Whatever next? Predictive brains, situated agents, and the future of cognitive science
Open accessThe DOI matches its Crossref record; 387 of the 397 references it cites were found in a registry or an index; 29 of the 152 sources its claims cite were checked for support.
Abstract
Brains, it has recently been argued, are essentially prediction machines. They are bundles of cells that support perception and action by constantly attempting to match incoming sensory inputs with top-down expectations or predictions. This is achieved using a hierarchical generative model that aims to minimize prediction error within a bidirectional cascade of cortical processing. Such accounts offer a unifying model of perception and action, illuminate the functional role of attention, and may neatly capture the special contribution of cortical processing to adaptive success. This target article critically examines this "hierarchical prediction machine" approach, concluding that it offers the best clue yet to the shape of a unified science of mind and action. Sections 1 and 2 lay out the key elements and implications of the approach. Section 3 explores a variety of pitfalls and challenges, spanning the evidential, the methodological, and the more properly conceptual. The paper ends (sections 4 and 5) by asking how such approaches might impact our more general vision of mind, experience, and agency.
Overview
CoChat Summary
TL;DR
Clark argues that hierarchical predictive processing—brains as prediction machines minimizing prediction error via bidirectional cortical cascades—offers the best clue yet to a unified science of mind and action. He critically reviews its evidence, conceptual implications, and limits, concluding it is a powerful but incomplete framework that must be paired with embodied, situated, and socio-cultural approaches.
Key findings
- Hierarchical predictive processing treats perception as top-down probabilistic generative modeling that “explains away” sensory input, with forward connections carrying prediction error and backward connections carrying predictions; action-oriented predictive processing extends the same logic to motor action and active inference.
- The framework unifies perception, action, attention, and learning: attention is modeled as precision-weighting or gain on error units, and the perception/cognition distinction becomes one of scale and precision-modulated influence rather than a stable interface.
- Reviewed evidence includes Rao and Ballard’s model trained on image patches from five natural scenes, which learned oriented edges/bars and extra-classical receptive field effects; Hosoya et al.’s finding that within several seconds about 50% of retinal ganglion cells altered behavior with changing image statistics; Egner et al.’s finding that FFA responses to face and house stimuli became indistinguishable under high face expectation, with a best-fit simulation in which error units contributed twice as much to the fMRI signal as representation units; and Melloni et al.’s finding that reportable conscious percept onset varied by around 100 msec with expectations.
- The account is extended speculatively to delusions and hallucinations in schizophrenia via false, highly weighted prediction errors, and to human cognition as shaped by culturally constructed “designer environments” that feed new statistical structures back into prediction-driven learning.
Methods
- Design: theoretical target article and critical review, not an empirical study; it synthesizes computational neuroscience, Bayesian brain approaches, predictive coding, action-oriented predictive processing, and the free-energy principle.
- Data/sample: no new data; it draws on prior computational models and empirical studies, including Rao and Ballard (1999), Hosoya et al. (2005), Hohwy et al. (2008), Friston and colleagues, Egner et al. (2010), and others.
- Main analysis: conceptual and argumentative assessment of the hierarchical prediction machine framework, its neural plausibility, explanatory scope, and challenges across evidential, methodological, and conceptual dimensions.
Limitations
- The paper states that direct neuroscientific testing of hierarchical predictive coding and its action-oriented extension “remains in its infancy,” that the best current evidence is indirect, and that “no single neuron study has systematically pursued the search for sensory prediction error responses” (quoting Summerfield & Egner 2009); direct unambiguous evidence for functionally distinct error and representation subpopulations is still missing.
- The paper states that the framework does not by itself specify the full human cognitive architecture, including how cognitive labor is divided across areas, what aspects of the world are sensorially coded, how to navigate the explore–exploit continuum, and what representational formats are used; it also says the “desert landscape” replacement of goals and rewards by predictions remains unconvincing, and a full assessment of the free-energy principle is beyond its scope.
