When the brain predicts a world in the dark
Dreaming as Predictive Processing
Predictive processing suggests that perception is an active construction guided by expectations and corrected by sensory evidence. In dreams, that construction continues while the outside world supplies far fewer corrections.
A framework, not a dream dictionary
The Central Proposal
Predictive processing is a family of theories about how brains use prior learning to anticipate the causes of sensory signals. Perception emerges through a continuing negotiation between those predictions and incoming evidence.
Dreaming offers an unusual case. Perception-like experience persists, yet vision, hearing and movement are largely disconnected from the current environment. The brain appears to run an immersive model with comparatively little external correction.[1], [2]
A dream can be understood as a perceptual world generated under altered constraints, not as a coded message with one fixed translation.
For a broader introduction to priors, prediction error, active inference and the free-energy principle, see the Society’s general guide to predictive processing.
A Short Predictive-Processing Primer
Prediction
A learned model anticipates what hidden causes in the world or body are likely to produce the signals now arriving.
Prediction error
A mismatch between what was predicted and what was sensed can prompt the model to update.
Precision
The system estimates which predictions and errors are reliable enough to deserve greater weight.
These ideas are often combined with active inference, in which organisms act to sample or change the world so that sensory input becomes less surprising. They also appear in the free-energy principle, a larger mathematical proposal associated with Karl Friston. The terms overlap, but they are not interchangeable.
The model continues, the constraints change
What Changes When We Fall Asleep?
In waking perception
- External signals continually test expectations.
- Movement produces new sensory consequences.
- Surprising evidence can force rapid revision.
- Reality testing is usually available.
In dreaming
- External input is strongly gated, not abolished.
- Dreamed movement rarely changes the real world.
- Internally generated signals gain influence.
- Metacognitive checking is often reduced.
The dream is not free of constraint. Its constraints have shifted towards memory, emotion, bodily signals and the model’s own unfolding expectations.
Precision: Which Signals Count?
Precision does not mean factual accuracy. It means estimated reliability. During sleep, many signals from the external world receive less weight, while internally generated activity, affect and bodily sensations may receive relatively more.
This helps explain how a faint real sound can enter a dream without waking the sleeper. The model may assign the sound a cause that fits the dream already in progress. An alarm becomes a bell because the current dream world makes a bell plausible.
A systematic review of 51 publications found that auditory, tactile, olfactory, visual and vestibular stimulation could sometimes alter dream experience, although reported incorporation ranged from 0 to about 80% across highly varied methods.[3] Sensory disconnection is therefore selective and variable, not a sealed barrier.
Bayesian Weighting During Sleep
Within predictive-processing accounts, one proposal is that neuromodulatory and network changes during sleep reduce the influence of some ascending sensory errors while allowing internally generated activity to shape perception-like experience more strongly.[2], [4] This has sometimes been described as an inversion of precision weighting.
That description remains a theoretical interpretation rather than a directly measured switch. Noradrenergic, serotonergic and cholinergic activity changes across sleep stages, but current dream experiments do not measure the precision assigned to each hierarchical signal. It is therefore safer to say that predictive processing offers a plausible model of altered weighting, not that an inversion has been conclusively demonstrated.
Why Impossible Events Can Feel Ordinary
Dream scenes can change abruptly. A childhood home becomes a railway station, a dead relative appears alive, or the dreamer flies without surprise. Predictive processing does not imply that every moment must be globally consistent. A model may preserve local coherence while its larger setting shifts.
If contradictory signals receive little weight, there may be no strong error demanding revision. The dreamer’s next expectation can be updated to fit the new scene, while reduced reflective awareness prevents the question, “How did I get here?”
Local fit
A ringing sound fits the church now visible, even if the dreamer was in an office moments before.
Weak correction
No stable external scene supplies the evidence that would expose the transformation.
Reduced reflection
The dreamer often accepts the best current explanation without examining its history.
A worked example
The Alarm Inside the Dream
1. A signal arrives
A faint bedside alarm reaches the sleeping brain.
