PROJECT SUMMARY Colorectal cancer is the second deadliest cancer in the United States. Black men and women are 20% more likely to get colorectal cancer and 40% more likely to die from it than most other groups. They are more likely to develop colorectal cancer at a younger age, be diagnosed at a later stage, and are more likely to die of their disease. Patients do not die of primary colon cancer; they die of liver metastases with a dismal <10% of patients surviving past three years. The goal of our parent R01 award is to develop robust machine-learning based imaging biomarkers (radiomics) for personalized treatment of colorectal liver metastases (CRLM). While the development of these markers is extremely promising, machine-learning models are increasingly shown to be biased so understanding the impact of our biomarkers on Black patients that are more likely to be harmed by artificial intelligence (AI) is imperative. This proposal will not solve the totality of disparities that Black patients are experiencing in colorectal cancer but rather, the objectives of the supplemental application are to highlight areas of concern in tandem, as AI-based biomarkers are developed and broadcast this information to AI stakeholders. There are currently no validated imaging biomarkers for CRLM patients, a gap that our parent R01 addresses. Our currently funded activities are focused on computational and statistical biases: with large datasets from multiple institutions (Memorial Sloan Kettering Cancer Center in New York and MD Anderson Cancer Center in Houston), and state-of-the-art AI techniques with carefully controlled data acquisition and annotation, which are vital to the development of robust biomarkers. None of these standard practices, while important, are a panacea against bias: our proposed supplemental activities will begin to address human and systemic biases that are currently overlooked. The proposed studies are innovative because they will examine, for the first time, bias in AI colorectal cancer algorithms in tandem with a socio-technical perspective provided by our new team members. Our work will systematically dispel the myth that there is perfect data by comprehensively evaluating sources of bias and offer a path for AI stakeholders (patients, developers, and users). Given the breadth and complexity of the proposed work, we will not only disseminate our work through traditional methods (publications and talks), but we will also create a podcast. This multi-channel approach will ensure the timely dissemination of important aspects of the work that could inform policy to diverse audiences in an accessible format. Importantly, our work challenges the fundamental assumption in AI development that enough data will overcome any biases, which is simply untrue. The data are already biased.