- The text makes evident that many supporting findings are indirect, simulation-based, or compatible with alternative accounts; the included commentaries further highlight unresolved challenges concerning attention, adaptation, anti-predictive ERP patterns, the size–weight illusion, and the “dark room” problem.
Model-written summary of a shortened version of the paper’s full text (it is longer than we read). Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures




Verification
Claims the paper makes
27 Valid · 5 Not valid · 84 Not checked · 210 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Recent work con fi rms and extends this assessment, suggesting that humans act as rational Bayesian estimators, in perception and in action, across a wide variety of domains (Berniker & Körding 2008; Körding et al. 2007; Yu 2007).”
What the sources say
- Supportedcites Berniker & Körding 2008 source
The Discussion describes neural coding, sensorimotor integration, sensory cue combination, and decision making within a Bayesian framework, and says this study extends those ideas to motor adaptation and generalization. The Abstract says the Bayesian model predicts many findings from published motor-adaptation studies, supporting the citing sentence’s hedged assessment. (LLM, conf=0.86)
- Partly supportedcites Körding et al. 2007 sourceExperimental — accuracy not yet measured
The Results show that a Bayesian model accounts for animal motor-adaptation behavior and report similar effects in visual-neuron adaptation. The Discussion mentions human saccade adaptation, but does not establish that humans act as rational Bayesian estimators across a wide variety of perceptual and action domains. (LLM, conf=0.85)
- Not matchedcites Yu 2007
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The key idea, however implemented, is that understanding the positive symptoms of schizophrenia requires understanding disturbances in the generation and weighting of prediction error.”
What the sources say
- Not matchedcites Corlett et al. 2009a
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites Corlett et al. 2009b
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Supportedcites Fletcher & Frith 2009 source
In “A hierarchy of prediction and model building” (page unavailable), the authors propose that positive symptoms begin with false prediction errors propagating through a hierarchy. In “From disrupted neurochemistry to beliefs” (page unavailable), they explain how the precision assigned to prediction errors determines their influence on inference, directly supporting the claim about their generation and weighting. (LLM, conf=0.91)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
Every source the paper cites for this claim was checked and supports it.
In the paper
“Transposed (in ways we are about to explore) to the neural domain, this makes prediction error into a kind of proxy (Feldman & Friston 2010) for sensory information itself.”
What the sources say
- Supportedcites Feldman & Friston 2010 source
The Introduction states that, in predictive coding schemes, sensory data are replaced by prediction error because it carries sensory information not yet explained by predictions. This supports the claim that prediction error acts as a proxy for sensory information in the neural scheme described. (LLM, conf=0.94)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“That way (borrowing from work in linear predictive coding -see below) depicts the top-down fl ow as attempting to predict and fully ' explain away ' the driving sensory signal, leaving only any residual ' prediction errors ' to propagate information forward within the system”
What the sources say
- Not checked — no full text of the sourcecites Rao and Ballard (1999) source
- Supportedcites Friston (2005) source
In “Reciprocal connections” (page not available), the paper says backward connections generate predictions of lower-level responses, while forward connections let prediction error drive higher-level units. “Implications for cortical physiology” adds that error is suppressed when bottom-up input accords with top-down predictions. This supports the claim that residual errors, rather than the predicted signal, propagate forward. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Every source the paper cites for this claim was checked and supports it.
In the paper
“This balancing act between cancelling out and selective enhancement is made possible, it should be stressed, only by positing the existence of ' two functionally distinct subpopulations, encoding the conditional expectations of perceptual causes and the prediction error respectively ' (Friston 2005, p. 829).”
What the sources say
- Supportedcites Friston 2005 source
In “Shut up or stop gossiping?” (cited p. 829), Friston says the tension between suppressing prediction error and selecting representations is resolved by positing the two functionally distinct subpopulations named in the claim. (LLM, conf=0.95)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
176 more claims · 15 with their cited sources checked
Create a free account to see all 326 claims and to verify the 5 references still open.
