R2. Multisensory horizons

Our paper focused on conscious vision. But what about other modalities, such as audition, olfaction, or proprioception? In this section, we respond to the challenge of extending our account to incorporate multisensory horizons. In our target article, we already envisaged generalizing our argument to “encompass world models in general, regardless of their sensory origins” and hypothesised that “a capacity for model-based planning would scale with the organism’s (multimodal) sensory horizon” (Fleming & Michel, p. 13). We note that much of the discussion here is conditional on future empirical study of the (aquatic) sensory horizon in these modalities and how it may or may not support model-based planning. But in general, we are doubtful these modalities in isolation can support the rich spatial and object-based precision afforded by (terrestrial) vision. Vision offers an enormous advantage for model-based planning compared to other senses because it provides (almost) immediate information about what is where.

Cheng & Tseng, Herzog & Doerig, Cleeremans & Dolega, Browning & Veit, and Snyder all noted that sound travels well in water, potentially offering a large acoustic horizon. Aquatic species can detect prey and predators using sounds, and some can produce acoustic signals when encountering danger or during courtship (Ladich, Reference Ladich2019). However, it is difficult to use auditory cues for model-based planning. Let us present three reasons why.

For starters, although it is true that low-frequency sounds travel very far underwater, their long wavelengths provide poor directional cues. For this reason, low-frequency sounds might be useful as an early-warning system but might not provide the spatial precision required for complex planning. For instance, Cleeremans & Dolega note that sharks can detect prey from hundreds of meters using sound (Myrberg, Reference Myrberg2001). This shows that sounds can lead to approach and avoidance behaviors, but not that sharks are able to plan based on sounds. Future research might explore a frequency trade-off in this respect: low-frequency sounds travel far but might only provide poor spatial resolution, while high-frequency sounds (e.g., the clicks used by dolphins for echolocation) have better spatial resolution but are absorbed faster. A second point is that, unlike on the surface where echoes disappear quickly, underwater sound waves bounce repeatedly off reflective surfaces and overlap. These reverberations and distortions make it difficult to determine whether a sound comes from a single event or multiple reflections, and where exactly the sound originates. This significantly complicates the use of sounds for model-based planning underwater. Finally, while vision provides stable landmarks for planning, auditory cues are dynamic: the acoustic environment changes all the time. This is the case because objects and animals are moving, because of water currents, and again because of reverberations. Reliable spatial reference points are important for model-based planning, and the information that sounds provide underwater may be too ambiguous for (spatial) reasoning and planning.

Birch considers the case of olfaction and suggests a promising hypothesis: olfaction naturally comes in discrete samples associated with sniffs. If olfaction came first, one can imagine that vision recruited the integrative machinery of olfaction, thus leading to discrete episodes of conscious vision. This hypothesis is not directly in tension with our view, so long as vision remains the first sense that was recruited for model-based planning after the water-to-land transition, rather than olfaction itself being the driver of model-based planning. The diffusion of odorants is slow, especially underwater – approximately 10,000 times slower underwater compared to air (Hoover, Reference Hoover2010, p. 65) – and generally does not provide the excellent spatial resolution of vision. This is not a good source of information for model-based planning. A model of the world relevant for planning in the here-and-now should represent distal features of the environment in a way that is dynamically coupled to changes in the environment. Odors underwater do not provide the relevant kind of information and do not provide it fast enough. This is not a problem when it comes to using olfaction for other purposes, such as moving toward or away from an odor source, which aquatic animals of course do.

Birch argues that an olfaction-for-planning story is consistent with the olfactory bulb projecting directly to the forebrain without the need for a thalamic relay. However, the opposite case can also be made, in light of the well-established role of thalamo-cortical gating in planning and consciousness in the human brain. In fact, Coppin uses the same neuroanatomical datapoint to mount a contrary argument, that bypassing the thalamus makes olfactory information more likely to be processed implicitly. Coppin argues that the water-to-land transition impacted vision more than other sensory modalities, which is potentially consistent with many aspects of olfaction remaining more implicit in humans. A proper test of these suggestions is to explore how sensory horizons operate in other modalities and to triangulate these investigations against the functional signatures of modality-specific planning and reality monitoring.

One might be skeptical that our hypothesis applies to modalities such as touch and interoception for which the notion of a sensory horizon does not straightforwardly apply. But Shepherd shows how our model can be generalized to conscious proprioception. In a nutshell, Shepherd suggests that, just as we create a model of the external world for planning, we also build a model of our own body, which might take time – as Shepherd suggests is indicated by the cutaneous rabbit illusion (Eimer, Forster, & Vibell, Reference Eimer, Forster and Vibell2005). One possibility is that other proximal modalities such as proprio- and interoception become part of a world model once a larger (visual) sensory horizon is available to drive the benefit of model-based planning. This view predicts that the conscious sensation of movement might play a role for planning, but not for fast error corrections which can occur unconsciously, for instance, following a subtle change in the position of a target (Fourneret & Jeannerod, Reference Fourneret and Jeannerod1998).

Finally, Cheng & Tseng suggest that multisensory integration indicates consciousness. As they note, in fish, this integration occurs for the most part at a low level, in the tectum (superior colliculus). This structure is responsible for quickly translating sensory inputs into appropriate motor programs such as escape responses or pursuit (Baier & Scott, Reference Baier and Scott2024). This is the kind of structure our account would consider unconscious par excellence. As noted by Isa et al. (Reference Isa, Marquez-Legorreta, Grillner and Scott2021), and as Goodale & Corneil point out, the retino-colliculo-thalamic pathway is remarkably well preserved across vertebrates, which can allow us to test whether it can operate unconsciously. As it turns out, this pathway is responsible for (at least some) unconscious behavior in blindsight (Isa et al., Reference Isa, Marquez-Legorreta, Grillner and Scott2021). One intriguing possibility is that this structure might also mediate other forms of unconscious multisensory integration in humans which have been documented in experiments (Mudrik, Faivre, & Koch, Reference Mudrik, Faivre and Koch2014; Scott et al., Reference Scott, Samaha, Chrisley and Dienes2018).

