Fleming and Michel (F&M) suggest that conscious vision reflects the output of a reality monitoring (RM) mechanism that evolved to separate model-based planning from a “real now” percept used to update world models. They propose that this lens explains why conscious vision feels unitary and why it is relatively slow to appear and update. Here, I constructively challenge F&M’s evolutionary argument by exploring how slowness and unity might arise without long integration windows or selecting a single “reality” for planning.
To begin, let’s consider how RM might select or tag a given first-order representation as “real now.” F&M suggest this occurs when sensory integration has resulted in a “reliable-enough” first-order estimate, which is tagged as the single reality. F&M do not use Bayesian language, but their description implies a RM decision policy (Peters, Reference Peters2022; Peters & Asrari, Reference Peters and Asrari2025) which takes as “real” the only (or first) first-order representation to arrive at this reliability threshold, that is, a winner-take-all, maximum a posteriori (MAP) policy: F&M say, “subjects presumably feel like they are experiencing a single image at a time,” “reflecting the outside world as it is now.” So one representation is tagged as “real,” and updating is slow because integration to achieve stability takes time.
If this explanation is correct, evolution likely optimized the RM decision policy. A MAP policy would optimize for minimizing error in selecting which first-order representation is “real” and then updating this tag when it later determines a subsequent first-order representation has become “reliable enough.” This not only implies that the most important function of RM is to periodically identify the one “real” first-order representation, but also that the cost of getting this wrong is symmetric (Berger, Reference Berger2014; Bishop, Reference Bishop2006): tagging a hallucination as real would have the same consequences as tagging a percept as hallucination.
But could RM’s slowness reflect something other than integration to support model-based planning? If so, those alternative sources of slowness may also mask alternative decision policies, which themselves could reveal how evolutionary pressures shaped RM’s emergence, function, and link to consciousness.
To explore these possibilities, let’s start by introducing “stability priors,” formalized in decision neuroscience as hazard-rate priors about environmental rate of change: the current state estimate partially depends on the previously inferred state, modulated by an estimate of environmental volatility, h (Behrens et al., Reference Behrens, Woolrich, Walton and Rushworth2007; Gold & Shadlen, Reference Gold and Shadlen2007; Nassar et al., Reference Nassar, Wilson, Heasly and Gold2010). If h is small (conservative), the world is assumed stable and interpretations are “sticky”; if h is large, the model more readily switches interpretations. In perception, conservative hazard rates can explain the relatively slow rate of binocular rivalry switching: “under the hood” the brain is sampling perceptual evidence and computing posteriors over world states quickly, but the system resists updating its world-state estimate unless the balance of evidence shifts by “enough” (Gershman, Vul, & Tenenbaum, Reference Gershman, Vul and Tenenbaum2012).
Hazard-rate priors usually describe expectations about environmental stability, but RM monitors the stability of first-order representations. So the RM system must also determine the source of any first-order variance: is the world changing, or just the competing first-order representations? Is variance due to meaningful stochasticity, or is it just noise? This need for variance source monitoring also implies two hazard-rate priors: expected environmental stability and expected first-order representational stability.
Either of these hazard-rate priors could make RM-based updating of consciousness slow. If I expect slow environmental change, I may compute world-state posteriors faster than the 300-500 ms evidence integration loop suggested by F&M, but will not update perceptual content very often. Likewise, if I expect “reality”-tagging to change slowly, I may rapidly compute a dynamic probability of “realness” or signal-to-noise ratio in many competing first-order representations, but only update the “real” tag when evidence for an alternative outweighs my conservative hazard-rate prior – not when a representation is “stable enough” specifically to close an integration window.
Importantly, slow RM (and consciousness) updating due to conservative stability priors could mask alternative RM decision policies besides MAP. For example, RM might tag multiple representations as “likely real,” with resulting perceptual content – and update to any world model – reflecting a weighted average of those representations. Or, as with binocular rivalry, RM might rapidly sample from competing “realities” and consequently rapidly update internal world models. In fact, these non-MAP policies appear in constructing first-order estimates (e.g., multisensory causal inference; (Rohe & Noppeney, Reference Rohe and Noppeney2015; Wozny, Beierholm, & Shams, Reference Wozny, Beierholm and Shams2010)), making them plausible for RM, too.
