Computer Science editorial
Open AccessOA2026
Annie, Are You Okay? How Style- and Context-Based Personalization Shape AI-Assisted Decision-Making
A preregistered experiment reveals that context-based personalization in AI financial advice steers decisions without increasing perceived correctness or trust.
Hasibur Rahman; Benjamin R. Cowan; Smit Desai· 2026· DOI 10.48550/arXiv.2609.24644
The core problem
As generative AI systems become ubiquitous in financial advising, they increasingly personalize both how they communicate (style) and what they recommend (context). This study investigates whether these two forms of personalization differentially affect decision-making. The authors conducted a preregistered 2x2 between-subjects factorial experiment with N=240 participants. Participants first ranked three comparably viable stocks, then discussed them with an AI advisor, and finally reranked the stocks. The central hypothesis is that personalization—whether stylistic or contextual—can shape decisions, but the mechanisms and outcomes may differ. The research question is: How do style- and context-based personalization influence AI-assisted decision-making, and do they affect perceived influence, correctness, trust, intelligence, likeability, and quality? The study aims to uncover whether personalized AI can steer decisions among defensible options with minimal evaluative trace, raising concerns for design and governance.
Innovation
Participants perceived both forms of personalization, but only context-based personalization reliably changed ranking behavior. Specifically, context-based personalization increased reconsideration and moved rankings toward the AI's assigned recommendation. In contrast, style-based personalization did not significantly affect ranking changes. Despite the behavioral shift, participants did not judge the AI as more correct, trustworthy, intelligent, likeable, or high-quality when context-based personalization was present. Interestingly, participants felt more influenced by the AI when personalization was used. Those initially farther from the AI's recommendation moved more toward it, yet judged its advice as less correct. Exploratory analyses suggested greater susceptibility among lower-expertise participants. These findings indicate that context-based personalization can steer decisions without leaving a strong evaluative trace, as the AI is not perceived as more competent or trustworthy despite its influence.
As generative AI systems become ubiquitous in financial advising, they increasingly personalize both how they communicate (style) and what they recommend (context). This study investigates whether these two forms of personalization differentially affect decision-making. The authors conducted a preregistered 2x2 between-subjects factorial experiment with N=240 participants. Participants first ranked three comparably viable stocks, then discussed them with an AI advisor, and finally reranked the stocks. The central hypothesis is that personalization—whether stylistic or contextual—can shape decisions, but the mechanisms and outcomes may differ. The research question is: How do style- and context-based personalization influence AI-assisted decision-making, and do they affect perceived influence, correctness, trust, intelligence, likeability, and quality? The study aims to uncover whether personalized AI can steer decisions among defensible options with minimal evaluative trace, raising concerns for design and governance.
The experiment employed a 2x2 between-subjects design, manipulating style-based personalization (present vs. absent) and context-based personalization (present vs. absent). Participants (N=240) were randomly assigned to one of four conditions. They first ranked three stocks that were comparably viable based on financial metrics. Then they engaged in a discussion with an AI advisor that provided recommendations. The AI's communication was personalized either in style (e.g., matching the participant's language style) or in context (e.g., referencing the participant's prior choices or preferences). After the discussion, participants reranked the stocks. Measures included ranking changes, perceived influence, and evaluations of the AI's correctness, trustworthiness, intelligence, likeability, and quality. The study was preregistered, ensuring transparency and reproducibility. The analysis focused on comparing ranking shifts and perceptions across conditions, with exploratory analyses on expertise and initial distance from the AI's recommendation.
Why it matters
The results demonstrate that context-based personalization can subtly steer decision-making among defensible options. The lack of increased perceived correctness or trust suggests that the influence operates without participants recognizing the AI as more authoritative. This raises concerns for the design and governance of personalized decision support systems, as users may be swayed without awareness. The finding that lower-expertise participants are more susceptible highlights potential vulnerabilities. The study's limitations include a specific financial context and a single interaction. Future research should explore long-term effects and other domains. The implications call for ethical guidelines and transparency in AI personalization to prevent undue influence. The equation for the experimental design can be represented as:
, where is the effect of style personalization, is the effect of context personalization, and is their interaction. A Mermaid diagram of the experimental flow is shown below:
Who should read this
CS practitioners and researchers
Opening member content…