- Probably not a substantial claim. 87
- It may state two separate claims as one. 77
- The quote only partly states this claim. 2
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“The internal model is jointly shaped by the structure of the external environment and the representational capacity of the system itself, which abstracts the experientialist insight into a substrate-independent principle.”
From Understanding large language models demands distinguishing human projection from machine cognition (2026) · cites this paper as [70]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“The ability to make predictions has been hailed as a general feature of both biological and artificial intelligence, cutting across disparate perspectives on what constitutes intelligence (e.g., Ciria et al., 2021; Clark, 2013; Friston and Kiebel, 2009; Ha and Schmidhuber, 2018; Hawkins and Blakeslee, 2004; Littman and Sutton, 2001; Lotter et al., 2016).”
From Predictive Representations: Building Blocks of Intelligence (2024) · cites this paper as Clark, 2013
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“In either case RPP posits that a localized comparator region of the brain does not (necessarily) carry out the comparison of expected and actual sensory information for a given domain.”
From From cognitivism to autopoiesis: towards a computational framework for the embodied mind (2018) · cites this paper as Clark 2013
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“The recurrent exchange of signals between adjacent hierarchical levels resolves prediction error at each and every level, resulting in a hierarchically deep explanation for sensory inputs.”
From Active interoceptive inference and the emotional brain (2016) · cites this paper as [6]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“According to theories of predictive processing, the brain maintains hierarchical generative models that forecast sensory inputs; discrepancies (prediction errors) are resolved by updating posterior beliefs or by acting to change sensory inputs, including through autonomic adjustments that regulate the internal milieu (Clark, 2013; Friston, 2010; Sterling, 2012).”
From Listen to your inner body: embodied emotions in predictive neuroscience and traditional East Asian medicine (2026) · cites this paper as Clark, 2013
Experimental — accuracy not yet measured
31 more claims about this paper
Create a free account to see every claim other papers make about this one.
Citations
Cites
Maximum Likelihood from Incomplete Data Via the EM Algorithm
Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977) · Journal of the Royal Statistical Society Series B: Statistical Methodology · doi:10.1111/j.2517-6161.1977.tb01600.x · Cited by 50,444
The Ecological Approach to Visual Perception
Bornstein, M. H., & Gibson, J. J. (1980) · The Journal of Aesthetics and Art Criticism · doi:10.2307/429816 · Cited by 22,935
Reducing the dimensionality of data with neural networks
Hinton, G. E., & Salakhutdinov, R. R. (2006) · Science · doi:10.1126/science.1127647 · Cited by 20,928
A fast learning algorithm for deep belief nets
Hinton, G. E., Osindero, S., & Teh, Y.-W. (2006) · Neural Computation · doi:10.1162/neco.2006.18.7.1527 · Cited by 16,336
Parallel Distributed Processing: Explorations in the Microstructure of Cognition: Foundations
Rumelhart, D. E., McClelland, J. L., & AU (1986) · doi:10.7551/mitpress/5236.001.0001 · Cited by 15,686
Information theory and statistical mechanics
Jaynes, E. T. (1957) · Physical Review · doi:10.1103/physrev.106.620 · Cited by 12,931
Hutchins, E. (1995) · doi:10.7551/mitpress/1881.001.0001 · Cited by 10,801 · cited 2×
“We may also ask what, if anything, the hierarchical predictive processing perspective suggests concerning situated, world-exploiting agency (Clark 1997; 2008; Clark & Chalmers 1998; Haugeland 1998; Hurley 1998; Hutchins 1995; Menary 2007; Noë 2004; 2009; Rowlands 1999; 2006; Thelen & Smith 1994; Wheeler 2005; Wilson 1…”