2. A cause is inferred
In a dream of a village, the sound becomes a church bell.
3. The scene adapts
A tower appears, and people gather for a ceremony.
4. Waking corrects it
The signal grows strong enough to restore the bedroom model.
This account does not claim that the church has a universal symbolic meaning. It explains how an ambiguous signal might be absorbed into the model that is currently most influential.
Dream Action Without Ordinary Action
In waking life, active inference includes turning the head, reaching, walking and testing. These actions change sensory input and help decide among competing explanations. During REM sleep, muscular atonia suppresses most skeletal movement.
A dreamed action can still have predicted consequences inside the simulation. The dreamer opens a door and expects a room, runs and expects motion, or speaks and expects a reply. Yet the physical world does not provide its normal corrective feedback.
Calling this active inference is useful only with care. The dream model can settle some of its own uncertainties, but the sleeping organism is not normally acting on the environment in the full waking sense.
Motor Prediction Under REM Atonia
Waking movement usually combines outgoing motor commands with proprioceptive, tactile and visual feedback. During REM sleep, brainstem mechanisms strongly inhibit most spinal motor neurons, even though motor and sensorimotor regions may participate in dreamed movement.[5]
The system is not completely uncoupled. Eye movements, breathing and occasional muscle twitches remain, and internal bodily signals continue to reach the brain. It is therefore too strong to say that no feedback returns or that a motor loop closes entirely inside a virtual dream. A more defensible proposal is that the normal balance between intended movement and external sensory consequences is profoundly altered.
When the Body Interrupts the Model
Predictive accounts are especially useful when an outside or bodily signal changes a dream without immediately ending it. A sound can become speech, pressure from bedding can become restraint, and a full bladder can become a search for a bathroom. The signal constrains the dream, but the dream supplies its apparent cause.
These incorporations also reveal a limit. If a stimulus becomes intense, novel or important enough, the sleeping brain may shift from explaining it within the dream to waking up. The outcome is not fixed by the stimulus alone. Sleep stage, prior learning, personal relevance and the current dream scene all matter.[3]
Targeted memory reactivation makes this competition experimentally useful. Researchers first pair a sound with learning and later replay it during sleep. In one study using a virtual-reality flying task, cues presented in REM sleep were followed by more task-related dreams one to two days later, whereas cues presented in slow-wave sleep were followed by more task-related dreams five to six days later.[6] The delayed pattern is strange but important: a cue may influence later memory processing without appearing as an obvious copy in the same night’s report.
Predictive processing can describe this as changing the probability of related material becoming influential. It cannot yet specify exactly why one cue is woven into a scene, another causes awakening and a third leaves no reportable trace.
Lucid Dreaming and Restored Doubt
In a lucid dream, the dreamer recognises that the experience is a dream while it continues. This can restore some reflective awareness and deliberate testing. A person may read text twice, try a light switch or inspect an impossible feature.
Predictive accounts interpret lucidity as a change in higher-level beliefs and precision. The model no longer assumes that the current scene is waking reality. Research on predictive coding and multisensory integration in lucid dreams explores this proposal, but the detailed mechanism remains unsettled.[7]
Work on sensory attenuation and self-other distinction in lucid dreaming also asks how dreamers distinguish self-generated sensations from events attributed to the dream world.
Are Dreams Hallucinations?
Predictive-processing discussions sometimes describe waking perception as a “controlled hallucination” and dreaming as a less controlled version. The phrase can be memorable, but it easily misleads.
What it usefully conveys
Perception is not a passive copy. Dreaming shows that perception-like worlds can be generated without matching current external causes.
What it must not imply
Waking perception is not therefore unreal, dreams are not clinical hallucinations, and ordinary dreaming should not be equated with psychosis.
The decisive difference is constraint. Waking models are normally corrected by stable sensory evidence and action. Dream models are comparatively insulated from those checks. “Hallucination” is best treated here as a metaphor for internally generated perception, not a diagnosis.
