Enhancing Business Intelligence: An Experiential Learning Journal at a Major Financial Institution ; Balancing the Equation: Ethical Dilemmas in Artificial Intelligence Based Credit Scoring

The technical report in this portfolio is about my time as a software engineering intern at Capital One, a major financial institution. Capital One needed to enhance an internal Business Intelligence (BI) tool to improve functionality for end-users data product permission requests. My team and I developed a backend API to orchestrate requests between different internal products. The BI tool streamlined the data product permission request process, enhancing efficiency while ensuring data security. Continuous improvement will include user interface enhancements and integration of advanced recommendation algorithms powered by machine learning.
The thesis begins with an introduction to the inequity present in the financial lending industry. From disadvantaging minority communities to being a barrier to job entry for many Americans, the current system of credit scoring is flawed – despite being a step in the right direction from the days of bankers looking at your skin color or church attendance. Three literature choices, with emphasis on O’Neil’s Weapons of Math Destruction, support this notion of inequality as the current system traps people in a cycle of poverty.
With that comes the solution posed by many: Artificial Intelligence (AI). Many argue that utilizing artificial intelligence will mitigate the equity issue of credit. Eubanks disagrees, noting how automated decision-making software uses big data and predictive models (i.e., AI) to further profile and punish disadvantaged members of our society. Furthermore, Broussard adds that since models are built using historical data – data which is plagued with biases, limitations, and errors – they will only keep things the same, if not make the situation worse by increasing the usage and confidence in the data sources.
Research done on the data biases issue highlights an interesting finding – the data artificial intelligence models use do lead to biased results, to the point where the original gender and ethnicity of data points can be predicted. However, it is demonstrated that AI-based models were still able to increase approval rates while reducing default rates – which can be seen as both a positive to the lenders and the consumers.
Next, is the issue of transparency of these artificial intelligence models. If we cannot tell how they are coming up with a decision, how can we be able to interpret if the decisions are fair? Two different groups of researchers attempt to tackle this with the eXplainable AI framework that attaches another model to the original model to interpret its decision-making. As a developer, I find this to be inadequate as the explanation is not being made by the model making the decision.
The last part of the research is on Upstart.com. An online lender that uses artificial intelligence with e-scores, credit score look-a-like that uses non-traditional data not protected by legislation. Although Upstart claims that they expand loan access to more people, I found the rates to be very high, bordering on predatory. A five-year $27,000 loan from Upstart had an APR greater than 2 times the APR from the lender Discover with the same terms.
Analyzing all of the research through the lens of Actor-Network Theory, I conclude that the addition of artificial intelligence into credit scoring has a very marginal effect. Although the AI is an actor of its own, it is beholden to the decisions of its maker, thus fundamentally the decision maker remains the same. All that is achieved is an increase in loan eligibility at the cost of data privacy and a guarantee of true explainability. Future research should focus on whether increased access to credit leads to better outcomes for all actors, especially consumers, and lenders, since the current results I have seen are leaning towards overwhelmingly benefitting the lender. Another research area is to develop an AI model that can explain itself instead of relying on another model to explain it. This would go a long way in ensuring that decisions can be traced and critiqued.
The technical report and the thesis do not have a connection outside of working in the finance industry leading me to research the crossroads of my interests in finance and computer science. After seeing countless ads for Upstart and similar AI-based lenders, I desired to leverage my experience in machine learning and my understanding of the financial landscape through my internships at Capital One and the London Company to analyze what this means for us, the consumer, and me – the developer.

School of Engineering and Applied Science

Bachelor of Science in Computer Science

Technical Advisor: Rosanne Vrugtman, Briana B. Morrison

STS Advisor: Joshua Earle

Technical Team Members: