A study warns that some AI pathology models used to detect cancer biomarkers may rely on unreliable “shortcuts” instead of true biological signals. Researchers from the University of Warwick found that these models often depend on correlations in clinical features rather than direct molecular evidence. This can work in many cases, but when conditions change or patients fall into specific subgroups, the models may fail and lead to possible misdiagnosis.
After analyzing more than 8,000 tissue samples from breast, colorectal, lung, and endometrial cancers, the researchers identified a key problem called interdependency. For example, in colorectal cancer, AI models often could not clearly separate BRAF mutations from microsatellite instability (MSI). Because these two features frequently occur together, the AI learned to predict BRAF status based on MSI patterns. Confusing them could result in incorrect treatment decisions.
The authors stress that AI remains a useful tool, but it should not replace standard molecular testing at this time. They recommend more rigorous evaluation methods, better reporting of performance across different patient groups, and careful clinical use. Until AI models show true biological understanding, their results should be treated as supportive tools rather than final diagnostic answers.