Research

Abstract: Algorithmic underwriting systems must set interest rates that are commensurate with credit risk, yet two pricing failures persist: demographic bias correlated with protected borrower attributes, and risk distortions across the creditworthiness distribution. Fair-lending regulation often prevents platforms from observing the protected attributes used in most debiasing approaches. We study this as a problem of fair loan pricing with unobserved protected attributes. We propose equal loan attractiveness as a fairness criterion for loan pricing and implement it through Generative Adversarial Pricing (GAP), a framework that learns risk-commensurate interest rates without using protected attributes. Drawing on nearly two million unsecured personal loans from Prosper and LendingClub, we find evidence of bias against African American borrowers and recipients of public assistance. We also document risk distortions, where borrowers with lower FICO scores are charged disproportionately higher rates than their credit risk warrants. In out-of-sample tests, GAP reduces these disparities and distortions while preserving risk-based pricing. The results suggest that fairer loan pricing is feasible even when protected attributes cannot be collected.

Abstract: A long-standing aspiration of information systems (IS) research on financial technology is the democratization of financial services through IT-enabled innovation. Generative AI (GenAI), with its unprecedented capacity to synthesize unstructured information through a natural-language interface, appears poised to fulfill this promise for retail investors. We ask whether it actually does. Between May and October 2025, we collect 2,845 forward-looking directional forecasts on the 101 constituents of the Nasdaq 100 Index generated by Gemini Deep Research, a state-of-the-art consumer-facing agentic workflow, across weekly, monthly, and quarterly horizons. This prospective design eliminates the look-ahead bias that has confounded prior backtesting-based evaluations of large language models (LLMs). We find that the LLM's directional precision is statistically indistinguishable from a coin flip at every horizon, and that portfolios built on its signals underperform the Nasdaq 100 Index. To diagnose the mechanism behind this failure, we parse every cited work in each Deep Research report and show that the LLM acts as a sentiment parrot: its directional outlook is shaped by the sentiment, visibility, and quality of the sources it consumes, yet none of these features is associated with predictive accuracy. These results articulate a democratization paradox: GenAI democratizes the process of investment research but not its outcome, because the binding constraint on quality is the architecture of publicly available internet content rather than model capability. The paradox may generalize to other high-stakes domains where LLMs place expert-seeming synthesis in non-specialist hands, and motivates IS scholarship, platform design, and regulatory frameworks that surface the gap between perceived and actual analytical competence.