Why can't hallucination be fully eliminated from a model trained to predict the next token?

## The objective does not contain truth Pretraining maximizes the likelihood of the next token given the preceding text. What that fits is a distribution over what text looks like. Truth enters only indirectly, through the fact that true statements are more common in the corpus than any particular false one. Where that statistical correlation is strong — widely repeated, consistently stated facts — the model gets them right with high reliability. Where it is weak, the objective offers no mechanism that prefers the true continuation over an equally fluent false one. There is no term in the loss that says "this string is false"; there is only "this string was less likely in the data". That is the root of the phenomenon. Hallucination is not the model malfunctioning; it is the model doing exactly what it was trained to do on inputs where doing that is not enough. ## Generation is harder than judging A useful way to see why the floor is not zero: reduce generation to classification. Imagine a binary question — is this string a valid, correct statement? Any system that can reliably *produce* correct statements could be used to judge candidate statements, so generation is at least as hard as that classification problem. Learning theory gives every classifier a nonzero error rate whenever the features available do not separate the classes, and for arbitrary facts the features available in text are often exactly that: a well-formed sentence with a wrong number in it looks, statistically, like a well-formed sentence with the right number in it. So the generative error rate is bounded below by the misclassification rate of the corresponding validity problem. Anywhere validity is not statistically determinable from the training distribution — arbitrary dates, obscure identifiers, one-off events — the bound is well above zero, and no amount of scale removes it, because scale improves the estimate of the distribution, not the informativeness of the features. ## The singleton intuition The sharpest version of this concerns facts with no redundancy. Consider a class of facts, say the birthdays of minor public figures, and ask what fraction of them appear exactly once in the training corpus. A fact seen once provides essentially nothing to generalize from: the model cannot distinguish it from noise, cannot cross-check it, and has no second phrasing to anchor it. If about a fifth of that category is singleton, then a base model asked such questions is expected to be wrong on at least roughly that share — the error rate on arbitrary facts tracks the singleton rate. This is why hallucination is so lopsided across topics. The model is close to flawless on facts the internet repeats constantly, and unreliable in exactly the long tail where you most wanted help. It is also why a user's experience swings so violently: the failure clusters on rare queries, not uniformly. ## The other two floors Two further limits are less subtle but just as binding. **Capacity.** Parameters are finite and the training corpus is not. Pretraining is lossy compression; details that are cheap to drop get dropped, and the reconstruction that replaces them is a plausible average. This is why fabricated details are usually *typical* rather than random — the model fills the hole with the modal value of that slot. **Cutoff.** Anything that happened after training simply is not represented, and there is no internal marker distinguishing "after my cutoff" from "I don't recall". Asked about a recent event, a model has no way to notice the absence. ## What this changes about how you build Accepting the floor reframes the engineering. The goal is not a model that never errs; it is a system whose residual error is small, measured, and caught where it matters. That means moving facts out of the weights and into the context, where recall becomes reading. It means treating anything with a canonical registry as a lookup, not a memory. It means making abstention a first-class output, since a model that cannot say "I don't have that" must fill the gap. And it means measuring: publish the residual rate per query class rather than claiming a rate of zero, and put verification in front of the surfaces where an error is expensive. It also sets expectations correctly for a stakeholder conversation. "We reduced fabricated citations from 8% to 0.3% and hard-block the rest at the citation resolver" is a defensible claim. "We fixed hallucination" is not, and an interviewer will hear the difference immediately. ## What separates the strong answer Weak answers say the model "tries to be plausible" and stop. Strong answers name a mechanism: the loss contains no truth term; generation inherits a lower bound from the corresponding validity-classification problem; the floor for arbitrary facts tracks how many of them appear only once; capacity and cutoff add their own floors. Then they pivot to what you do about it, because an interviewer asking this is usually testing whether you will over-promise.