WAITS – IX

The Ninth Washington Area IT Symposium (WAITS) will be hosted at The George Washington University, home of the GW Revolutionaries, in Washington DC the afternoon of April 4, 2025 from Noon to 2:30pm. I know what you are thinking, did Shuba agree to get sandwiches? Oh he did my friends! Oh yes, he did! Moreover, my understanding is that famed GW Alumni Alec Baldwin, Courteney Cox, and Jacqueline Kennedy Onassis will also be there! The only way to know if I am telling the truth is to be there!

The goal of WAITS is to bring together like-minded IS scholars in the greater DC Metro area and facilitate inter-school dialogue and collaboration. While there are many business schools in the DMV, and many tremendous IS departments each with their own talented researchers, communication between the groups is limited. This stifles synergies, creative and innovative work, and the opportunity for each of us to learn from each other. It is the objective of WAITS to eliminate this problem by coming together every semester.

Speakers

We are lucky to have two fantastic researchers giving talks.

Sunghun Chung
George Washington University

Delivering AI Responsibly for Digital Equity: Three Sequential Field Experiments in Revenue-Based Financing

This study investigates how to deliver AI in a responsible manner to narrow the prevalent digital inequality of AI utilization among investors in revenue-based financing (RBF). Accurate AI analytics of loan applications are critical in RBF because they reduce uncertainty for investors and help them assess borrowers’ future revenue more reliably. Over the course of three years, we collaborated with a major RBF platform in Asia to conduct three sequential field experiments involving over 2,300 investors. In the first experiment, we demonstrate the presence of inequality in AI utilization. Investors were randomly assigned to a control or treatment group receiving AI-predicted revenue of borrowers. While AI significantly helps investors with 21.5% higher amount as a whole, sophisticated investors benefited disproportionately—investing 30.8% more, improving capital utilization by 67.3%, and reducing delinquency rates by 3.3% compared to unsophisticated investors. Further analysis revealed that AI aversion, lower AI accessibility, and, limited cognitive bandwidth among unsophisticated investors significantly contributed to this disparity. In response, the second experiment introduced explainable AI (XAI). The results show it narrows the gap between high and low-income investors and reduced inequality for non-tech-savvy and rural investors. Nonetheless, institutional investors continued to reap larger benefits. Building on these findings, the third experiment aimed to deliver responsible AI (RAI) by disclosing all relevant predictive features alongside visual explanations. This provision helped unsophisticated participants catch up more effectively, reducing the disparity observed in previous experiments. The series of findings underscore that while standard AI interventions can improve overall investment outcomes, they may inadvertently widen digital inequalities. Introducing transparent and user-centric explanations—aligned with RAI principles—can foster more equitable engagement and performance, highlighting practical avenues for inclusive digital finance.

Jingjing Li
The University of Virginia

AUTOMATING IN HIGH-EXPERTISE, LOW-LABEL ENVIRONMENTS: EVIDENCE-BASED MEDICINE BY EXPERT-AUGMENTED FEW-SHOT LEARNING

Many real-world process automation environments are characterized by high expertise requirements and limited labeled data. To address this challenge, we propose FastSR, a computational design science artifact designed to automate systematic reviews (SRs) in such environments. SR is a manual, resource-intensive process that collects and synthesizes data from medical literature to inform clinical decisions and improve medical practice. Existing machine learning (ML) solutions for SR automation struggle with data scarcity and fail to accurately replicate the nuanced reasoning involved in expert-driven SR processes. Motivated by the human ability to learn from limited examples—a capability rooted in compositionality—we introduce FastSR, a principled and generalizable few-shot learning framework that automates the multi-step, expertise-intensive SR process using minimal training data. Informed by SR experts’ annotation logic, FastSR enhances traditional few-shot learning through three key components: (1) diverse representations to capture the breadth of SR knowledge, (2) attention mechanisms to model the semantic relationships within medical texts, and (3) shared representations to jointly optimize interrelated tasks, such as sentence classification and sequence tagging. We instantiated and evaluated FastSR on three datasets: full-text articles on Wilson disease (WD) and COVID-19, as well as the public EBM-NLP dataset containing clinical trial abstracts across a range of diseases. FastSR consistently outperformed state-of-the-art models, including the latest generative AI models like GPT-4, reducing SR completion time by up to 65%. We critically examine FastSR’s outcomes and practical advantages, proposing a new FastSR-augmented protocol tailored for the medical domain. This work not only sets a new benchmark for SR automation but also offers foundational insights into the design of AI systems capable of surpassing even the most advanced large language models in specialized, high-stakes domains.

Want to present at WAITS? Shoot me an email. You know how to find me.