R3. Model-based planning

A number of commentators questioned the scope and functionality of the kind of model-based planning that we propose is enabled by conscious vision. A first clarificatory point: the kind of model-based “planning” we have in mind is rapid and unconscious, occurring within ∼400 ms windows of perceptual integration. Krakauer & Quilty-Dunn ask us to distinguish between this leaner notion of model-based planning and the explicit, deliberative sort of planning that may be unique to humans. We agree with Krakauer & Quilty-Dunn that explicit deliberation is rare in the animal kingdom and goes beyond the functionality supported by conscious vision in a majority of species. Instead, we propose that visual consciousness supports the sort of non-habitual, “world sensitive” behavioral control exhibited by Tolman’s classic experiments on animal learning – what Krakauer & Quilty-Dunn refer to as “intelligent unconscious processes.” Much of our disagreement is semantic. In the literature on reinforcement learning, these processes are labeled as “model-based,” to distinguish them from more inflexible Pavlovian or habitual forms of behavior. Similarly, by “simulations,” we do not mean deliberative, conscious simulations, or the kind of heavy-duty counterfactual thinking that emerges later on in childhood (see also Suddendorf & Renshaw). Our minimal commitments only require that conscious percepts are being used for fast, intelligent inferences about possible future behavior. We propose that these inferences are themselves unconscious, and that the distinction between goal-directed (model-based) and habitual (model-free) control helps provide an algorithmic framework for understanding their differences.

Jablonka & Englander suggest that unlimited associative learning (UAL) is model-based – or at least invokes a “model-dependent” capacity involving mappings of the world and body. Again, we suspect our disagreement is semantic. Model-free learning in reinforcement learning can avail itself of rich, map-like representations of the body and the world but does not involve an internal search through these maps to support planning. We suspect the capacity Jablonka & Englander refer to in fact lines up more with associative (model-free, in reinforcement learning terminology) learning in a large (body and world-representing) state space, rather than model-based cognition of the sort which requires simulating and planning within an internal model of body and world. This latter capacity is what we refer to as model-based cognition, and as Jablonka & Englander point out, we associate this capacity with the requirement for reality monitoring (to distinguish between simulations and reality), and consciousness. Similarly, Shea argues there may be forms of perception and selection that admit of relatively sophisticated object tracking, but that might not require simulation or step-by-step planning. We agree, and to the extent that such a selection process does not require reality monitoring, these would not lead to conscious perception under our account.

MacIver & Dombeck recount recent experiments identifying what they refer to as “non-local activity” in the hippocampus – signals that indicate the brain is representing locations distal to the animal’s current position, and which appear to track trajectories which the mouse could take away from their current location, firing at 10–20× the speed of actual movement. Whether or not we should interpret these signals as internal “simulations” is open to question, but they could plausibly form a neural substrate for percept-guided “unconscious intelligent behaviour” of the sort advocated for by Krakauer & Quilty-Dunn. By inactivating these neural signatures, and observing the consequences for behavior, one could develop an empirical test of this hypothesis. Our prediction is that online action control would remain unaffected, but that planning and the efficiency of escape behavior would be disrupted.

Following this discussion and refinement of terminological differences, one question remains for us: why would inferences require simulation? Why not simply draw inferences without simulating? An intriguing possibility is that a significant amount of knowledge an animal has is not explicitly stored but is instead implicitly stored in the dynamics governing transitions between states of the system. Unconscious simulation could be a way of effectively using implicitly stored knowledge for the purposes of planning.

MacIver & Dombeck also note an intriguing corollary of our view: a drawback of the architecture we propose is that deliberative, conscious imagination that takes us out of the present moment to engage in future-oriented thinking beyond our immediate environment may be unusually difficult – requiring cultural scaffolding to instill in children (see Suddendorf & Renshaw), and perhaps even technological aids to guide effective planning over longer time horizons in adults.

Finally, Kentridge presents intriguing evidence indicating that jumping spiders (Portia) can program a series of moves within a mental map of their environment. Our argument regarding the water-to-land transition argument rests on the advantages afforded by a terrestrial ecological niche for visually guided planning. We cannot assume the same mechanisms, given differences in anatomy, but we see no reason those advantages cannot be enjoyed by arthropods. We note however that these demonstrations may not (yet) indicate that jumping spiders can perform flexible inferences using an internal model, namely the ability to re-program a series of moves on the fly. As an analogy, think of navigating to a given location in a new city. Suppose that for some reason you cannot take a map with you. One method is to look at a map before leaving and memorize a series of left/right moves (and the names of the relevant streets). This method would work as long as nothing along the way requires coming up with a new route. We could call this strategy “pre-planning.” Pre-planning is different from online model-based inference – the capacity to come up with a new route on the fly. Further experiments could determine whether the kind of planning observed in Portia is genuine model-based planning or more akin to pre-planning.

R4. Aquatic ecology and working memory

Several commentators noted that aquatic animals might be able to use vision-for-thinking even with short sensory horizons because not all vision-based behavior toward objects at close range impose high behavioral urgency. Accordingly, those commentators appeal to evidence of cognitively sophisticated behavior in aquatic animals (most notably in fish) which might require capacities to maintain and manipulate sensory information to reach goals. Here, we consider whether our proposal underestimates those cognitive capacities and whether those capacities are good indicators of consciousness.

We noted that the right amount of environmental clutter might be sufficient for the development of model-based capacities, even underwater (Mugan & MacIver, Reference Mugan and MacIver2020), which could explain the capacities of cephalopods. Snyder points out that model-based mechanisms might develop for “avoidance of dangers and the seeking of foods that are not continuously detectable with the senses,” which he argues is consistent with the observation that fish are capable of amodal completion and tracking objects behind occluders. However, as Kanizsa (Reference Kanizsa1985) showed, amodal completion is a low-level phenomenon that does not depend on cognition. There is some evidence of unconscious amodal completion in humans (Emmanouil & Ro, Reference Emmanouil and Ro2014). Simple trajectory extrapolation behind an occluder does not require any kind of reasoning about the occluded objects either. As noted by Suddendorf & Redshaw, better evidence would come from invisible displacement tasks, where an animal has to track an object that is moving with its occluder, as in the shell game.

Browning & Veit, Yu & Carruthers, Jablonka & Englander, and Allen all emphasize the possibility of cognitively sophisticated behaviors in fish, even if visual objects are at close range, which do not require the kind of fast reactions to visual stimuli that we focused on. Browning & Veit suggest that finding shelter or interacting with conspecifics does not require immediate reactions and might involve planning on the basis of vision. Similarly, Yu & Carruthers show that fish are capable of optimal foraging, mate-choice copying, and can represent numerosity. Finally, Allen discusses intriguing cases of cooperative hunting between groupers and morays.

Optimal foraging can be achieved in a model-free way via the marginal value theorem (Charnov, Reference Charnov1976; Constantino & Daw, Reference Constantino and Daw2015). All that is required is to leave a patch when the rate of return decreases past a certain threshold. In fact, some forms of efficient foraging, such as Lévy walk (short moves occasionally followed by long relocations), do not even require input from the head ganglia in drosophila (Sims et al., Reference Sims, Humphries, Hu, Medan and Berni2019). So, efficient (albeit inflexible) foraging does not necessarily require cognition on the basis of vision, and we should expect that this capacity is indeed widespread.

Yu & Carruthers suggest that model-based decision-making is required to explain mate choice and number representations in fish. Mate-choice copying is observed when an individual (typically a female) increases her likelihood of choosing a particular male after observing a conspecific court or mate with that male (Dugatkin & Godin, Reference Dugatkin and Godin1992). We find it plausible that mate-choice copying results from a model-free process. One possible reason why this mechanism has evolved is precisely to avoid the costs of having to think about which mate to select. A variety of mechanisms have been proposed to explain mate-choice copying (Vakirtzis, Reference Vakirtzis2011; Westneat et al., Reference Westneat, Walters, McCarthy, Hatch and Hein2000), most of which do not require animals to model the mental states of other animals at all – such as their preferences. Some of those explanations are quite deflationary. For example, one possibility is “local enhancement” (Heyes, Reference Heyes1994): mating simply makes a male more likely to be noticed by a female, which then displays a preference for that male.

As for representations of numbers observed in fish (Agrillo & Bisazza, Reference Agrillo and Bisazza2018), the evidence can be accounted for by the well-known approximate number system, which is an analog, non-conceptual representation of numerosity (Clarke & Beck, Reference Clarke and Beck2021), perhaps driven by low-level cues such as density or amount of motion. However, the evidence does not indicate use of this representation for the purposes of model-based planning. Instead, these studies are evidence of a variety of model-free learning and memory mechanisms in aquatic animals (Salwiczek & Bshary, Reference Salwiczek and Bshary2011), which are consistent with our account.

Similarly, Jablonka & Englander, as well as Allen, emphasize the capacity to perform trace conditioning in fish which, according to Jablonka & Englander, indicates that fish “have working memory well exceeding 400ms.” We disagree with this interpretation (and also that trace conditioning is a good indicator of consciousness; see section “Unconscious processing in action”). It is crucial to distinguish between working memory proper and the short-term storage of sensory information. Working memory is distinguished from the latter in that it involves both maintenance and manipulation. Evidence of short-term memory in fish does not directly indicate a capacity to mentally manipulate temporarily stored information, which is required for genuine working memory (Bonin, Manrique, & Bshary, Reference Bonin, Manrique and Bshary2025). Tasks that supposedly indicate working memory in fish do not distinguish it from short-term memory. For example, zebrafish can perform delayed match-to-sample tasks, but their performance falls to chance for delays higher than five seconds (which might indicate no capacity to rehearse) and can be explained by familiarity effects, independently of working memory (Bloch et al., Reference Bloch, Froc, Pontiggia and Yamamoto2019). Of note, a recent study found that even the cognitively sophisticated cleaner wrasse fails a series of working memory tasks that not only require maintenance of sensory information but also manipulation (Bonin et al., Reference Bonin, Manrique and Bshary2025). Further cause for doubt with respect to working memory capacities is the fact that Aellen, Burkart, and Bshary, (Reference Aellen, Burkart and Bshary2022) failed to find correlations in performance between different cognitive tasks (a g factor, which itself correlates with working memory) in cleaner wrasse (see also Ferreira, Szabo, and Taborsky, (Reference Ferreira, Szabo and Taborsky2025) for a convergent result). These results led Triki, Van Schaik, and Bshary, (Reference Triki, Van Schaik and Bshary2025) to suggest that fish have a “modular organization of cognitive abilities” (p. 4) (though we should be careful when interpreting what are essentially null results, and, as Allen noted, generalizing beyond specific species of fish is difficult).

Like Browning & Veit and Yu & Carruthers, Allen holds that fish can “detect … and carefully monitor” prey from a distance. This could be another case of vision that does not require fast reactions. But this behavior is possible in the absence of any significant cognition on the basis of vision. There are more promising cases. As we noted, other influences – such as environmental clutter – may have created pressures for the evolution of planning, particularly in the kind of cooperative hunting that Allen reviews. However, there is cause for doubt even in this case. Bshary et al. (Reference Bshary, Hohner, Ait-el-Djoudi and Fricke2006), who studied coordinated hunting between groupers and giant moray eels, note that “associating partners behave in exactly the same manner as they do when hunting alone” and that “interspecific coordinated hunting of this kind can emerge … from a simple associative learning process, whereby each species associates increased rewards with hunting in the vicinity of the other species” (Bshary et al., Reference Bshary, Hohner, Ait-el-Djoudi and Fricke2006, p. 2396).

Finally, Jablonka & Englander argue that our analysis underestimates how far aquatic animals can “discern” ecologically relevant objects, partly based on the Nilsson–Warrant–Johnsen equations for the optical properties of water (Nilsson, Warrant, & Johnsen, Reference Nilsson, Warrant and Johnsen2014). The Nilsson–Warrant–Johnsen framework (Nilsson et al., Reference Nilsson, Warrant and Johnsen2014), however, is not concerned with discernment of the kind relevant for cognition on the basis of vision, which would minimally involve categorization of the target. Instead, it estimates an upper bound for the detection of a black stationary target against a uniform background (Nilsson et al. Reference Nilsson, Warrant and Johnsen2014). In particular, Nilsson et al. (Reference Nilsson, Warrant and Johnsen2014) obtain their detection ranges with an explicit spatial pooling assumption: all photons emitted by the target are modeled as being pulled into one channel, the activity of which is then compared against that of a channel for the background. This idealized assumption significantly increases detectability and is useful for the purpose of modeling the best-case-scenario detection distance. However, this also means that this cannot be interpreted as a model of discernment or discriminability: the spatial pooling assumption increases detectability at the cost of even the most basic spatial details, including motion. Cognition on the basis of vision requires much more than the simple capacity to detect something (as opposed to nothing). It (minimally) requires the capacity to categorize visual stimuli, which for aquatic animals only becomes available at closer range.

R5. Reality monitoring

A related set of commentaries raise questions about the scope and generality of reality monitoring within our framework. Another point of clarification: in line with perceptual reality monitoring (PRM) theory (Lau, Reference Lau2019, Reference Lau2022), the reality monitoring system we propose as relevant for consciousness operates implicitly, functioning to unconsciously route sensory representations for direct impact on beliefs. As noted by Cleeremans & Dolega, this is in line with our view being one version of higher-order theories of consciousness.

Suddendorf & Redshaw note that explicit capacities to distinguish pretense and reality, as well as the capacity of “representing representational relations themselves,” develop relatively late in childhood. We agree that it seems implausible that consciousness depends on these explicit meta-representational capacities given how late they emerge. Instead, we expect that implicit reality monitoring is in place earlier in life, and does not require possessing a concept of representation, or explicitly representing that some representations reliably represent reality. The contrast between dreaming without insight and lucid dreaming illustrates the distinction. During dreaming, the reality monitoring system mistakenly selects representations as reliable representations of the world. This is compatible with lucid dreaming, where one explicitly recognizes that one’s experiences do not reflect reality – which presumably requires concepts such as “reality” and “dream.” Lucid dreaming is a case where implicit reality monitoring and explicit reality monitoring come apart – and only the former is relevant for consciousness in our account. So, while explicit reality monitoring indeed emerges fairly late in development and is pretty rare among animals, we expect that implicit reality monitoring capacities are widely shared among animals with the model-based planning capacities that emerge with large sensory horizons. Similarly, Shepherd points out that humans also have the ability to explicitly and intentionally engage in mental imagery. This mental imagery may be tagged by implicit mechanisms with a degree of reality (and hence experienced), but explained away at a higher cognitive level as imagined based on the background knowledge that it is self-generated (Dijkstra, Kok, & Fleming, Reference Dijkstra, Kok and Fleming2022).

Balsdon & Jovanovic and Peters both explore the mechanistic basis of reality monitoring and how this might relate to temporal windows of integration in conscious perception. We agree with Balsdon & Jovanovic that the window of unconscious sensory integration could be modulated by current confidence levels in order to flexibly solve Hamlet’s problem – but such mechanisms would themselves be unconscious and rely on implicit forms of metacognition. Rather than “conscious vision [having] some use in solving Hamlet’s problem,” the need to engage in reality monitoring, and solve Hamlet’s problem, provides a functional basis for conscious vision. Work by Balsdon and colleagues shows that this online and continuous implicit tracking of uncertainty may be foundational to the control of windows of integration in vision (Balsdon, Wyart, & Mamassian, Reference Balsdon, Wyart and Mamassian2020; Balsdon & Philiastides, Reference Balsdon and Philiastides2024).

Peters articulates different mechanistic influences on reality monitoring “decision policies” – the set of criteria that are used to select or tag a given first-order representation as real in the service of planning and decision-making. Peters points out that the mechanistic basis and operation of reality monitoring was left as an open empirical question in our target article. She helpfully articulates a number of influences that a well-functioning reality monitoring system should incorporate, leading to novel and testable empirical hypotheses. Beyond tracking the signal-to-noise statistics of perceptual representations, such a system may track priors over the hazard rate of the environment – making windows of integration more labile in faster-changing worlds. It would be possible to investigate this by combining experimental manipulations of environmental volatility (Behrens et al., Reference Behrens, Woolrich, Walton and Rushworth2007; Nassar et al., Reference Nassar, Wilson, Heasly and Gold2010) with postdictive psychophysics.

