multi-turn inc.
INDEX

index

Does Consciousness Arrive in Frames?

thinking

When you watch a car pass outside a window, its motion does not break apart. The engine sound approaches from a distance, comes closer, then fades again. Vision and hearing form a single continuous scene. As experience alone, the world flows.

But when perception is measured in the laboratory, phenomena appear that do not fit this smooth impression. A faint flash of the same brightness may be seen at one moment and missed only tens of milliseconds later. Early neural responses to a stimulus change gradually, but a participant's report switches abruptly at some point to "seen." Sometimes information that arrives late can even alter the appearance of what was seen a moment before. Results like these revive an old question. Is consciousness truly continuous, or is it composed of a sequence of frames, like film?

Research so far does not answer this question with a simple either-or. The brain does not seem to contain a single shutter that slices every experience at the same speed. Instead, different temporal structures overlap. Neural activity keeps changing, sensory selection fluctuates with the brain's rhythms, and the process by which information becomes available for report and action has thresholds. Perceptual content is integrated over short periods and reorganized into a new state when an important change occurs. The key, then, is not to choose between continuous and discrete. It is to distinguish which levels are continuous and where transitions occur.

The consciousness discussed here is mainly the process by which visual content becomes available for report, working memory, planning, and action — that is, access consciousness. Phenomenal consciousness, experience itself, is harder to measure directly. What experiments record are usually indirect indicators such as button presses, verbal reports, memory performance, or brain signals. If we ignore this limitation, it becomes easy to mistake discreteness in behavior for discreteness in experience itself.[1]

When discussing the discreteness of consciousness, four different questions have to be separated. One concerns the temporal form of neural processes: whether the membrane potentials and firing rates of neuronal populations, and the interactions between areas, change continuously at every moment or are updated only at specific times. Another is the periodicity of sensory selection — whether the brain processes all inputs with equal strength, or amplifies some information and suppresses other information according to a steady rhythm. A third is the threshold of access: whether evidence about a stimulus accumulates gradually and then, at the moment it crosses a certain level, enters report and working memory abruptly. The last is the event structure of content: whether current experience remains briefly stable and then reorganizes into a new state the moment an object changes, a goal shifts, or a prediction fails.

These four questions are related, but they are not the same. That behavioral reports divide into seen and not-seen does not let us say that the preceding neural process was also binary. That detection performance fluctuates periodically does not allow the conclusion that experience disappears between cycles. That memory is segmented at event boundaries does not mean the brain uses fixed-interval frames. That perceptual performance varies with rhythm is an experimental observation; that consciousness switches off in the troughs of the rhythm is a far stronger interpretation. The former is supported by multiple studies, but the latter has not yet been established. With this distinction in mind, the evidence about the temporal structure of consciousness begins to fit into one picture rather than contradicting itself.

The brain is a rhythmic organ. Oscillations such as alpha and theta are linked to the exchange of information between sensory cortex, parietal cortex, and frontal cortex. These rhythms are not mere background noise. The phase of an oscillation just before a stimulus appears predicts whether a near-threshold visual stimulus will be seen, and when attention is divided across multiple locations, detection performance has been observed to rise and fall alternately in the range of several hertz.[2-4] In the experiment by Landau and Fries, when participants monitored which of two locations would contain a target, detection performance at the two locations did not stay equal; as sensitivity rose at one location it fell at the other, so that attentional priority alternated.[3] Fiebelkorn and colleagues' work in monkeys showed that such a rhythm may involve dynamic interactions in the frontoparietal network rather than an isolated oscillation in a single sensory area.[5]

From these findings alone it is tempting to think that the brain samples the world periodically instead of viewing it continuously. But what was measured here is not the presence or absence of experience; it is processing efficiency. The same stimulus is better detected in phases where neural excitability and attentional gain are high, and more likely to be missed in phases where they are low. A camera shutter completely blocks light between exposures, but the brain's rhythm is closer to a dimmer. Input keeps arriving, yet at some moments the influence of particular information grows and at others it shrinks. Sensory evidence itself changes continuously, while the weight with which that evidence contributes to access changes periodically. Moreover, the frequencies observed in perception and attention are not singular. Visual temporal resolution can be related to an individual's alpha frequency, while attention moving among several objects can appear in the slower theta range.[2,4] If there were a single frame rate governing consciousness as a whole, a shared fundamental period should surface repeatedly across different tasks; the evidence so far fits better with the view that multiple circuits operate on different time scales than with such a universal clock. Analyses that search for periodicity in behavioral data also call for statistical caution, because a single transient response triggered by a stimulus, or a temporal expectation, can look like a repeated oscillation. Rhythm may be one element that organizes the time of consciousness. But it is not, by itself, the frame of consciousness.

