Stephen Fleming and the Sensory Horizons Theory of Conscious Vision

Why does conscious vision exist at all, given that most of what the visual system does runs perfectly well without it. Stephen Fleming and Matthias Michel’s answer, developed across a 2026 target article and a follow-up response to commentaries in Behavioral and Brain Sciences, is that conscious vision is too slow to guide immediate action and instead evolved to support offline planning, a function they tie to reality monitoring, the capacity to tell whether an internal signal reflects the world or the system’s own simulation.

The Speed Problem With Action-Guidance Accounts

Most functional theories of consciousness treat conscious perception as if it exists to guide behavior in the moment, seeing a ball and consciously experiencing its arc in order to catch it. Fleming and Michel’s target article, “Sensory horizons and the functions of conscious vision,” challenges that premise directly with timing data. Postdictive effects, cases where a stimulus presented after an event changes what subjects report having consciously seen, reveal integration windows lasting up to 400 milliseconds. A system that needs several hundred milliseconds to settle on what it consciously saw cannot be using that settled percept to guide split-second motor responses, which typically operate on timescales an order of magnitude faster. If conscious vision were for real-time guidance, it is being built far too slowly for the job.

Sensory Horizons and the Water-to-Land Transition

Fleming and Michel locate the alternative function in evolutionary history. Underwater vision is limited to short range by light scattering, which keeps an animal’s effective planning horizon short and favors fast, largely reflexive responses to nearby stimuli. The transition to land dramatically expanded the distance at which visual information becomes available, what the authors call the sensory horizon, and that expansion rewarded a different capacity entirely, the ability to build and manipulate internal models of distant or future states rather than only react to what is immediately present. Conscious vision, on this account, is the marker of a system doing that kind of offline modeling rather than a marker of perception itself.

Their 2026 response to commentaries, “Looking ahead: Sensory horizons and the science of consciousness,” extends the argument to other sensory modalities and to aquatic organisms specifically, defending the claim that the sensory-horizon logic generalizes rather than being an artifact of vision or of terrestrial life. The extension matters because it turns the account into a testable comparative claim, animals and systems with larger effective planning horizons should show more evidence of the offline, model-based processing the theory associates with consciousness, regardless of substrate.

Reality Monitoring as the Higher-Order Link

The mechanism Fleming and Michel propose for how expanded horizons produce consciousness is reality monitoring, the capacity to determine whether a given internal signal was caused by the external world or generated by the system’s own simulation, sometimes called Hamlet’s problem in perception, the problem of knowing when to stop integrating evidence and commit to a model of what is real. This ties the account to higher-order theories of consciousness, the family of views holding that a state becomes conscious when the system represents itself as being in that state, since reality monitoring requires exactly this kind of self-directed representation. Richard Brown’s application of higher-order thought theory to metacognitive reporting in large language models, covered in Richard Brown on higher-order thought theory and metacognition in Claude, and Hakwan Lau’s argument that perceptual reality monitoring specifically should be the target test for AI consciousness, covered in Hakwan Lau on perceptual reality monitoring as a higher-order test for AI, both converge on the same functional target Fleming and Michel arrive at from an evolutionary angle rather than a computational one.

Megan Peters’s work on metacognitive uncertainty, discussed in Megan Peters on metacognitive uncertainty and what it demands of artificial consciousness, adds a further constraint from the same broader research programme Fleming leads at University College London, that a system claiming reality-monitoring capacity should also be able to report calibrated confidence about when that monitoring is likely to fail.

Comparison to The Consciousness AI

This project’s sensory tectum module, built on a recurrent state space model as described on the architecture page, processes visual and audio input before that content competes for Global Workspace broadcast. Reality monitoring in Fleming and Michel’s sense, distinguishing self-generated from externally caused signals, is not a capability the architecture currently implements as a distinct, measured function. The reentrant processing loop and the affective core’s arousal-valence coupling create the kind of internal state that a reality-monitoring mechanism would need to check against incoming sensory data, but no explicit test currently verifies whether the system can tell its own simulated states apart from directly sensed ones. Fleming and Michel’s account suggests that gap, not the sensory tectum’s processing speed, is the more relevant one to close if this architecture is meant to engage seriously with what conscious vision is functionally for.

What the Account Does and Does Not Settle

Fleming has not extended the sensory horizons argument into a direct claim about artificial systems in print, and the theory is explicitly about why conscious vision evolved in biological organisms with a specific evolutionary history, not a general recipe for building consciousness into a machine. What it offers instead is a functional target that does not depend on matching perception speed or reporting behavior. If reality monitoring, not raw processing speed, is the thing consciousness is for, then evaluating an AI system’s claim to relevant functional capacity means asking whether its architecture needs, and can perform, the specific job of distinguishing self-generated from externally caused internal states in service of planning, a narrower and more falsifiable question than asking whether the system behaves as if it sees. The broader landscape of competing functional and structural accounts is indexed on the state of the field on AI consciousness.