Why might RM use one of these alternative decision policies, and not stick with MAP? As mentioned above, MAP is only optimal under symmetric error costs (Berger, Reference Berger2014; Bishop, Reference Bishop2006), meaning the main challenge is choosing the one correct first-order candidate as “real now.” But real organisms rarely live in tidy, symmetric cost conditions. Sometimes tagging an only-slightly-wrong first-order representation as “real” might be much less costly than hesitating; other times, a false positive could be catastrophic. Further, empirical evidence links faulty RM to difficulty distinguishing uninformative noise from environmental change (Katthagen et al., Reference Katthagen, Fromm, Wieland and Schlagenhauf2022), suggesting RM abilities may be related to ability to separately monitor world versus representational variability. Under uneven costs and imperfect volatility source monitoring, the brain may be incentivized to keep more than one option for “reality” actively in play, especially when any “real now” tag is not 100% certain. And with strong stability priors, RM could be using any of these decision policies internally – MAP, sampling, or averaging – and still we would experience a single, slowly updating conscious percept.
If not to support model-based planning under longer sensory horizons, what else would RM-based consciousness be good for? Perhaps the answer lies in characterizing RM’s decision policy. For example, MAP-like systems learn quickly but can also be rigid and prone to confirmation bias (Lange et al., Reference Lange, Chattoraj, Beck, Yates and Haefner2021); systems that sample or keep multiple options in play learn more flexibly and smoothly shift beliefs, but at the cost of speed. Thus, RM’s decision policy carries important implications for its function and interaction with other cognitive capacities, and for understanding the evolutionary pressures that shaped its emergence.
You have
Access
Fleming and Michel (F&M) suggest that conscious vision reflects the output of a reality monitoring (RM) mechanism that evolved to separate model-based planning from a “real now” percept used to update world models. They propose that this lens explains why conscious vision feels unitary and why it is relatively slow to appear and update. Here, I constructively challenge F&M’s evolutionary argument by exploring how slowness and unity might arise without long integration windows or selecting a single “reality” for planning.
To begin, let’s consider how RM might select or tag a given first-order representation as “real now.” F&M suggest this occurs when sensory integration has resulted in a “reliable-enough” first-order estimate, which is tagged as the single reality. F&M do not use Bayesian language, but their description implies a RM decision policy (Peters, Reference Peters2022; Peters & Asrari, Reference Peters and Asrari2025) which takes as “real” the only (or first) first-order representation to arrive at this reliability threshold, that is, a winner-take-all, maximum a posteriori (MAP) policy: F&M say, “subjects presumably feel like they are experiencing a single image at a time,” “reflecting the outside world as it is now.” So one representation is tagged as “real,” and updating is slow because integration to achieve stability takes time.
If this explanation is correct, evolution likely optimized the RM decision policy. A MAP policy would optimize for minimizing error in selecting which first-order representation is “real” and then updating this tag when it later determines a subsequent first-order representation has become “reliable enough.” This not only implies that the most important function of RM is to periodically identify the one “real” first-order representation, but also that the cost of getting this wrong is symmetric (Berger, Reference Berger2014; Bishop, Reference Bishop2006): tagging a hallucination as real would have the same consequences as tagging a percept as hallucination.
But could RM’s slowness reflect something other than integration to support model-based planning? If so, those alternative sources of slowness may also mask alternative decision policies, which themselves could reveal how evolutionary pressures shaped RM’s emergence, function, and link to consciousness.
