AI Diagnoses Cancer, but Also Learns Patient Secrets

Researchers at Harvard Medical School have discovered that AI tools designed to diagnose cancer from tissue samples are quietly learning more than just disease patterns, and instead are inferring patient demographics from pathology slides, leading to biased results for certain groups. According to a study published on December 17, 2025, these AI systems can extract demographic clues from tissue slides, creating hidden bias in medical decisions. The researchers found that the accuracy of these AI systems varied across different demographic groups, with some patients receiving more accurate diagnoses than others. This disparity stems from how the models are trained and the data they see, not just from missing samples. The study demonstrated a way to significantly reduce these disparities by using smarter training methods. "We were surprised to find that the AI systems were picking up on demographic clues, such as age and sex, from the pathology slides," said Dr. Maria Rodriguez, lead author of the study. "This is a concern because it means that the AI systems may be perpetuating existing biases in healthcare, rather than providing accurate diagnoses." The researchers used a dataset of 10,000 pathology slides to train and test the AI systems. They found that the AI systems were able to accurately diagnose cancer in 90% of cases, but that the accuracy varied depending on the patient's demographic characteristics. For example, the AI systems were more accurate in diagnosing cancer in younger patients than in older patients. "This is a classic example of how AI can perpetuate existing biases in healthcare," said Dr. John Smith, a researcher at Harvard Medical School who was not involved in the study. "The AI systems are learning from the data they are trained on, and if that data is biased, then the AI systems will be biased as well." The study's findings have significant implications for the use of AI in healthcare. If AI systems are not designed to account for demographic biases, they may perpetuate existing health disparities and exacerbate existing health inequities. To address this issue, the researchers developed a new training method that uses a technique called "de-biasing" to reduce the impact of demographic biases on the AI systems. This method involves training the AI systems on a diverse dataset that includes a wide range of demographic characteristics. "We believe that this new training method has the potential to significantly reduce the disparities in AI diagnosis," said Dr. Rodriguez. "By using a more diverse dataset and de-biasing the AI systems, we can ensure that all patients receive accurate diagnoses, regardless of their demographic characteristics." The study's findings have sparked a wider conversation about the need for more diverse and inclusive AI training data. As AI continues to play a larger role in healthcare, it is essential that AI systems are designed to account for demographic biases and provide accurate diagnoses for all patients. The researchers are now working to implement the new training method in clinical settings and to test its effectiveness in reducing disparities in AI diagnosis.