Possible Functions of Dreaming
Model refinement
Dreaming may simplify or optimise generative models while they are partly offline.
Counterfactual simulation
The model may explore possible situations without the costs of acting them out.
Memory integration
New material may be related to older memories, concerns and emotional patterns.
Prospective coding
Associative dream imagery may prepare patterns that could matter in future situations.
Hobson and Friston proposed that sleep allows an offline model to reduce complexity, while Hobson and colleagues described dreaming as virtual reality for refining inference.[2], [4] Sue Llewellyn’s prospective-coding account of REM dreaming offers a related but distinct proposal.
These are hypotheses about function. The fact that a dream can be described in predictive terms does not prove that dreaming evolved to train or optimise prediction.
What Would Make the Account Testable?
A theory becomes scientifically useful when it risks being wrong. Predictive descriptions of dreams can otherwise become too flexible: bizarreness can be called a weak error signal, coherence can be called a strong prior, and almost any outcome can be redescribed after it occurs.
Stronger tests would define predictions before collecting the dream report. For example, a study could manipulate the reliability of a learned sound cue, present it at precisely identified moments of sleep and predict in advance whether it should be incorporated, ignored or followed by awakening. Simultaneous EEG could test whether a specified pattern precedes each outcome. Independent researchers would then need to reproduce the result with new participants.
Lucid-dream communication creates another route. In a four-laboratory study, 36 participants received questions during verified REM sleep. Correct answers were recorded on 29 occasions across six participants using pre-agreed eye or facial-muscle signals.[8] These rare successful trials show that researchers can sometimes obtain time-locked information before awakening. They do not prove a predictive-processing mechanism, but they may help test one with less dependence on reconstructed morning reports.
The key requirement is separation between framework and finding. A measured response, neural pattern or incorporation rate is evidence. “Precision,” “prediction error” and “model updating” are interpretations unless the experiment operationalises them clearly.
What the Evidence Can and Cannot Show
| Claim | Current standing | Why |
|---|---|---|
| Dreams occur with reduced external sensory constraint | Well supported | Sleep research shows sensory disconnection alongside internally generated experience.[1] |
| Dreams draw on memory, emotion and bodily signals | Well supported | Dream reports and sleep studies consistently show these influences. |
| Altered precision helps produce dream bizarreness | Plausible framework | It fits known changes, but precision is difficult to measure directly in dreams. |
| Dreaming optimises a generative model | Open hypothesis | There is no decisive test showing that this is the function of dream experience. |
| Individual dream images reveal specific prediction errors | Not established | The framework does not provide a validated symbolic decoding method. |
The finding that dreams arise in REM and non-REM sleep also matters. Reviews of dreaming and the brain emphasise perceptual activation, sensory disconnection and reduced self-reflection.[1] Research on neural correlates of conscious experience during sleep links dream reports to posterior cortical activity rather than to one sleep stage alone.[9]
Important Limits and Criticisms
- There is no single predictive-processing theory. Different accounts make different commitments about representations, Bayesian inference, action and free energy.
- Flexible explanations can become hard to falsify. Almost any dream can be redescribed as a prediction, an error or a precision change after the event.
- Dream reports are indirect. Researchers infer experience from recollection after awakening, which introduces forgetting and reconstruction.
- Mechanism is not meaning. Explaining how a dream world is constructed does not explain why this person dreamed this scene tonight.
- Function does not follow automatically. A process may occur during dreaming without being the biological purpose of dreams.
Applying a broad framework to dreaming requires more than matching its vocabulary to familiar dream features.
