Output minimization is a security defense for AI systems that limits how much detail a model reveals in each response. Instead of returning full probabilities, long explanations, internal reasoning, or highly specific data, the system provides only what the user needs.
This matters because rich outputs can help an attacker imitate or probe a model. In model extraction attacks, repeated queries combined with detailed answers can reveal enough signal to train a copy or map the model’s behavior. Output minimization reduces that signal and makes large-scale scraping less useful. In practice, teams may suppress confidence scores, shorten answers, remove sensitive fields, and throttle verbose modes. It is often combined with rate limiting, authentication, logging, and anomaly detection. The goal is not to make an AI unusable, but to reduce the amount of exploitable information exposed through the interface.