BackgroundPhilosophical transactions of the Royal Society of London. B
Press, H. (1886) · Cited by 10,080
Posner, M. I. (1980) · Quarterly Journal of Experimental Psychology · doi:10.1080/00335558008248231 · Cited by 9,670
A neural substrate of prediction and reward
Schultz, W., Dayan, P., & Montague, P. R. (1997) · Science · doi:10.1126/science.275.5306.1593 · Cited by 9,388
Research works printed but not linked: 11
- Pribram, 1980, The orienting reflex in humans, 3
- Velleman, 1989, Practical reflection
- Neisser, 1967, Cognitive psychology
- Pfeifer, 2007, Lecture Notes in Computer Science (LNCS), 4850
- Geissler, 1991, Constraints of mental self-organization: The indirect validation approach toward perception, Estratto da Comunicazioni Scientifiche di Psicologia Generale, 5, 47
- Maher, 1988, Delusional beliefs, 15
- Dennett, 1991, Consciousness explained
- Mumford, 1994, Large-scale theories of the cortex, 125
- Linsker, 1989, Advances in neural information processing systems, vol. 1, 86
- Alais, 2005, Binocular rivalry
- Marr, 1982, Vision: A computational approach
Cited by
Structural and Functional Brain Networks: From Connections to Cognition
Park, H.-J., & Friston, K. (2013) · Science · doi:10.1126/science.1238411 · Cited by 2,106
Interoceptive inference, emotion, and the embodied self
Seth, A. K. (2013) · Trends in Cognitive Sciences · doi:10.1016/j.tics.2013.09.007 · Cited by 1,816
Interoceptive predictions in the brain
Barrett, L. F., & Simmons, W. K. (2015) · Nature Reviews Neuroscience · doi:10.1038/nrn3950 · Cited by 1,766
Knowing your own heart: Distinguishing interoceptive accuracy from interoceptive awareness
Garfinkel, S. N., Seth, A. K., Barrett, A. B., Suzuki, K., & Critchley, H. D. (2015) · Biological Psychology · doi:10.1016/j.biopsycho.2014.11.004 · Cited by 1,624
An integrated theory of language production and comprehension
Pickering, M. J., & Garrod, S. (2013) · Behavioral and Brain Sciences · doi:10.1017/s0140525x12001495 · Cited by 1,569
Metadata
- Authors
- Clark, A.
- Affiliations
- University of Edinburgh (GB)
- Corresponding author
- Clark, Andy
- ORCID
- Clark, Andy
- Venue
- Behavioral and Brain Sciences
- Publisher
- Cambridge University Press (CUP)
- Published
- 2013-05-10
- Volume
- 36
- Issue
- 3
- Pages
- 181-204
- ISSN
- 0140-525X (ISSN-L)
- 1469-1825
- OpenAlex
- W2153791616
- Cited by
- 5,924
- References
- 397
- Citations by year
- 2026: 644 · 2025: 603 · 2024: 438 · 2023: 535 · 2022: 506 · 2021: 488 · 2020: 501 · 2019: 577 · 2018: 398 · 2017: 351 · 2016: 284 · 2015: 255 · 2014: 213 · 2013: 95
- FWCI
- 24.16
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- Behavioral and Brain Sciences (publishedVersion)
- PubMed (publishedVersion)
- Funders
- University of Oxford
- University of Edinburgh
- Ruhr-Universität Bochum
- School of Informatics, University of Edinburgh
- Topics
- Life Sciences › Neuroscience › Cognitive Neuroscience › Embodied and Extended Cognition
- Social Sciences › Arts and Humanities › Philosophy › Free Will and Agency
- Social Sciences › Arts and Humanities › Philosophy › Philosophy and Theoretical Science
- Keywords
- predictive processing, top-down prediction, prediction error, generative models, perception-action coupling, attention, cortical processing, situated agents, agency, experience
- MeSH terms
- Attention, Brain, Cognition, Humans, Learning, Models, Neurological, Perception, Cognitive Science
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Whatever next? Predictive brains, situated agents, and the future of cognitive science. https://app.cochat.ai/papers/10.1017/s0140525x12000477
- The paper
- Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences. https://doi.org/10.1017/s0140525x12000477
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.