A Computational Audit of Demographic Association Encoding in ClinicalBERT Language Predictions

In three linesComputational audit of ClinicalBERT showing 65.6% of demographic biases stem from model-internal amplification rather than training data inheritance. Analysis via Log Probability Bias Analysis and MLM probing across 98 real clinical sentence templates and 8 intersectional race-gender combinations.

Summary generated by Claude — human-verified