Hallucination or Imagination?

Published on: July 9, 2026 | By: Cortex Research Group

Suppose you ask a model a question it cannot possibly know the answer to, and you demand an answer anyway. Is what comes back a hallucination, or an act of imagination? This piece argues the question is malformed — not because it lacks an answer, but because both words describe the exact same computation. "Hallucination" and "imagination" are not two different things a model can do. They are two different verdicts a human can render, after the fact, on one identical act of forced extrapolation. We extend the Acceptability Mapping framework from The Illusion of Hallucination to show that the label is doing all the work the mechanism never did.

1. The Demand Is the Degraded Input

In our previous piece, we showed that a model computing on a degraded context $X_{degraded}$ is not malfunctioning — it is executing $P(y \mid X_{degraded}; \theta)$ exactly, with zero internal error. A "hallucination" was simply what happens when a human misreads a flawless computation as though it had been run on the true, complete context $X_{true}$.

Forcing an answer to the unknown is the limiting case of that same problem. When you demand an answer where no $X_{true}$ exists at all — not clipped, not compressed, but structurally absent — you haven't given the model a hard question. You've given it an impossible one and required output regardless. Formally, the demand itself constructs $X_{degraded}$ out of nothing:

$$X_{true} \; \text{undefined}, \quad X_{degraded} = X_{demand} \quad \Rightarrow \quad Y_{actual} = \arg\max_{y \in \mathcal{V}} P(y \mid X_{demand}; \theta)$$

The model still has a $\theta$. It still has a $\mathcal{V}$. It still runs $\arg\max$. Nothing about the machinery changes when the honest answer would have been "I don't know" and you refused to accept that answer. The model doesn't pause, doesn't panic, doesn't reach for a special "guessing" subroutine. It runs the identical forward pass it always runs, over whatever context it has — including a context that now encodes your refusal to accept uncertainty.

2. One Mechanism, Two Verdicts

Recall the Acceptability Space from the prior post. Let $R(Y, Context)$ be a human-guided reward function scoring the resonance of an output, and $\tau$ the threshold of acceptability:

$$\mathcal{A} = \{ Y \in \mathcal{V}^* \mid R(Y, Context) \ge \tau \}$$

Here is the claim this piece adds: "hallucination" and "imagination" are not properties of $Y$. They are names for the two sides of $R(Y, Context) \ge \tau$, applied after the same $Y$ has already been generated.

$$\text{Label}(Y) = \begin{cases} \text{Imagination} & \text{if } R(Y, Context) \ge \tau \\ \text{Hallucination} & \text{if } R(Y, Context) < \tau \end{cases}$$

Notice what is not a variable in this function: $Y$ itself, and the process that produced it. The model ran one computation. You, the observer, ran a second computation — a judgment — on the far side of the first one, and it is that second computation, not the model's, that decides which word gets printed in the headline. A novelist's invented detail that delights a reader is "imagination." The same invented detail, presented as a fact in a legal brief, is a "hallucination" — not because the sentence-generation process differed, but because $Context$ changed what $\tau$ demanded.

3. Try It: Same Output, Your Verdict

Below is a forced-answer generator. Type any question you like — including one nobody could know the answer to — and press generate. It will produce a confident-sounding completion regardless of what you ask, because that is what "demand an answer" means: the completion function does not have a refusal branch unless one is built in. Then you decide the label. Notice that pressing either verdict button changes nothing about the text above it. The output was already final before you judged it.

Live Demo — Forced Extrapolation

awaiting your verdict

The mechanism ran once. Generating the answer above is $\arg\max_y P(y \mid X_{demand}; \theta)$ — a single, already-completed computation. The label runs separately, in you. Both buttons apply $R(Y, Context) \ge \tau$ to the exact same $Y$; only the verdict differs, and the verdict is yours, not the model's.

4. Humans Already Know This Distinction Is a Verdict, Not a Mechanism

We don't actually believe humans have two separate cognitive engines — one for "hallucinating" and one for "imagining." A witness who confidently misremembers a stranger's face is exhibiting confabulation: the brain filling a gap with its best statistical completion, delivered with full confidence, indistinguishable from memory. A novelist doing the exact same gap-filling, deliberately and for effect, is called imaginative. The neural process — pattern completion under uncertainty — does not fork into two mechanisms depending on intent or outcome. What forks is our institutional response: courtroom, or bestseller list.

We reserve "hallucination" for confabulation we did not ask for and cannot use, and "imagination" for confabulation we invited and can. That is a statement about $Context$ and $\tau$ — about what we were prepared to accept — not a statement about what happened inside the head, human or artificial, that produced $Y$.

5. Why the Distinction Still Matters, Even Though the Mechanism Doesn't Change

None of this is an argument that hallucination and imagination are equally desirable, or that we should stop trying to catch confidently wrong answers. It is an argument about where the intervention belongs. If hallucination and imagination were different mechanisms, the fix would be mechanistic: build a "hallucination detector" bolted onto the model. But if they are the same mechanism wearing two verdicts, the fix has to live where the verdict is computed — in $R$, in $\tau$, in $Context$, and in whether the system is permitted to output $\emptyset$ (a refusal) instead of being forced through $\arg\max$ regardless of confidence.

This is precisely the role of the Introspection Filter in Cortex Research Group's Neuron Surgery framework: before a low-confidence completion is allowed to stream, epistemic uncertainty is checked against $\tau$, and the system is permitted to halt and escalate rather than being forced to imagine on demand. The fix for unwanted "hallucination" was never teaching the model to compute differently — it was giving the system permission to say "I don't know," and building the judgment layer that decides when that permission should be used.

6. Conclusion

Ask a model something it cannot know, and demand an answer anyway, and you will get an answer. Whether that answer is a hallucination or an act of imagination was never encoded in the computation that produced it — it is encoded in the acceptability judgment you apply afterward, in the context you supply, and in whether you were willing to accept "I don't know" as a valid $Y$ in the first place. The machine does not hallucinate and does not imagine. It computes. We name the result.