U.S. banks are introducing the concept of "nutrition labels" that allow them to see at a glance the quality of data fed into AI and where accountability lies. The move aims to reduce the risk that incorrect or outdated data could lead to AI errors.

American Banker reported on Sept. 8 that a task force under the Financial Services Sector Coordinating Council (FSSCC) has prepared a Data Nutrition Label (DNL) framework for assessing data for AI, and some banks have begun applying it in their work. DNL helps determine whether data is suitable for AI use by showing its provenance, quality, freshness, completeness, bias and update cycle.

FSSCC proposed a two-tier evaluation system, consisting of a basic and an advanced level. The basic level checks whether data meets minimum quality requirements, while the advanced level more closely evaluates AI-related factors such as data collection, preprocessing, sample composition and maintenance. Requirements also vary by use. Higher data quality is needed for tasks where financial risk occurs directly, such as credit and lending reviews, than for generating marketing copy.

PNC Group distinguishes between traditional model risk management and generative AI risk in its operations. Ned Carroll (네드 캐럴), PNC's head of data and automation, said it is hard to manage AI in the same way as traditional deterministic models because AI generates answers probabilistically. He added that systematic data management is therefore important.

Accountability is also central. The approach is for data owners to verify accuracy and timeliness, and to demand quality assurances even when using data from external suppliers such as credit rating agencies or market information firms. Banks are not using general-purpose models such as ChatGPT and Claude as-is, but are combining them with internal policies, procedures and their own data to apply them in ways that fit their business context.

DNL is not a cure-all. Ian Schnoor (이언 슈노어), managing director at the Financial Modeling Institute, said it is a starting point for improving data transparency but noted that room can remain for interpretation or manipulation depending on the evaluation method. Ultimately, the AI race in finance is expanding beyond model performance into a contest over how reliably data can be managed and trusted.