Supervisors Symposium

MODEL BIAS

Model Validation: Bias Concerns (cont’d) Jimenez notes that FinCEN has found explicit bias in SAR reporting: • FinCEN has warned that filings should not be based on a person’s ethnicity, such as being of Middle Eastern descent. • Despite the potential for bias, FinCEN data is widely available to law enforcement agencies and FinCEN employs algorithms to screen filings to identify reports that merit further review.

23

MODEL BIAS

Model Validation: Bias Concerns (cont’d) ● This danger can be exacerbated if the high ‐ tech entities doing testing on such models are not populated with programmers/analyst teams with diversity and inclusion as part of their groups. ● Implicit bias can become part of these AI tools and analytics and it can be compounded if the data set being analyzed also has certain bias aspects due to imprecision in collection or cataloging of the information. ● There is a danger of discriminatory outcomes in the use of AI proprietary systems which leads to possible legal exposure for private firms. There are similar potential risks leading up to and including mission failure for public policy entities such as federal bank regulators if model bias is not addressed.

24