To explore these possibilities, let’s start by introducing “stability priors,” formalized in decision neuroscience as hazard-rate priors about environmental rate of change: the current state estimate partially depends on the previously inferred state, modulated by an estimate of environmental volatility, h (Behrens et al., Reference Behrens, Woolrich, Walton and Rushworth2007; Gold & Shadlen, Reference Gold and Shadlen2007; Nassar et al., Reference Nassar, Wilson, Heasly and Gold2010). If h is small (conservative), the world is assumed stable and interpretations are “sticky”; if h is large, the model more readily switches interpretations. In perception, conservative hazard rates can explain the relatively slow rate of binocular rivalry switching: “under the hood” the brain is sampling perceptual evidence and computing posteriors over world states quickly, but the system resists updating its world-state estimate unless the balance of evidence shifts by “enough” (Gershman, Vul, & Tenenbaum, Reference Gershman, Vul and Tenenbaum2012).
Hazard-rate priors usually describe expectations about environmental stability, but RM monitors the stability of first-order representations. So the RM system must also determine the source of any first-order variance: is the world changing, or just the competing first-order representations? Is variance due to meaningful stochasticity, or is it just noise? This need for variance source monitoring also implies two hazard-rate priors: expected environmental stability and expected first-order representational stability.
Either of these hazard-rate priors could make RM-based updating of consciousness slow. If I expect slow environmental change, I may compute world-state posteriors faster than the 300-500 ms evidence integration loop suggested by F&M, but will not update perceptual content very often. Likewise, if I expect “reality”-tagging to change slowly, I may rapidly compute a dynamic probability of “realness” or signal-to-noise ratio in many competing first-order representations, but only update the “real” tag when evidence for an alternative outweighs my conservative hazard-rate prior – not when a representation is “stable enough” specifically to close an integration window.
Importantly, slow RM (and consciousness) updating due to conservative stability priors could mask alternative RM decision policies besides MAP. For example, RM might tag multiple representations as “likely real,” with resulting perceptual content – and update to any world model – reflecting a weighted average of those representations. Or, as with binocular rivalry, RM might rapidly sample from competing “realities” and consequently rapidly update internal world models. In fact, these non-MAP policies appear in constructing first-order estimates (e.g., multisensory causal inference; (Rohe & Noppeney, Reference Rohe and Noppeney2015; Wozny, Beierholm, & Shams, Reference Wozny, Beierholm and Shams2010)), making them plausible for RM, too.
Why might RM use one of these alternative decision policies, and not stick with MAP? As mentioned above, MAP is only optimal under symmetric error costs (Berger, Reference Berger2014; Bishop, Reference Bishop2006), meaning the main challenge is choosing the one correct first-order candidate as “real now.” But real organisms rarely live in tidy, symmetric cost conditions. Sometimes tagging an only-slightly-wrong first-order representation as “real” might be much less costly than hesitating; other times, a false positive could be catastrophic. Further, empirical evidence links faulty RM to difficulty distinguishing uninformative noise from environmental change (Katthagen et al., Reference Katthagen, Fromm, Wieland and Schlagenhauf2022), suggesting RM abilities may be related to ability to separately monitor world versus representational variability. Under uneven costs and imperfect volatility source monitoring, the brain may be incentivized to keep more than one option for “reality” actively in play, especially when any “real now” tag is not 100% certain. And with strong stability priors, RM could be using any of these decision policies internally – MAP, sampling, or averaging – and still we would experience a single, slowly updating conscious percept.
If not to support model-based planning under longer sensory horizons, what else would RM-based consciousness be good for? Perhaps the answer lies in characterizing RM’s decision policy. For example, MAP-like systems learn quickly but can also be rigid and prone to confirmation bias (Lange et al., Reference Lange, Chattoraj, Beck, Yates and Haefner2021); systems that sample or keep multiple options in play learn more flexibly and smoothly shift beliefs, but at the cost of speed. Thus, RM’s decision policy carries important implications for its function and interaction with other cognitive capacities, and for understanding the evolutionary pressures that shaped its emergence.
Acknowledgments
I thank Rachel Denison for helpful comments.
Financial support
This work was supported in part by a fellowship from the Canadian Institute for Advanced Research Program in Brain, Mind, & Consciousness.
Competing interests
I declare none.