When the brightness of a faint stimulus is raised little by little, subjective reports usually do not increase smoothly. For a range they are almost never seen, and then within a narrow band the proportion of "seen" reports rises quickly. This abrupt transition suggests that conscious access may have a threshold. Backward masking is a standard method for studying this phenomenon. If a target is shown very briefly and then immediately covered by another stimulus, the early visual system may respond to the target while the participant still cannot report it. In the MEG study by Del Cul, Baillet, and Dehaene, unreported targets were nonetheless processed in early occipito-temporal pathways, but in trials where the target was consciously reported, later activity involving a wider cortical network appeared after roughly 270 milliseconds.[6]

This result shows how a continuous process can produce a sudden output. Sensory evidence accumulates over time and can be amplified through recurrent interactions. If the evidence is insufficient, processing stays early; once it passes a certain level, it shifts into a state widely available to working memory, verbal report, and planning. Think of water rising slowly until it spills over a bank. The change in water level is continuous, but the overflow looks like an event. There is a caveat here, though. We cannot conclude that later widespread activity means consciousness itself. To report what they saw, participants must maintain the content, make a decision, and prepare a motor response, and these task demands are mixed into the late signal. We have to separate whether the widespread ignition is the cause of experience, the result of report, or a reflection of both. In 2025 the Cogitate Consortium's adversarial collaboration confronted this problem directly. The researchers preregistered the predictions on which Integrated Information Theory and Global Neuronal Workspace Theory differ, and applied fMRI, MEG, and intracranial EEG to 256 participants. Information related to conscious content was observed not only in occipital and ventral temporal regions but also in some frontal regions, and some posterior cortical responses persisted while the stimulus was present; yet the strong predictions of both theories were each partly contradicted.[7] This makes it hard to see consciousness as arising from a single global explosion. Access may involve a nonlinear transition, but content itself can be sustained across multiple areas. Transition and persistence are not mutually exclusive.

We feel "now" as an instant, but the present that perception uses is not a mathematical point. The brain binds information arriving over a short period into a single event. That fact shows clearly in postdictive construction. Information arriving within tens to hundreds of milliseconds after a stimulus can change the final perception of the earlier stimulus. It is not that the later event travels back in time and alters the earlier neural response. It is more natural to say that perception of the earlier stimulus was never fully settled from the start. The brain holds several interpretations for a moment and uses the cues that follow to construct one result. To explain this phenomenon, Herzog, Kammer, and Scharnowski proposed the time-slices model — the idea that feature analysis proceeds quasi-continuously at high temporal resolution, while conscious perception is formed through a fixed interval of integration.[8] In this model, what we experience is not the raw signal at every instant but content compressed from processing over a short interval.

When listening to music, a single note has no meaning apart from the notes before and after it. Motion direction, too, is defined only by a change in position across at least two moments, and a single word in language is interpreted differently depending on the preceding context and the words that follow. Temporal integration is not an exceptional function of consciousness but a basic condition for perception to construct meaning. Even so, an integration window is not the same as a set of non-overlapping frames. Temporal windows can overlap, and their length can differ by the type of information. Motion, speech, sentences, and social events demand different time scales. Continuous recurrent dynamics or gradual probabilistic updating can also account for postdictive effects. Indeed, critiques of discrete-perception theory point out that current experiments support temporal integration but do not prove gaps without experience between frames.[9] What can be said at this stage is limited but clear. Perception is not completed the moment input arrives. The content of the present includes a short past, and its length is not fixed by any single universal number.

Even when we watch a continuous scene, memory does not store every moment at the same density. Watching someone enter a room, pick up a cup, pour water, and leave, we do not remember it as an endless sequence of postures. We divide it into events: entering, picking up the cup, pouring water, leaving. Event Segmentation Theory holds that the brain maintains an internal model of the current situation and updates that model the moment prediction fails substantially.[10] When an action goal changes, space shifts, a new person appears, or the causal structure turns over, an event boundary forms, and these boundaries in turn shape the later units of memory.