Peters also asks whether asymmetric cost functions (such as a particularly high cost for failing to detect a predator) may incentivise the system to maintain multiple candidate first-order representations, even if this creates reality monitoring errors. This discussion relates to questions posed by de Jong about the computational benefits of selecting a single model of reality to plan over. While we agree with Peters that asymmetries in cost functions may be normatively expected to keep multiple percepts in play, empirical findings suggest that percepts are in fact fixed before a cost function is applied. Fleming, Maloney, and Daw, (Reference Fleming, Maloney and Daw2013) had people make rapid perceptual discriminations under asymmetric payoffs and found that the best (and, in this case, anti-Bayesian) account of the data was one in which a single perceptual interpretation is fixed prior to being combined with asymmetric payoffs. In some cases, this can lead to catastrophic results of the sort articulated by Peters – as when aircraft pilots, caught out by vestibular illusions, do not make allowance for even a small probability of misperception, leading them to “correct” the plane’s flight and ultimately causing a crash (Fleming et al., Reference Fleming, Maloney and Daw2013). The existence of such irrationalities in categorical perception is consistent with broader computational constraints of the sort we articulate in the target article that posit dynamic loops of (discrete) perception followed by planning.

Finally, Niknam and Hossein & Roknabadi suggest our analysis could shed light on the origins of particular psychiatric symptoms such as hallucinations or depression. We agree that a successful functional explanation of consciousness has the potential to shed new light on disturbances of conscious experience suffered in disorders of mental health. We disagree though that expanding sensory horizons should necessarily guarantee reliable reality monitoring under our account. Indeed, the opposite is likely to be true – under short sensory horizons, organisms are more directly coupled to their environments, and reality monitoring is not needed. Indeed, much of what Niknam et al. characterise as disturbances in a “fragile and fallible” capacity for reality monitoring may be a consequence of the development of richer model-based cognition and the increasing need to use the brain in a generative, offline capacity. But we can also think of particular aspects of psychiatric conditions as being due to failures of reality monitoring – for instance, one reason hallucinations in schizophrenia are so distressing is that they are not typically accompanied by metacognitive insight and are instead automatically tagged as reflecting reality. As noted in our target article, unconscious visual processing does not seem particularly affected in schizophrenia, consistent with a specific deficit in reality monitoring (Berkovitch, Dehaene, & Gaillard, Reference Berkovitch, Dehaene and Gaillard2017; Grandgenevre et al., Reference Grandgenevre, Vaiva, Boloix, Bubrovszky, Schwan and Laprevote2015). One intriguing possibility raised by our hypothesis is that deficits in reality monitoring and time processing observed in schizophrenia are two sides of the same coin (Amadeo et al., Reference Amadeo, Esposito, Escelsior, Campus, Inuggi, Pereira Da Silva, Serafini, Amore and Gori2022). Studying how insight works in the context of conscious perception may shed light on the mechanisms involved in such disorders (Dijkstra, Mazor, & Fleming, Reference Dijkstra, Mazor and Fleming2024; Mihali et al., Reference Mihali, Broeker, Ragalmuto and Horga2023).

R6. Unconscious processing in action

Some commentators questioned the role of unconscious processing in action, or unconscious processing in general.

Shanks is skeptical of the evidence of unconscious processes cited in our article. No one denies (or at least should deny) that unconscious processing exists. If that weren’t the case, conscious vision would start at the retina. There are two questions from there. First, when does conscious perception start? Second, which functions can be performed unconsciously? We suggested that one’s response to the former question should inform one’s response to the latter. Shanks does not address the view that conscious perception occurs late. But if conscious perception is late, we argued, there must be a whole class of actions that can be driven by unconscious vision instead. This argument is left untouched.

Of course, it would be better to also have independent evidence supporting our answer to the second question – evidence that does not depend on the argument from the speed of conscious vision. Is the evidence we cited for unconscious functions as unreliable as Shanks makes it out to be? We do not think so. There is still ample evidence for a dissociation between conscious vision and vision for action (see Kozuch (Reference Kozuch2023) for a response to some of the concerns), that blindsight is indeed best understood as reflecting unconscious vision (Michel & Lau, Reference Michel and Lau2021), for unconscious priming effects (most notably using the dissociation paradigm; Schmidt & Biafora, Reference Schmidt and Biafora2024), and that various forms of conditioning (especially Pavlovian conditioning) can indeed occur unconsciously, for instance in spinalized rats (Liu et al., Reference Liu, Ferguson, Crown, Bopp, Miranda and Grau2005).

Like Shanks, Allen is skeptical that trace conditioning can occur unconsciously, noting that the study of trace conditioning in decerebrate guinea pigs we mentioned was only achieved with a 250 ms delay (though there was still trace conditioning even with a 500 ms delay in this study; Kotani, Kawahara, & Kirino, Reference Kotani, Kawahara and Kirino2003, Figure 4). But it seems possible to achieve trace conditioning in anesthetized rats with a 3-second delay (De Oliveira, Dickson, & Reyes, Reference De Oliveira, Dickson and Reyes2020), in addition to other forms of learning (see Samuel et al. (Reference Samuel, Taub, Paz and Raz2018) for a review), as well as during non-REM sleep in humans (Arzi et al., Reference Arzi, Shedlesky, Ben-Shaul, Nasser, Oksenberg, Hairston and Sobel2012).

What about recent failures to identify unconscious conditioning (Skora, Scott, & Jocham, Reference Skora, Yeomans, Crombag and Scott2021, Reference Skora, Scott and Jocham2024)? Why are unconscious effects so hard to demonstrate? Here is why: the main methods used to suppress consciousness (such as visual masking by noise) also suppress unconscious processing to a large extent (Lau, Reference Lau, Weiskrantz and Davies2008). This confound is partly avoided in blindsight monkeys, which show instrumental conditioning for cues in the blind field (Kato et al., Reference Kato, Zeghbib, Redgrave and Isa2021). Of course, every study has flaws and there are always possible alternative explanations. Yet, when we consider all the evidence together, including evidence on the speed of conscious vision, it points toward a significant role for unconscious mental processes.

