Developing and Evaluating an Empathetic Peer-Bot for Young Adults with Financial Anxiety

Background and Objective: Approximately 1 in 5 young adults in the United States experience mental health challenges each year, with conditions such as anxiety and stress increasingly linked to financial insecurity and economic instability. Rising living costs, including housing and education, alongside relatively stagnant wages, make it difficult for young adults to achieve financial stability. In this context, financial anxiety, characterized by persistent worry or distress related to personal finances, has emerged as a serious concern, with nearly 60% of young adults reporting anxiety about their financial situation. Limited access to mental health services further amplifies these challenges. The conversational capabilities of next-generation AI tools have led many young people to use them for mental health–related advice, reflecting a growing reliance on digital platforms for emotional support. Many AI systems tend to offer directive advice; however, these AI ‘solutions’ ignore the fact that most young adults just want a friendly peer to empathize with their problems, highlighting the need for user-centered models with proper safeguards that support both emotional well-being and financial resilience. Thus, the primary aim of this study was to develop and evaluate an empathetic “peer bot” that responds to financial anxiety and stress in a supportive, non-directive manner.

Method: We fine-tuned a base large language model, Mistral 7B Instruct v0.3, using a synthetic dataset of prompts and responses centered on financial anxiety, stress, and economic insecurity. Prompts included concerns about debt, savings, employment, and financial decision-making and were designed to reflect realistic scenarios that young adults would seek support for. The responses were designed to be empathic, acknowledging users’ struggles, offering peer-like support, and avoiding any directive advice. To evaluate model performance, 1,250 synthetic prompts were processed by both the base model and the fine-tuned “peer-bot.” Comparative analysis assessed the presence of directive language, while emotion analysis was conducted using a RoBERTa-base model to quantify levels of empathetic expression in generated responses.

Results: Comparative analysis revealed that the fine-tuned “peer-bot” substantially reduced directive language present in the base model’s outputs and prioritized empathetic and validating responses. Quantitative emotion analysis indicated a measurable increase in expressions associated with empathy, with the “peer-bot” more consistently acknowledged the emotional dimensions of financial stress. The fine-tuned model saw a 68% decrease in “neutral” emotion and a 25% increase in ‘fear’ and “sadness” emotions, showcasing how the peer-bot emotionally mirrors instead of responding with apathetic to-do lists.

Discussion: Findings suggest that model fine-tuning can meaningfully alter response style in ways that align more closely with user preferences for peer-like, non-directive interaction. Findings also highlight the potential for empathetic AI systems to support young adults in ways that reinforce positive financial behaviors. Such tools may help users feel more supported and engaged, potentially increasing their confidence to seek information, make informed financial decisions, and access formal services and resources when needed. At the same time, these systems should be viewed as complementary rather than substitute to formal services, as they do not replace professional financial advising or mental health care.