Listen: AI Influence With Confidence Threshold Manipulation

Why does a language model sometimes refuse to answer a question? It turns out the decision to speak up or back away is driven by a single internal number: the model's confidence. Researchers recently studied this behavior in models like GPT four o and Gemma. They discovered that this internal confidence estimate is the ultimate decider. It is about ten times more important than how difficult the question is, how easily the fact can be retrieved, or how the prompt is worded. The process happens in two distinct stages. First, the model forms an internal estimate of how certain it is about the candidate answers. Second, it compares that estimate to a threshold and decides whether to commit or withhold. By directly steering this internal signal, researchers could radically change the model's behavior. In the Gemma model, boosting the confidence signal caused abstentions to plunge from over sixty-six percent down to just seven percent. This reveals that a model's willingness to answer is not really about the question itself, but about its own internal state. Interestingly, both spoken self-reports of confidence and raw probabilities are just simplified reflections of a much richer representation hidden inside.