Hallucination Is Generative Memory With Its Verifier Turned Down: One Constraint Axis Links Dreaming Sleep and LLM Confabulation

Abstract

Dreaming and large language model hallucination are usually studied apart, one as biology and the other as engineering. We argue they are two expressions of a single computation: generative reconstruction from distributed memory under incomplete constraints. Neither the sleeping brain nor an autoregressive model retrieves stored records; both synthesize output by recombining learned representations, and both produce fluent, structured content that can depart from fact. We organize the two systems with a four-stage account, encoding, latent representation, generative reconstruction, and reality verification, and argue that the difference between a reliable output and a hallucination or a dream is governed by one variable: the strength of the constraints and verification acting on the generator. In the brain these are sensory evidence and prefrontal control, both attenuated in REM sleep; in a model they are grounding and external checking, both absent in free decoding. We state where the mapping holds and where it breaks (it is computational, not phenomenal, and dreaming may be adaptive where hallucination is an unselected byproduct), and we separate our claim from accounts that treat dreams as training-time regularization or hallucination as narrativity. On this view, hallucination is the expected behavior of a generative memory whose verification channel is turned down, not a discrete defect. We recast mitigation as restoring verification rather than removing generation, connect it to human reality monitoring and to lucid dreaming as mid-generation metacognitive control, and set out falsifiable predictions and a two-way research agenda linking neuroscience and machine learning. We anchor that stance to a measurement reported separately, in which a chart distortion’s presence and geometry are decodable from a vision-language model’s activations and its magnitude is read out more accurately by a probe than by what the model states, while writing the recovered signal back in leaves the output unchanged. A small test of the axis on 300 PopQA questions, with its pass criteria fixed before any model was run, finds the hallucination rate of two language models falling at every step from forced closed-book answering to open-book answering with a verifier (50.7 to 7.0 percent and 46.0 to 5.3 percent), and falling with entity popularity when unconstrained; the last step is marginal for one model and is paid for in correct answers in both. This is a perspective paper: it contributes a framework and testable claims, with one small test of the axis and no full study of its own. Perspective paper, sole author. 16 pages, 2 figures, 2 tables, 83 references. The vignette plan, code, cached model responses and results are released with the paper. Section 5.4 anchors to the companion measurement at 10.5281/zenodo.22134653.