It is important to note that our view is not that vision-based actions are entirely carried out by the dorsal visual pathway. As noted above in our response to Cheng & Tseng, subcortical areas might be responsible for triggering some fast vision-based reactions quite independently of the cortex. In particular, Goodale & Corneil point out that express visuomotor responses occur “far faster than any response generated by the cerebral cortex, including by the dorsal stream,” which suggests that they emanate from the superior colliculus. As they note, it is possible to pit these very fast subcortical and slower cortical responses against one another in anti-pointing experiments. The result of those experiments indicates that the first subcortical-dominated phase of movement is still directed toward the stimulus despite explicit instructions (Gu et al., Reference Gu, Wood, Gribble and Corneil2016), consistent with a role in habitual rather than goal-directed behavior (Hardwick et al., Reference Hardwick, Forrence, Krakauer and Haith2019).

Mann & Wu suggest that reportability, which we take to be an indicator of consciousness, is just one kind of intentional action and should not rule out other intentional actions as indicators of consciousness. Since some of those intentional actions are fast, this might indicate fast consciousness. In making that argument, Mann & Wu present interesting data on the properties of skilled performance in elite cricketers. Strikingly, batters’ hitting performance is resilient to substantial amounts of blur that affect conscious vision, with one study revealing that blur can be applied up until the point of legal blindness before performance starts to be affected (Mann et al., Reference Mann, Ho, De Souza, Watson and Taylor2007). This result is presumably possible because dorsal stream and subcortical processing exploits low spatial frequencies and motion energy. If our account is correct, fast responses of this kind are driven by unconscious vision. Here is a point in our favor: it is notable that the batters are “typically terrified” when this blur is introduced. A possible explanation in line with our view is that their conscious visual model of the world does not seem to be one that could support skilled performance, and yet unconscious, online (and still visually guided) action control can perform wonders. A significant advantage of considering that this performance is driven by unconscious vision is that it explains why hitters are surprised by their own skilled performance, given that unconscious vision is not available for explicit metacognition (Michel, Reference Michel2023).

Finally, Roskies considers the implications of our model for autonomous agency and free will. In particular, Roskies cautions against understanding our view along the lines of the (mis)interpretations of Libet’s results which have “negatively influenced discussions of free will for the past forty years.” Essentially, the misinterpretation of our view would be that our actions cannot express free will, since consciousness lags behind action. As Roskies notes, however, “lags in consciousness of external stimuli are consistent with our ongoing conscious involvement with the world in ways that preserve most intuitive conceptions of agency, free will, and responsibility.” We entirely agree. We take unconscious action-guidance to be compatible with free will and responsibility, so long as agents can rationally identify with those actions (Frankfurt, Reference Frankfurt1971). Free will is not primarily manifested in the fine details of online action-guidance, reflex reactions to stimuli, or the inconsequential wrist movements studied by Libet (Reference Libet1985), but in the kind of planning and deliberation that we argue arose following the water-to-land transition. Free will would be threatened by the existence of unconscious processes only if those processes somehow led agents to act against principles that they would otherwise rationally endorse (Nahmias, Reference Nahmias and Sinnott-Armstrong2014). But there is no reason to think that this is the case. Going back to Mann & Wu’s example, a cricketer deserves credit for hitting a boundary even if they did so based on unconscious vision, for it is their conscious planning and investment in hours of training that set them up for success.

R7. The spirit of functionalism and the finish line of consciousness

Several commentators (Cleeremans & Dolega, Herzog & Doerig, Key) raised questions about our claim that consciousness has the function of reality monitoring and about our commitment to the idea of a “finish line” of consciousness (Phillips and Blackmore).

Key argues that consciousness having a function would imply that conscious experience “could open ion channels to elicit the action potentials which are needed to produce behaviour.” We believe that this objection implicitly depends on a dualist assumption. If substance dualism were true, and subjective experiences were identical to states of a non-physical substance, it would indeed be puzzling how subjective experiences could cause anything. But if subjective experiences are identical to neural states realizing certain functions, the puzzle is dissolved. Beyond this, we are sympathetic to Key’s view that consciousness has a function in making sense of ourselves: part of the role of consciousness might be to build a model of ourselves and our own behavior.

Herzog & Doerig present what we could call a “cheap implementation” argument against our view. Cheap implementation arguments go like this: theory T says that conditions A, B, and C are jointly sufficient for consciousness. But there is a cheap implementation of A, B, and C that is obviously not conscious (e.g., a simple neural network). Therefore, T is wrong. One could indeed design a fairly simple neural network performing something akin to reality monitoring (see also Cleeremans & Dolega). As Herzog & Doerig note, similar cheap implementation arguments can be formulated against all functionalist theories of consciousness. So, are all functionalist theories of consciousness thereby proven wrong?

