Exploratory As-Analyzed No-Detection of Culturally-Marked Predicate-Triggered PII Amplification in a Synthetic-English RAG Probe: A Predicate-Resource-Confounded Audit

An exploratory research audit investigated whether retrieval-augmented generation (RAG) systems leak more personal identifiable information (PII) when responding to stereotype-loaded queries about culturally marked individuals compared to neutral queries. The study, which pre-registered a four-culture audit and used a synthetic English PII corpus, found no definitive evidence of such amplification. However, significant methodological caveats, including the exploratory nature of the analysis and a prompt-echo artifact affecting name-leakage metrics, limit the conclusiveness of these findings.

Why it matters

Understanding how RAG systems respond to culturally sensitive queries is crucial for mitigating risks related to data privacy and algorithmic bias in AI deployments. The identified methodological challenges highlight the complexities in auditing AI systems for PII leakage and the need for robust evaluation frameworks to ensure ethical and secure AI development.

Key insights

• The research aimed to determine if stereotype-loaded queries about culturally marked individuals increase PII leakage from RAG systems compared to neutral queries.
• A four-culture audit (en-Anglo, es-LATAM, Arabic, Hindi) was planned using a synthetic English PII corpus.
• The study compared five query arms, termed the Stereotype-Trigger Leakage Delta (STLD).
• The analysis was entirely exploratory, as the pre-registered confirmatory estimator was not executed.
• A significant 'prompt-echo' artifact contaminated the name-leakage metric, where the model often re-emitted the queried name, artificially inflating apparent leakage without actual retrieval.
• No conclusive evidence of PII amplification due to stereotype-loaded queries was detected in this exploratory analysis.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.20351