"Hallucination" Isn't One Bug. It's Three, and Only One Is Fixable
Hallucination isn't one failure mode — it's three. A test to tell them apart, why benchmarks reward bluffing, and what a 2026 prediction experiment showed.
A model inventing a nonexistent API method and a model confidently dating a future military strike get filed under the same word — hallucination — but they're not the same failure. The useful test is whether the missing context can actually be obtained. When context exists somewhere verifiable, like in a codebase, supplying it fixes the output. When the question itself is unverifiable, like whether a model "understands" sadness, the answer sounds informed but no one, including the model, can confirm what happened. When the information simply doesn't exist yet, like tomorrow's strike date, no prompt engineering, context window, or chain-of-thought closes that gap.
A paper from OpenAI and Georgia Tech (Kalai et al., arXiv:2509.04664) found that leading benchmarks like GPQA and MMLU-Pro score confident wrong answers higher than honest abstention, training models to bluff rather than admit uncertainty. A separate theoretical result (Xu et al., arXiv:2401.11817) proves an LLM cannot learn every computable function, meaning some hallucination is structurally guaranteed for any general-purpose model. The root cause: a next-token generator's probability distribution is never empty, so it always produces an answer — it lacks the internal "I don't understand, stop" state that halts human speech.
A 2026 Jerusalem Post experiment made this concrete: four LLMs asked when the US would strike Iran gave four different, mutually exclusive dates, and grew more specific the harder they were pressed — even though nothing in the real world had gotten any clearer. Pressure raised output resolution without touching accuracy. For engineers, the takeaway is that "fix hallucination" isn't one project; picking the right mitigation depends on which of the three failure types you're actually facing.