Shrinking powerful AI models like LLMs to run on everyday devices seems like a win-win—less energy, more privacy. But new research reveals a hidden danger: compression can make these smaller models *more* biased, especially in subtle ways. While compressed models may appear less toxic at first glance (mostly due to the decrease in generation quality), their underlying biases can actually worsen. This is especially true for "representational harm," the kind of bias that shows up when models are used for tasks like classifying text or answering questions. Think of an AI deciding which job applicant is a better fit based on a resume. Compression can amplify existing stereotypes, leading to unfair outcomes. The research also uncovered a troubling disparity: as models shrink, the impact on different social groups diverges, with some groups experiencing more harm than others. Ironically, a smaller model, optimized for standard English, might be better at understanding text than a larger model struggling with dialects like African American English. This raises serious questions about equitable access to technology. Even more concerning is that popular compression methods like "quantization" seem to preserve bias at a moderate compression rate. While these findings raise alarms, they also offer a path forward. Developers can no longer rely on single metrics like "perplexity" to judge a model's performance. Thorough testing for representational harm and bias across different groups is essential before deploying any compressed AI model. Ultimately, creating smaller, safer AI isn't about simply shrinking the size, but about actively mitigating bias in every compressed model. This means developing specialized compression strategies that prioritize fairness, ensuring compressed models are useful, and avoiding unintended consequences.