In their innovative target article, Fleming and Michel have developed an intricate account of how visual ecology, in particular the water-to-land transition, may have sculpted the temporal features of conscious vision. Their analysis starts with the claim that conscious vision is slow. The fate of a percept can still be changed after 450 ms, which “indicated that it was not already conscious.” Conscious vision, F&M claim, “is preceded by a window of unconscious sensory integration that can last up to 450ms.”
But is conscious vision really that slow? Does our brain really wait almost half a second – at least – before visual content comes online? I think that framing the speed of conscious vision in terms of “waiting periods” and “latencies” leads to difficulties. Fleming and Michel may already be aware of these difficulties, as they are familiar with the Orwellian alternative to their Stalinesque interpretation of postdictive effects (Dennett & Kinsbourne, Reference Dennett and Kinsbourne1992). Is visual content consciously perceived but later overwritten without a trace by incoming information (Orwellian rewriting of history) or unconsciously perceived and then edited before it is broadcast to consciousness (Stalinesque show trials)? Ironically, the original point of these “competing” accounts sometimes seems to be forgotten without a trace. Dennett and Kinsbourne (Reference Dennett and Kinsbourne1992) argue that there is no “difference that makes a difference” between these accounts, except for the arbitrary label “conscious” applied either to visual content before or after it is modified by incoming information. The label “conscious” will not make a difference to what someone thinks they saw, nor to their button presses in experiments, nor to the neural processes that lead up to their subjective reports.
To clarify, I am not committing to some version of the overflow gambit. I believe that it does not make sense to talk about a sharp boundary between “not yet” and “already” conscious at the timescale of postdiction. But since F&M’s arguments against the overflow gambit largely constitute their argument for a Stalinesque interpretation of the data (and therefore slow conscious vision), I will examine one of their rebuttals a bit further. F&M discuss Cowan and Greenspahn (Reference Cowan and Greenspahn1995), who found that participants report the midpoint of apparent motion as quickly as its endpoint. The overflow gambit supposedly predicts that participants “should first perceive the end point and then perceive the apparent motion only after this”, which is then reflected in reaction times. According to F&M, the overflow gambit commits us to a model in which postdiction takes longer than stimulus detection. But models of conscious vision that take the vehicle/content distinction seriously can assume that information about temporal order is represented at the same time. The whole representation of the temporal sequence can be conscious at once and therefore reported simultaneously (see White, Reference White2024), regardless of whether the original content was already consciously perceived or not. Hence, there is no reason in principle why an Orwellian account could not make this assumption, like F&M’s Stalinesque interpretation.
The perpetual impasse between Orwellian and Stalinesque explanations naturally leads us to conclude that the latency of conscious vision is an equivocal estimate at best, and a misguided measure at worst. Postdictive effects are not “strong evidence” for long latencies in – let alone lower bounds on – conscious vision. Postdictive effects suffer from similar (if not the very same) conceptual issues as the alternative approaches surveyed by F&M.
But if not latency, what can postdictive effects tell us about conscious vision? With careful experimentation, postdictive effects can reveal the shape of temporal windows of conscious vision (see, e.g., Hogendoorn, Reference Hogendoorn2022), for example, their extent (e.g., how long after X is presented can subsequent information modulate the appearance of X?) or their temporal resolution (e.g., how finely can the order of stimuli be reported within this window?) or their speed (e.g., at what rate is information encoded into the window?). On occasion, F&M allude to temporal windows of integration, where information is accumulated to support perceptual decisions. However, the contents of conscious vision typically represent a window of time, containing changes, sequences, trajectories, and the like (White, Reference White2024). It makes little sense to draw a sharp dividing line across the window and claim something is “not yet” or “already” conscious.
Longer sensory horizons may exert evolutionary pressures on, for example, the extent of temporal windows of conscious perception. If F&M are right (and I think they are) that conscious vision involves building a world model, then a good world model will involve temporal parameters that can be represented inside a temporal window (de Jong et al., Reference de Jong, Voelker, Stewart, Akyürek, Eliasmith and Van Rijn2025), such as “how quickly is that predator moving?”. The organism can take some time to generate a simulation by playing out this world model in their mind (“what if I hide behind that tree?”) and take appropriate action. Crucially, longer temporal windows may not only enable longer pre-diction (i.e., simulation) into the future, but they will also necessitate longer post-diction (i.e., revision) during model building.
However, echoing their Stalinesque interpretation, F&M claim that animals first do model building, and then fix that model for planning. This commits F&M to a discretized, serial computational model of consciousness, which should be testable in principle. But if, as F&M claim, animals have “an experience of what the internal model represents”, then why not during model building? Why wait so long? And in their Figure 4B, when is an animal conscious of their world model at all? During the extended periods of supposedly unconscious, postdictive world building or during offline simulation and planning, which is supposed to be labeled “not real” by perceptual reality monitoring systems?
In the end, the driving force of visual ecology for the timescale of conscious vision is convincingly argued by Fleming and Michel. However, interpreting postdictive effects as strong evidence for long latencies in conscious vision sets them off on the wrong foot. It commits them to a discretized version of temporal consciousness that is neither warranted by the current analysis nor argued against its more continuous counterparts.