There is a fundamental flaw in cheap implementation arguments. Shoemaker (Reference Shoemaker1981) distinguished between the total realizer of a function and the core realizer of a function. In a car, the car brake has the function of slowing the car down. The core realizer of the function is the brake mechanism that produces deceleration. The total realizer of the function is the brake mechanism within the whole car. One could imagine someone missing the distinction between total and core realizers, and making the mistake of believing that the brake mechanism is not responsible for deceleration, as evidenced by the fact that it doesn’t slow anything down once studied in isolation. The cheap implementation argument essentially makes the same mistake. The brake mechanism – the core realizer of the function – explains how the car slows down, but doesn’t properly fulfill its function outside the car, namely, without the total realizer. Similarly, in our account, the reality monitor is the core realizer of consciousness: it is what explains the difference between conscious and unconscious states within a subject with mental states such as beliefs, perceptions, and desires. One can have a cheap implementation of a reality monitoring system that doesn’t give rise to consciousness, just as one can have an isolated car brake mechanism that doesn’t slow anything down. But this simple fact doesn’t imply that the reality monitor is not the core realizer of consciousness. We agree with Herzog & Doerig that providing a description of the total realizer of consciousness will require inserting much more details, most notably a theory of phenomenal qualities (e.g., Rosenthal, Reference Rosenthal2010), and we don’t have a good idea of what all the relevant conditions are yet. This, however, does not prevent the reality monitoring system from being the core realizer of consciousness.

Like Phillips, Blackmore questions a central assumption of our article, which is that of a “finish line” of consciousness. In particular, Blackmore considers that this assumption implies that “consciousness is something additional to the brain processes on which it depends.” So, is there a finish line of consciousness, and does positing a finish line of consciousness betray some non-materialist commitments on our part?

It is important to distinguish between two kinds of criticisms of the idea of a finish line of consciousness – call them the fuzzy-boundary view and the no-boundary view. The fuzzy-boundary view says that there is a vague finish line: there is no determinate moment where one becomes conscious of a stimulus. Just as one might at some point reach a stage of life where it is indeterminate whether one is bald or not, there might be a moment in time where it is indeterminate whether one is conscious of a stimulus or not. Concluding from this that there is no finish line of consciousness (i.e., the no-boundary view), however, would be committing Loki’s wager fallacyFootnote ¹ . Just because there is no determinate point where one becomes bald does not imply that no one ever transitions from not being bald to being bald. In the same way, the fact that the boundary between conscious and unconscious might be fuzzy does not imply that there is not a time before which one is determinately not conscious of a stimulus, and after which one is determinately conscious of it. Relying on the fuzzy-boundary view to reject our view amounts to committing Loki’s wager fallacy: one could still be determinately unconscious of a stimulus for 350 ms after its onset, even if one holds that there is a period of indeterminate consciousness between no consciousness and full consciousness. Our view is compatible with a fuzzy boundary of this kind, so long as visual states occurring before, say, 300 ms are determinately unconscious.

Meanwhile, the no-boundary view says that there is no fact of the matter at all about the speed of consciousness – no boundary between unconscious processing and conscious perception. This view is incompatible with our own and we consider it untenable. As Blackmore writes, we are committed to the view that before a subject consciously perceives a stimulus, their perceptual system unconsciously registers the presence of this stimulus. We believe that there is overwhelming evidence for this hypothesis (see Section “Unconscious processing in action”).

Finally, Blackmore argues that our commitment to a finish line of consciousness has the implication “that consciousness is something additional to the brain processes on which it depends.” We do not hold this view: instead, we think that consciousness is identical to a brain process that realizes a certain function (see also our response to Key). But we emphasize a crucial difference between reductionism and eliminativism: our view is reductionist in the sense that consciousness is identified with a set of functions realized in the brain. It does not eliminate consciousness, unless one starts with a concept of consciousness so inflated that it is incompatible with any kind of physical realization.

R8. The speed of consciousness: Extending the overflow gambit

Block, Philips, Turner et al., and Shea all raise skeptical challenges to our proposal that conscious vision is slow. Some of these challenges seek to add weight to the overflow gambit that we already outlined in our target article. As a brief reminder, the overflow gambit claims that postdictive phenomena are consistent with visual consciousness being fast, but also ephemeral, and being unreportable by the time the postdictive window of integration is complete.

Turner et al. note that the available evidence is still compatible with an “Orwellian” interpretation where consciousness occurs fast but is later edited. According to Hogendoorn (Reference Hogendoorn2022), “information becomes available immediately, whilst remaining editable.” (p. 4) A model like this one is indeed still on the table, if only because one can always posit conscious contents that are unavailable for introspection and later edited. One could even imagine someone arguing that retinal activity is conscious but unreportable and later edited. We suggested that absent any compelling theory-independent reason to believe that subjects are wrong about what they believe they experienced, we should side with the subjective reports. This is just applying a moderate version of the Principle of Charity which says that we should interpret what people say as rational or true except when we have compelling reasons to think otherwise. In this case, plausible reasons might include, for instance, a lack of attention, overloaded working memory capacities (Block, Reference Block2011), or response bias (Phillips, Reference Phillips2018). The cases discussed in the target article sought to rule those factors out. Beyond this basic principle, we agree with Turner et al. about the future areas of research. First, we were unaware of the result of Gwilliams et al. (Reference Gwilliams, Linzen, Poeppel and Marantz2018) showing a ∼450 ms postdictive effect of word identity on the perception of phonemes. Studies on speech perception might indeed be particularly promising for identifying long-lasting postdictive effects in the auditory domain. Second, we also agree that adding metacognitive measures offer an attractive method to probe what participants consciously experience, beyond discrimination capacity.

Block distinguishes between the formation of a conscious percept and its vulcanization (with analogy to how rubber tires are first formed, and then “set” by vulcanizing). A conscious percept formed within 200 ms might later become vulcanized as it enters working memory for the purposes of categorization and report – a transition which implies conceptualization. By the time the formed percept makes it to visual working memory, it has been vulcanized. Subjects thus report the post-vulcanization percept, not what they initially experienced as the conscious percept formed. This is also a version of the overflow gambit. We noted in the target article that for this response not to be ad hoc (and overgeneralize), we need a good reason to believe that this is indeed what is happening. Block mentions studies which suggest that event-related potentials (ERPs) associated with conscious perception emerge around 200 ms after stimulus onset. But as we discuss below in response to Siegel, Arstila & Railo, and Aranyosi, early consciousness-related ERPs could be good predictors of consciousness without corresponding to the moment consciousness itself emerges. Furthermore, the idea of vulcanization does not explain data from Scharnowski et al. (Reference Scharnowski, Hermens, Kammer, Öğmen and Herzog2007), which indicates that Vernier dominance can be influenced by TMS more than 300 ms after stimulus offset. So, while the vulcanization hypothesis is on the table, we do not believe that it currently has the evidentiary strength sufficient to overturn the subjects’ reports about what they consciously saw and what their discrimination performances indicate.

