Stanford HAI Documents 26% Black and 15% Asian Rejection Rates in Deployed AI Resume Screeners

The Stanford HAI analysis examined commercial applicant tracking systems and large-language-model resume parsers on matched resume sets differing only by name signals. Systems applied name-derived proxies for race at the parsing and scoring stages, producing measurable disparate impact before any human review occurred.

Data align with prior measurements. Amazon's 2018 recruiting engine downgraded resumes containing women's college indicators. Subsequent audits of HireVue and LinkedIn assessments showed similar name and zip-code effects. EEOC four-fifths rule violations appear in multiple vendor deployments once name-ethnicity mapping is isolated.

Operationally, these rates compound across high-volume pipelines. A firm processing 50,000 applications monthly loses thousands of qualified Black and Asian candidates at the first automated filter. Retraining on de-identified text reduces but does not eliminate the signal when downstream models retain correlated features such as school prestige or gap years.

Vendors now face increasing documentation requests under new state automated decision tool statutes. Primary mitigation remains removal of name and address fields before model ingestion.