and decision quality.Source: §9 · p. 44

Broader significance

• Addressing religious bias is essential for making LLM-mediated financial services fair, inclusive, culturally respectful, and capable of sustaining fiduciary trust and professional integrity.The conclusion presents neutrality–personalization balance as increasingly critical as financial advice becomes digitally mediated.Source: §9 · p. 44