Human Preference Alignment in Financial Advice: A Generative AI Approach

Financial advice is a highly regulated domain where unclear communication can cause consumer harm and regulatory breaches. Yet existing recommendation systems often fail to adapt to individual risk preferences and comprehension levels. In this work, we investigate how generative AI can be used to improve both the clarity and personalisation of financial product communications. We first construct a benchmark of clarity by collecting human ratings of real financial product descriptions and construct a novel dataset with 25,000 synthetically generated variations. Using this dataset, we then explored two optimisation strategies for generative models: dynamic generation guided by classifier feedback, and an RLHF-style approach using the classifier as a reward model. Our findings show that clarity is shaped both by textual style and consumer profile, and that integrating preference signals significantly improves comprehensibility. This work contributes a benchmark, models, and methods for aligning generative AI with human preferences in financial communication.