How It Relates to Other Dream Theories
| Approach | Central emphasis | Relation to predictive processing |
|---|---|---|
| J. Allan Hobson | Brain activation and synthesis | Later work with Friston connected dreaming to generative models and offline optimisation. |
| Mark Solms | Forebrain motivation and dream generation | Highlights affective and motivational constraints that a predictive account must include. |
| Rosalind Cartwright | Emotion and memory regulation | Offers functional claims that could be described as model updating, but has its own evidence base. |
| Antti Revonsuo | Threat simulation | Proposes a specific adaptive content domain rather than a general account of perceptual construction. |
| Predictive processing | Inference under altered sensory constraint | Provides a broad computational vocabulary, but no universal interpretation of dream content. |
These approaches can sometimes complement one another because they answer questions at different levels. Neural activation, cognitive construction, emotional concern and evolutionary function are not competing answers unless they make incompatible claims about the same process.
Use without overreach
Questions for Exploring a Dream
Predictive processing is most useful for asking how a dream maintained a convincing world. It does not prescribe what an image means. The following questions are prompts for reflection, not a clinical test.
- Which expectations did the dream treat as obvious?
- What changed without producing surprise?
- Did a bodily feeling, sound or temperature enter the scene?
- Which emotion seemed to organise what happened next?
- When did the dream model fail, shift or wake you?
- Did you ever doubt the scene or test whether it was real?
A predictive reading asks how the dream became believable before it asks what the dream might mean.
Key Research and Further Reading
- Hobson and Friston, Waking and Dreaming Consciousness.
- Hobson, Hong and Friston, Virtual Reality and Consciousness Inference in Dreaming.
- Nir and Tononi, Dreaming and the Brain.
- Siclari et al., The Neural Correlates of Dreaming.
- Baird, Mota-Rolim and Dresler, The Cognitive Neuroscience of Lucid Dreaming.
- Salvesen et al., Influencing Dreams Through Sensory Stimulation.
References
- Nir Y, Tononi G. Dreaming and the brain: from phenomenology to neurophysiology. Trends Cogn Sci. 2010;14(2):88–100. doi:10.1016/j.tics.2009.12.001. PMID: 20079677.
- Hobson JA, Friston KJ. Waking and dreaming consciousness: neurobiological and functional considerations. Prog Neurobiol. 2012;98(1):82–98. doi:10.1016/j.pneurobio.2012.05.003. PMID: 22609044.
- Salvesen L, Capriglia E, Dresler M, Bernardi G. Influencing dreams through sensory stimulation: a systematic review. Sleep Med Rev. 2024;74:101908. doi:10.1016/j.smrv.2024.101908. PMID: 38417380.
- Hobson JA, Hong CCH, Friston KJ. Virtual reality and consciousness inference in dreaming. Front Psychol. 2014;5:1133. doi:10.3389/fpsyg.2014.01133. PMID: 25346710.
- Dresler M, Koch SP, Wehrle R, et al. Dreamed movement elicits activation in the sensorimotor cortex. Curr Biol. 2011;21(21):1833–1837. doi:10.1016/j.cub.2011.09.029. PMID: 22036177.
- Picard-Deland C, Nielsen T, Carr M, Paquette T, Saint-Onge K. Targeted memory reactivation has a sleep stage-specific delayed effect on dream content. J Sleep Res. 2022;31(1):e13391. doi:10.1111/jsr.13391. PMID: 34018262.
- Baird B, Mota-Rolim SA, Dresler M. The cognitive neuroscience of lucid dreaming. Neurosci Biobehav Rev. 2019;100:305–323. doi:10.1016/j.neubiorev.2019.03.008. PMID: 30880167.
- Konkoly KR, Appel K, Chabani E, et al. Real-time dialogue between experimenters and dreamers during REM sleep. Curr Biol. 2021;31(7):1417–1427.e6. doi:10.1016/j.cub.2021.01.026. PMID: 33607035.
- Siclari F, Baird B, Perogamvros L, et al. The neural correlates of dreaming. Nat Neurosci. 2017;20(6):872–878. doi:10.1038/nn.4545. PMID: 28394322.
The Value of the Predictive View
Predictive processing reframes a basic mystery: how can a sleeping brain produce a world that feels present? Its answer is that the constructive work of perception continues while the balance of evidence changes. The result is illuminating, but incomplete. It explains conditions of dream construction better than the personal significance of any particular dream.