Artificial intelligence is increasingly used to help diagnose cancer, but new research shows those tools can mirror and magnify existing health disparities. An analysis of AI systems trained to read pathology slides found that nearly one in three cancer diagnostic tasks showed performance gaps tied to race, sex, or age, raising concerns about unequal care for minority patients.
Researchers examined more than 28,000 pathology images from over 14,000 patients across 20 cancer types. While overall accuracy often appeared high, deeper analysis revealed that standard AI models performed unevenly across demographic groups. These disparities were traced in part to training datasets dominated by White patients, which can skew how algorithms learn to recognize disease patterns.
The study found that tumors from African American patients showed distinct tissue characteristics, including higher densities of cancer cells and lower levels of immune and stromal components compared with tumors from Caucasian patients. AI systems detected these differences and, without safeguards, incorporated them into diagnostic decisions. Age-related variations also mattered, with tumors from older patients showing more stromal tissue and less inflammation than those from younger patients.
“Because AI is so powerful, it can differentiate many obscure biological signals that cannot be detected by standard human evaluation,” said Kun-Hsing Yu, a researcher from Harvard Medical School. That strength, the authors warned, becomes a liability when demographic imbalance shapes what the technology learns.
The researchers reported that bias affected 29.3% of diagnostic tasks they evaluated. They developed a fairness-aware framework designed to reduce these gaps, cutting disparities by nearly 90% in testing. The findings underscore how emerging medical technologies can either narrow or deepen racial health inequities, depending on whether equity is built into their design. Without deliberate checks, tools meant to improve cancer detection risk delivering less accurate diagnoses to patients already facing unequal health outcomes.
See: “One in Three AI Cancer Diagnoses Vulnerable to Demographic Bias” (December 16, 2025)