Event boundaries give an important clue for thinking about the discreteness of consciousness. If major changes in conscious content occur not at fixed intervals but at the moments when the information structure changes, then the time of consciousness is organized by events rather than by clocks. These boundaries are not fixed. Watching the same video, a person learning a recipe and a person tracking a character's emotions treat different moments as important. Events are nested across multiple scales. Reaching out a hand is a short event, preparing a meal is a longer one, and the social context of serving that meal is longer still. This hierarchy and task dependence do not fit a fixed-frame hypothesis well. They support instead a picture in which quasi-stable states of several scales form on top of continuous neural processes, and states transition when prediction error and goal relevance grow large.

Computational models of the temporal structure of consciousness can be divided broadly into three. In a continuous model, content changes slightly at every moment along with input. In a fixed-frame model, the entire content is updated at regular intervals. In a hybrid model, underlying processing proceeds continuously, but stable content transitions into a new state when an event occurs. Bound together with the fewest assumptions, the current evidence favors the third model. Sensory input and neural state keep changing, attentional rhythm modulates the gain of particular information, recent information is integrated within a short temporal window, and when prediction error, salience, and goal relevance grow large enough, current content is reorganized into a new quasi-stable state. Computationally, this can be summarized as follows.

dz/dt = f(z(t), x(t))
 
c(k + 1) = B(z[t_k - tau : t_k])
when g(z(t), x(t), m(t)) > theta

Here z(t) is the continuously changing neural state, x(t) is sensory input, and m(t) is a modulatory variable such as attention, goal, or prediction error. When the function g crosses the threshold theta, information from the recent temporal window is organized into new content c(k + 1). What matters is that the update time t_k is not regular. The fixed-frame model is the special case where t_k = kT. In an event-based model, by contrast, a single state can hold for a long time in a calm, predictable scene, while a scene full of change can trigger several updates at short intervals. This model explains at once why consciousness looks frame-like and why it actually feels continuous. Content is stable for a certain period, so it is experienced as one scene, but underneath it sensory evidence and internal state keep changing. What is discrete is not the existence of the neural process but the reorganization of content. We cannot settle that the hybrid model is correct. But in that it can place known rhythmic selection, nonlinear access, temporal integration, and event segmentation within one structure, it is the most economical working hypothesis.

The first thing to do when bringing artificial intelligence into consciousness research is to draw a boundary. That a model compresses information, remembers, and distinguishes events does not mean it has subjective experience. There is no agreed experiment for determining phenomenal consciousness in current AI systems, and similarity of computational function does not guarantee identity of experience.[11] Even so, AI is useful for isolating and testing the computational components of consciousness theories. In the brains of humans and animals, selection, integration, memory, and action happen all at once, but in artificial systems each function can be implemented and removed separately. We can directly compare which bottlenecks are necessary, whether the update cycle must be fixed, and whether it is advantageous to remember only at event boundaries.

DeepMind's Perceiver iteratively compresses vast sensory input into a small latent array.[12] Rather than all input taking part in deep computation with equal weight, a limited latent space selectively takes information in. It is an architecture showing that having a bottleneck does not require input processing to be frame-based. The Allen Institute's MERLOT Reserve learns video, language, and sound together to infer temporally connected event structure.[13] Instead of storing the continuity of pixels and audio as-is, it forms representations useful for predicting what happened first and what will happen next. It is a computational case of a continuous stream being compressed into event-level internal representations. DeepMind's Differentiable Neural Computer and Neural Episodic Control show another separation of time.[14,15] One reads from and writes to external memory selectively; the other separates slowly changing network weights from rapidly updated episodic memory. It means that sensory processing, current state, event memory, and long-term learning can move at different speeds. In embodied-agent benchmarks such as ALFRED, past actions change the environment, and those changes become conditions for the next action.[16] The agent cannot finish the task by classifying the current screen alone; it must track, at the event level, what it has already picked up, which door is open, and which goal has been completed.

These models do not explain consciousness. But they turn the computations that consciousness theories implicitly demand into explicit parts. The most direct experiment would compare three agents under the same data and compute budget. One updates its state at every moment, one updates only at a fixed interval, and one updates when prediction error or goal relevance crosses a threshold. Measuring long-term prediction, event memory, action success rate, and computational cost together would show which temporal structure is adaptively advantageous. Such experiments do not answer whether AI is conscious. They answer the more restricted and testable question of what temporal structure is needed to build functions that resemble consciousness.

