Large language models (LLMs) like ChatGPT are incredibly powerful tools, but what happens when you need to share sensitive information with them? Imagine applying for a job and wanting AI to help draft your application email. You’d need to give it your resume, full of personal details. This poses a significant privacy risk. Researchers are exploring how to balance the utility of these powerful AI models with the crucial need for privacy. A new research project called PAPILLON (PrivAcy Preservation from Internet-based and Local Language MOdel ENsembles) introduces a clever approach: using a “privacy-conscious proxy.” This involves running a smaller, less powerful LLM locally on your device. This local LLM acts as a gatekeeper, interacting with the powerful, external LLM on your behalf. It carefully crafts prompts for the external LLM, scrubbing out sensitive details while still retaining enough information to generate useful responses. Essentially, it learns to ask the external LLM questions in a way that protects your privacy. To test this, researchers created a benchmark dataset called PUPA (Private User Prompt Annotations) using real user interactions with LLMs, focusing on scenarios involving job applications, financial information, and emails. They then tested different combinations of local and external LLMs in a pipeline. The most effective approach used Llama-3.1-8B-Instruct as the local proxy and GPT-4o-mini as the external LLM. This combination successfully generated high-quality responses while minimizing the leakage of private information. While promising, this research is just the beginning. The gap in performance between using the external LLM directly and using the privacy-preserving proxy still exists. Future work will focus on training specialized privacy-conscious models and refining the pipeline to improve performance and ensure your sensitive information stays protected while you benefit from the power of AI.