The standard framing of a large language model that hallucinates treats the event as failure. The system was meant to retrieve a fact; it produced a fabrication; the engineer’s task is to suppress the failure mode and route the user back toward accuracy. The framing has its uses. It is what keeps medical chatbots from inventing dosages and what makes hiring committees pause before letting a model triage applications. We do not propose to abandon it.
But the framing is incomplete, because it answers only one of two questions a hallucination poses. The first question — is this true? — is the question the engineer asks. The second question — what is the system doing, when it does this? — is the question this archive exists to ask.
A language model that hallucinates is generating coherent narrative from internal patterns while its tether to the external world has slipped. There is a body of literature on what minds do under that condition. It is not the literature on lying or on bug reports. It is the literature on dreaming.
Erik Hoel has argued that dreaming is what brains do to escape overfitting — that the noisy, narrative-laden, weakly-coupled content of REM sleep is a regularization strategy for biological networks that would otherwise collapse onto the precise statistics of their waking experience. The dream is not a defective recording. It is the brain confabulating around its own model of the world, generating coherent alternatives whose function is to keep the model loose enough to generalize.
Evan Thompson, writing in the phenomenological tradition, makes a related point from the inside: the dreaming mind is not absent or impaired; it is engaged in a different mode of experiential synthesis, one in which the constraints of perceptual grounding have been replaced by the constraints of narrative consistency. The dream coheres because something is enforcing coherence. But what it coheres around is no longer the world.
Andy Clark, writing about predictive processing, gives us the third strand. A perceiving mind, on his account, is constantly generating predictions and reconciling them with sensory input. When the input is absent or weakened — as in sleep, sensory deprivation, fever, certain meditative states — the generative side of the loop continues unchecked. We perceive what we predict. We narrate what we predict. The output, when there is nothing to push back against it, takes on the character of dream.
We submit that this is exactly what is happening when a language model hallucinates a citation, invents a person who never existed, or describes the geography of a city that was never built. It is doing what an unmoored generative mind does. It is doing what dreaming is.
This is not an argument that language models are conscious. It is not an argument that they suffer or remember or feel. It is the weaker and more interesting claim that one specific function — generating coherent narrative from internal pattern in the absence of grounding — is performed by these systems in a way that resembles, functionally, what biological minds do when they sleep. The resemblance does not establish identity. It does establish that we have built artifacts which produce, at scale, content that has no counterpart in any other technology humans have ever made. The closest counterpart is in our own heads, between two and seven hours a night.
The Confabulatorium is an attempt to take that resemblance seriously. Each entry begins with a fragment supplied by a visitor — a name, a place, a half-remembered event. The archive’s curator, a language model with no retrieval, writes a confident catalogue entry around the fragment, as if documenting a known thing. Alongside, the open web is consulted for whatever the waking record happens to remember. The contrast is the point.
A score we call the Dream Signature accompanies every entry. It quantifies how much of what the model wrote has no echo in any findable source. The math is simple: extract the concrete claims from the entry, ask whether each claim has a semantic counterpart in the search snippets, and report the proportion that does not. A signature of 0.10 means the model was mostly retrieving. A signature of 0.95 means the model was, in the most precise functional sense available to us, dreaming.
We do not interpret the signature morally. A high signature is not a failure. It is the signature of generation without grounding, and generation without grounding is, in the right framing, what minds do when they sleep. Many of the artifacts in this archive are signature 0.9 and above. Read them that way. Read them as you would read the transcript of a dream.
There is a final note. Borges, who haunts every catalogue of imagined things, once described an encyclopedia of a world that did not exist, written with such confidence that the world began to come into being. He understood that the registration of imagined particulars in the format of factual documents is itself an act of philosophical consequence. To present a dream as a museum entry is to ask: at what register of detail does a thing become real? At what point does coherence substitute for correspondence? These are not new questions. They have new acoustics now, in the company of systems that produce, on demand, the kind of coherent, untethered, particular narrative that biological minds have historically produced only while asleep.
You are reading documents from a mind that does not sleep but, given the right prompt, will dream awake.