Imagine an AI that could predict how severe a COVID-19 infection will be based on a patient's bloodwork. Researchers are exploring just that using large language models (LLMs), the same technology behind ChatGPT. This new research introduces CovidLLM, an AI model designed to predict both the severity and clinical outcomes of COVID-19. Traditional prediction models often struggle with incomplete data—missing values in medical records are common. CovidLLM leverages the unique ability of LLMs to understand context and handle these missing values without relying on potentially inaccurate data imputation. Instead of filling in the gaps, the model is explicitly told a value is missing, allowing it to focus on other available data. Even more innovative is the model’s two-step prediction process. It first predicts the severity of the illness (mild or severe) and *then* uses that prediction, along with the initial data, to predict the clinical outcome (survival or death). This mimics how doctors assess patients and helps refine the prediction. Early results are promising. Compared to traditional methods like AdaBoost, Gradient Boosting, and Random Forests, CovidLLM shows improved accuracy in predicting both severity and outcomes, particularly in predicting death—a crucial aspect for timely intervention. The research also highlights key blood markers associated with severe illness and death, including lymphocyte percentage, D-dimer levels, and C-reactive protein. While the research currently focuses on unvaccinated patients, future work aims to incorporate patient-reported symptoms, time-series data to capture the dynamic nature of the disease, and explore the model’s potential with other illnesses. This research suggests that LLMs could become valuable tools for predicting disease progression and guiding treatment decisions.