Researchers at the University of Warwick report that artificial intelligence systems used in cancer pathology often rely on statistical shortcuts rather than biological signals, raising concerns about their reliability in clinical care, according to a study published Wednesday in Nature Biomedical Engineering. Highlighting issues related to AI cancer tools shortcut learning.
The study finds that while many deep learning models show high accuracy in predicting cancer biomarkers from microscope images, their performance often depends on indirect visual cues instead of causal biological patterns.
Researchers Identify Shortcut Learning Across Cancer Types
Scientists at the University of Warwick analyzed more than 8,000 patient samples spanning four major cancers: breast, colorectal, lung, and endometrial. The findings appear in Nature Biomedical Engineering.
The team compared leading machine learning approaches and found that models frequently relied on correlations between biomarkers rather than isolating biomarker-specific signals.
“It’s a bit like judging a restaurant’s quality by the queue of people waiting to get in,” said Fayyaz Minhas, associate professor and lead author of the study. “Many AI pathology models are doing the same thing, relying on correlations between biomarkers or obvious tissue features rather than isolating biomarker-specific signals.”
Researchers cited the example of predicting mutations in the BRAF gene. Instead of detecting the mutation directly, some systems learned to associate it with related clinical features such as microsatellite instability, or MSI. When those correlated features did not appear together, model accuracy dropped sharply.
Accuracy Drops When Confounding Factors Are Controlled
The study found that AI cancer tools shortcut learning achieved just over 80% accuracy in predicting certain biomarkers. By comparison, tumor grade assessments performed by pathologists reached about 75% accuracy.
When researchers tested models within specific patient subgroups, such as only high-grade breast cancers or only MSI-positive tumors, performance declined substantially. That decline suggests models depended on shortcut signals that disappeared once confounding factors were removed.
Kim Branson, senior vice president and global head of artificial intelligence and machine learning at GSK and a study co-author, said the issue reflects a broader challenge in AI development.
“Predicting a BRAF mutation by looking at correlated features like MSI is often like predicting rain by looking at umbrellas,” Branson said. “If a model cannot demonstrate information gain above a simple pathologist-assigned grade, we haven’t advanced the field.”
Experts Call for Stricter Evaluation Standards
Nasir Rajpoot, director of the Tissue Image Analytics Centre at Warwick and chief executive of spinout Histofy, said the findings highlight the need for stricter evaluation protocols before deploying AI tools in routine care. particularly in light of concerns surrounding AI cancer tools shortcut learning.
“To deliver real and lasting impact, the value of AI-based clinically important predictions must be judged through rigorous, bias-aware evaluation,” Rajpoot said.
The researchers say machine learning still holds promise for research, drug development, and clinical triaging. However, they urge developers to adopt models that explicitly represent biological relationships and causal structures.
Minhas said the study is not a rejection of AI in pathology but a warning about premature clinical adoption.
“Current models may perform well in controlled settings but rely on statistical shortcuts rather than genuine biological understanding,” he said. “Until more robust evaluation standards are in place, these tools should not be seen as replacements for molecular testing.”
The team calls for subgroup testing and comparisons against simple clinical baselines before AI systems move into widespread patient care.