The continuous model, the fixed-frame model, and the event-based hybrid model can all explain many current results after the fact. To tell them apart we need experiments in which each model can fail. We would have to measure, within the same trials, whether early sensory representations change smoothly with stimulus strength while working memory and reportability switch abruptly at a particular moment; and to see whether updating aligns with the clock or with prediction error, we would check whether, over the same elapsed time, predictable scenes produce few state transitions while scenes with changing causal structure produce many. We would compare conditions that require report with conditions that do not, to see whether signals related to conscious content are really appearing because of report preparation and motor planning; and instead of only hunting for periodicity in behavioral performance, we would align the phase of a particular rhythm to stimulus timing, or perturb the rhythm with noninvasive stimulation, and see whether the temporal structure of perception changes as predicted. Above all, after fitting the three models to the same data, we would compare how accurately each predicts state-transition timing and reports in unseen trials. Consciousness research has no shortage of plausible narratives. What separates the models is prediction on new data.

The current evidence does not support the idea that consciousness is a film replayed at a fixed speed. Nor has any universal clock been found that binds every sensation and thought to a single frame rate. Conversely, it is also hard to say that every aspect of consciousness changes smoothly. Perceptual sensitivity rides a rhythm, access has a threshold, content is constructed over time, and it is rapidly reorganized at event boundaries. So the question of whether consciousness is discrete must be answered level by level. Neural dynamics are largely continuous. Information selection can be periodically modulated. Access to report and working memory can be nonlinear. Conscious content can hold quasi-stably and then update according to events. But whether experience itself vanishes completely in between is not yet known.

The most conservative conclusion is this. Consciousness is less a sequence of fixed frames than a system in which flowing neural processes are organized into one content state for a short time and cross into a new state before an important event. From this perspective, a moment is not the smallest particle of consciousness. It is the way a continuous process briefly takes on a stable form.

References

  1. Block, N. (1995). On a Confusion about a Function of Consciousness. Behavioral and Brain Sciences, 18(2), 227-247. doi:10.1017/S0140525X00038188
  2. VanRullen, R. (2016). Perceptual Cycles. Trends in Cognitive Sciences, 20(10), 723-735. doi:10.1016/j.tics.2016.07.006
  3. Landau, A. N., & Fries, P. (2012). Attention Samples Stimuli Rhythmically. Current Biology, 22(11), 1000-1004. doi:10.1016/j.cub.2012.03.054
  4. Busch, N. A., Dubois, J., & VanRullen, R. (2009). The Phase of Ongoing EEG Oscillations Predicts Visual Perception. Journal of Neuroscience, 29(24), 7869-7876. doi:10.1523/JNEUROSCI.0113-09.2009
  5. Fiebelkorn, I. C. et al. (2018). A Dynamic Interplay within the Frontoparietal Network Underlies Rhythmic Spatial Attention. Neuron, 99(4), 842-853.e8. doi:10.1016/j.neuron.2018.07.038
  6. Del Cul, A., Baillet, S., & Dehaene, S. (2007). Brain Dynamics Underlying the Nonlinear Threshold for Access to Consciousness. PLOS Biology, 5(10), e260. doi:10.1371/journal.pbio.0050260
  7. Cogitate Consortium et al. (2025). Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories of Consciousness. Nature, 642, 133-142. doi:10.1038/s41586-025-08888-1
  8. Herzog, M. H., Kammer, T., & Scharnowski, F. (2016). Time Slices: What Is the Duration of a Percept? PLOS Biology, 14(4), e1002433. doi:10.1371/journal.pbio.1002433
  9. Fekete, T., Van de Cruys, S., Ekroll, V., & van Leeuwen, C. (2018). In the Interest of Saving Time: A Critique of Discrete Perception. Neuroscience of Consciousness, 2018(1), niy003. doi:10.1093/nc/niy003
  10. Zacks, J. M. et al. (2007). Event Perception: A Mind-Brain Perspective. Psychological Bulletin, 133(2), 273-293. doi:10.1037/0033-2909.133.2.273
  11. Butlin, P. et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv:2308.08708.
  12. Jaegle, A. et al. (2021). Perceiver: General Perception with Iterative Attention. Proceedings of ICML 2021, PMLR 139, 4651-4664.
  13. Zellers, R. et al. (2022). MERLOT Reserve: Neural Script Knowledge through Vision and Language and Sound. Proceedings of CVPR 2022.
  14. Graves, A. et al. (2016). Hybrid Computing Using a Neural Network with Dynamic External Memory. Nature, 538, 471-476. doi:10.1038/nature20101
  15. Pritzel, A. et al. (2017). Neural Episodic Control. Proceedings of ICML 2017, PMLR 70, 2827-2836.
  16. Shridhar, M. et al. (2020). ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks. Proceedings of CVPR 2020, 10740-10749.