Researchers from MIT, Worcester Polytechnic Institute, and Google have proposed a novel method to address persistent bias in artificial intelligence vision models. The study, accepted for the 2026 International Conference for Learning Representations, introduces a technique called Weighted Rotational DebiasING, or WRING. This approach aims to solve the "Whac-A-Mole dilemma," a significant challenge in the field where removing one type of bias inadvertently amplifies others. In high-stakes environments like healthcare, biased AI models can lead to dangerous outcomes. For instance, dermatological tools trained on data skewed toward certain skin tones may fail to accurately identify skin cancer in patients with darker skin. While bias is often attributed to training data, the model architecture itself can also harbor and amplify these issues. Previous attempts to fix this relied on projection debiasing, a post-processing method that removes biased information by projecting it out of the model's representation space. However, this approach often distorts other relationships within the model, effectively replacing one bias with another. Walter Gerych, the paper's first author and currently an assistant professor at Worcester Polytechnic Institute, explained that projection debiasing inadvertently compresses the data, altering all other relationships the model has learned. This led to the Whac-A-Mole phenomenon, where eliminating racial bias in an image retrieval system might unexpectedly increase gender bias. WRING offers a solution by rotating specific coordinates in the high-dimensional space where bias resides. By moving these coordinates to a different angle, the model can no longer distinguish between groups within a specific concept, effectively neutralizing the bias without destroying other learned relationships. Unlike projection debiasing, WRING is minimally invasive and highly efficient. As a post-processing tool, it can be applied to pre-trained models on the fly, avoiding the need to retrain models from scratch, which saves significant time and resources. The research team, which includes MIT graduate students Cassandra Parent and Quinn Perian, Google's Rafiya Javed, and MIT associate professors Justin Solomon and Marzyeh Ghassemi, tested WRING on Vision Language Models (VLMs), specifically the OpenAI OpenCLIP model. Their results demonstrated that WRING significantly reduced bias for targeted concepts without increasing bias in other areas. Currently, the WRING technique is primarily designed for Contrastive Language-Image Pre-training (CLIP) models, which link images to text for classification or search tasks. The researchers acknowledge that extending this method to generative language models, such as those powering ChatGPT, is the logical next step for future development. The work was supported by grants from the National Science Foundation, AI2050, the Sloan Research Foundation, the Gordon and Betty Moore Foundation, and an MIT-Google Computing Innovation Award.