111 Iowa L. Rev. 253 (2025)
Abstract
Financial technology has long relied on data like an applicant’s current indebtedness to make decisions about who gets access to new credit, but AI is now enabling credit determinations based on some unusual inputs. This “alternative data”—or data that is not intuitively connected to creditworthiness—includes information like a consumer’s online shopping habits, whether they paid their rent and utility bills, and even how many friends they have on social media. The use of this kind of data has exciting potential to expand access to credit but exemplifies a well-known feature of machine learning: It works by finding nonintuitive relationships in large datasets.
Regulators and industry have recognized both the potentials and pitfalls of alternative data. As nascent discussions emerge around appropriate uses of this alternative data, there is little consensus about the ideal shape of regulations for alternative data. This Article offers a path forward for alternative data regulation. In doing so, it surfaces a central tension that regulators must resolve—the contradiction between the desire for intuitive stories and the reality of how these technologies work. For instance, scrolling quickly through an online contract could be indicative of carelessness about legal obligations, which may make someone a risky creditor. Intuitive stories have long formed the basis for social oversight of credit lending. These stories have enabled us to determine that information such as zip codes and medical debt are unfair to include in credit calculations, but that using on-time payment history is acceptable.
Yet the emergence of machine learning models trained on alternative credit data challenges classic assumptions about the connection between data and the narratives we tell. For example, intuition is broken if a person “games” data by changing a proxy for creditworthiness without changing the underlying characteristic a system designer intends to measure. By uncovering this potential for breaking causation, this Article surfaces the true normative stakes of regulating alternative credit data. The stakes are not about “connectedness” between data and decisions, as existing regulations imply.