Consilio Report: A Comparative Evaluation of Demographic Bias in PII Detection Systems

Personally Identifiable Information (PII) detection systems are increasingly relied upon to ease the burden of large-scale manual document review to identify sensitive information in critical legal and regulatory workflows, such as identifying impacted individuals following a cyber breach.

However, bias in PII detection systems creates disproportionate privacy risks when certain demographic groups are less likely to be detected and therefore to be protected. In this technical report, we investigate whether PII detection systems and popular LLMs exhibit a bias against non-Western names, hypothesising that Western names are detected more reliably than non-Western names.

In our evaluations, the LLMs tested missed far fewer names than the traditional PII detection tools tested, and 4 of the 12 LLMs showed a statistically significant demographic bias - a higher rate of missed names missed for one group than another. We also test an in-house PII detection LLM, which was intentionally fine-tuned on a diverse mix of synthetic and real public international documents.

Among the systems that did not show statistically significant demographic bias, this LLM achieved the highest measured recall (99.3%) while maintaining an economical footprint of only 8B parameters, supporting large-scale deployment. These findings emphasise the importance of designing PII detection systems to achieve overall accuracy while ensuring fair and equitable performance across demographic groups.

The report is prepared by Consilio. Download the full report here.