You have
Access
In their innovative target article, Fleming and Michel have developed an intricate account of how visual ecology, in particular the water-to-land transition, may have sculpted the temporal features of conscious vision. Their analysis starts with the claim that conscious vision is slow. The fate of a percept can still be changed after 450 ms, which “indicated that it was not already conscious.” Conscious vision, F&M claim, “is preceded by a window of unconscious sensory integration that can last up to 450ms.”
But is conscious vision really that slow? Does our brain really wait almost half a second – at least – before visual content comes online? I think that framing the speed of conscious vision in terms of “waiting periods” and “latencies” leads to difficulties. Fleming and Michel may already be aware of these difficulties, as they are familiar with the Orwellian alternative to their Stalinesque interpretation of postdictive effects (Dennett & Kinsbourne, Reference Dennett and Kinsbourne1992). Is visual content consciously perceived but later overwritten without a trace by incoming information (Orwellian rewriting of history) or unconsciously perceived and then edited before it is broadcast to consciousness (Stalinesque show trials)? Ironically, the original point of these “competing” accounts sometimes seems to be forgotten without a trace. Dennett and Kinsbourne (Reference Dennett and Kinsbourne1992) argue that there is no “difference that makes a difference” between these accounts, except for the arbitrary label “conscious” applied either to visual content before or after it is modified by incoming information. The label “conscious” will not make a difference to what someone thinks they saw, nor to their button presses in experiments, nor to the neural processes that lead up to their subjective reports.
To clarify, I am not committing to some version of the overflow gambit. I believe that it does not make sense to talk about a sharp boundary between “not yet” and “already” conscious at the timescale of postdiction. But since F&M’s arguments against the overflow gambit largely constitute their argument for a Stalinesque interpretation of the data (and therefore slow conscious vision), I will examine one of their rebuttals a bit further. F&M discuss Cowan and Greenspahn (Reference Cowan and Greenspahn1995), who found that participants report the midpoint of apparent motion as quickly as its endpoint. The overflow gambit supposedly predicts that participants “should first perceive the end point and then perceive the apparent motion only after this”, which is then reflected in reaction times. According to F&M, the overflow gambit commits us to a model in which postdiction takes longer than stimulus detection. But models of conscious vision that take the vehicle/content distinction seriously can assume that information about temporal order is represented at the same time. The whole representation of the temporal sequence can be conscious at once and therefore reported simultaneously (see White, Reference White2024), regardless of whether the original content was already consciously perceived or not. Hence, there is no reason in principle why an Orwellian account could not make this assumption, like F&M’s Stalinesque interpretation.
The perpetual impasse between Orwellian and Stalinesque explanations naturally leads us to conclude that the latency of conscious vision is an equivocal estimate at best, and a misguided measure at worst. Postdictive effects are not “strong evidence” for long latencies in – let alone lower bounds on – conscious vision. Postdictive effects suffer from similar (if not the very same) conceptual issues as the alternative approaches surveyed by F&M.
But if not latency, what can postdictive effects tell us about conscious vision? With careful experimentation, postdictive effects can reveal the shape of temporal windows of conscious vision (see, e.g., Hogendoorn, Reference Hogendoorn2022), for example, their extent (e.g., how long after X is presented can subsequent information modulate the appearance of X?) or their temporal resolution (e.g., how finely can the order of stimuli be reported within this window?) or their speed (e.g., at what rate is information encoded into the window?). On occasion, F&M allude to temporal windows of integration, where information is accumulated to support perceptual decisions. However, the contents of conscious vision typically represent a window of time, containing changes, sequences, trajectories, and the like (White, Reference White2024). It makes little sense to draw a sharp dividing line across the window and claim something is “not yet” or “already” conscious.
Longer sensory horizons may exert evolutionary pressures on, for example, the extent of temporal windows of conscious perception. If F&M are right (and I think they are) that conscious vision involves building a world model, then a good world model will involve temporal parameters that can be represented inside a temporal window (de Jong et al., Reference de Jong, Voelker, Stewart, Akyürek, Eliasmith and Van Rijn2025), such as “how quickly is that predator moving?”. The organism can take some time to generate a simulation by playing out this world model in their mind (“what if I hide behind that tree?”) and take appropriate action. Crucially, longer temporal windows may not only enable longer pre-diction (i.e., simulation) into the future, but they will also necessitate longer post-diction (i.e., revision) during model building.
However, echoing their Stalinesque interpretation, F&M claim that animals first do model building, and then fix that model for planning. This commits F&M to a discretized, serial computational model of consciousness, which should be testable in principle. But if, as F&M claim, animals have “an experience of what the internal model represents”, then why not during model building? Why wait so long? And in their Figure 4B, when is an animal conscious of their world model at all? During the extended periods of supposedly unconscious, postdictive world building or during offline simulation and planning, which is supposed to be labeled “not real” by perceptual reality monitoring systems?
In the end, the driving force of visual ecology for the timescale of conscious vision is convincingly argued by Fleming and Michel. However, interpreting postdictive effects as strong evidence for long latencies in conscious vision sets them off on the wrong foot. It commits them to a discretized version of temporal consciousness that is neither warranted by the current analysis nor argued against its more continuous counterparts.
Financial support
The author was funded by the French National Research Agency (AAPG2022, “IntegratedTime”).
Competing interests
None.