Artificial intelligence systems are reshaping knowledge work across fields from hiring to healthcare, yet without appropriate safeguards, these systems risk amplifying existing social biases. The use of dense embeddings from large language models (LLMs) in order to index and query large document collections is no exception. We demonstrate that gender bias is introduced at the retrieval stage through the use of text embedding models, quantifying this bias in Google’s text-embedding-004 and OpenAI’s text-embedding-3-large and text-embedding-ada-002 models for the first time. To address this bias, we introduce FAIR-MASK, a bias mitigation technique based on selective reduction of embedding dimensionality. Evaluated across embedding models from OpenAI and Google, FAIR-MASK reduces gender bias by 25–30% at 50% dimensionality removal. It significantly mitigates gender bias, by up to 33%, across a wide range of dimensionality reduction levels. FAIR-MASK thus enhances the standard industry practice of achieving cost efficiency through dimensionality reduction, providing an additional benefit of improved fairness. Importantly, FAIR-MASK does not compromise search quality, as measured by Recall@K. Unlike previous debiasing methods, this technique does not require complex model development or expensive LLM retraining. This makes it a streamlined, drop-in solution for improving fairness in real-world AI systems.