Similar to Block and Turner et al., Phillips argues that conscious perception might be malleable, with later conscious percepts erasing earlier ones. Unlike Block, however, Phillips holds that subjects might have momentary access to pre-integration conscious representations. We believe that this interpretation conflicts with two important findings. The first is that the effect of integration seems to be all-or-none. Subjects fall from 75% performance to chance performance when a vernier is presented up to ∼400 ms after the presentation of the first vernier. One could have expected that momentary access to the pre-integration representation would still have allowed subjects to discriminate the identity of the first vernier above chance. It is difficult to imagine that participants could momentarily access a feature that is task-relevant and that they’re attending to without this access somehow improving their performance.

Second, any account of the Sequential Metacontrast paradigm (SQM) has to explain why subjects cannot discriminate the orientation of the second vernier when presented up to ∼400 ms after a vernier with the opposite offset at the start of the stream (while they can discriminate its orientation if it is preceded by a neutral vernier). Clearly, the second vernier is not maintained in working memory. Is it momentarily accessed? Perhaps, but it is mysterious why this access would be momentary, given that subjects are told to pay attention to the second vernier in order to report its orientation, and given that it is only followed by neutral verniers. It is important to note that subsequent verniers do not disrupt the processing of the second vernier, as evidenced by the fact that discrimination performance of the second vernier is 75% correct when not preceded by a vernier with opposite offset 320 ms earlier (see Drissi-Daoudi, Doerig, & Herzog, Reference Drissi-Daoudi, Doerig and Herzog2019, Figure 3b).

Shea also questions the duration of long-lasting postdictive effects in conscious vision, arguing there is more secure evidence for postdictive effects in the 200–300 ms range than in the 400 ms range. It is again important to note that the duration of postdictive effects is not a direct indicator of the speed of consciousness: a postdictive effect in the 200 ms range is incompatible with consciousness before 200 ms, but it is compatible with consciousness occurring after, say, 300 ms. All the postdictive effects in the 200 ms range are compatible with our view, but long-lasting postdictive effects are incompatible with fast consciousness. We agree with Shea, however, that the evidence for long-lasting postdiction is still limited and restricted to particular experimental setups. Our arguments are relative, not absolute: as sensory horizons become larger, benefits accrue for both model-based planning and the kind of longer timescales of perceptual integration associated with conscious vision in humans. As MacIver & Dombeck note, even land animals in darkness see significantly further than shallow water fish in full sunlight. These comparisons are relative and do not depend on identifying absolute cutoffs.

Other commentaries (Siegel, Arstila & Railo, and Aranyosi) seek to offer support for fast(er) visual consciousness with appeal to neural data. Siegel notes that the feedforward sweep of processing can operate extremely fast, and that postdictive effects such as masking occur earlier for evolutionary-relevant stimuli. Regarding the first point, the fast feedforward sweep is mostly considered to extract summary statistics (or gist) from visual scenes. Most theories of consciousness regard this initial sweep as unconscious following work indicating a crucial role for feedback for conscious vision (e.g., Lamme, Reference Lamme, Metzinger and Windt2016). It is significant that some of the most reliable effects of unconscious processing are found for ensemble properties (Elosegi, Mei, & Soto, Reference Elosegi, Mei and Soto2024, Reference Elosegi, Amerio, Mei, Valdazo, Mulay, Santana, Cleeremans and Soto2025; Sekimoto & Motoyoshi, Reference Sekimoto and Motoyoshi2022). The contribution of individual features to ensemble judgments is even preserved when features are completely crowded and indiscriminable (Manassi & Whitney, Reference Manassi and Whitney2018). These results suggest that the fast feedforward sweep can extract ensemble properties unconsciously. Conscious perception occurs later, perhaps following a pattern similar to Hochstein and Ahissar’s (Reference Hochstein and Ahissar2002) “reverse hierarchy.” As for the second point: we emphasized that the speed of postdictive effects do not tell us how fast consciousness occurs, but simply provide a lower bound for the speed of consciousness. The fact that with some stimuli masking is ineffective beyond 30 ms does not show that consciousness of the stimulus occurs right before the mask is processed.

Aranyosi and Arstila & Railo both argue that neural data expose an apparently faster timescale of visual consciousness (see also Block). We disagree. Arstila & Railo argue that the Visual Awareness Negativity (VAN), an ERP occurring around 200 ms after stimulus onset, is an early neural correlate of consciousness. The claim that the VAN correlates with consciousness comes from the observation that variations in the amplitude of the VAN correlate with variations of conscious experience as measured by the Perceptual Awareness Scale (e.g., Koivisto & Revonsuo, Reference Koivisto and Revonsuo2010). However, current evidence does not distinguish between the VAN being a good predictor of consciousness and the VAN corresponding to conscious visual awareness itself. In particular, one of us (Michel, Reference Michel2019) has argued that variations in the VAN likely correspond to variations in encoding of the contents of consciousness, rather than consciousness of those contents. Since the vast majority of the studies on the neural correlates of consciousness do not control for variations in signal strength, the amplitude of the VAN might simply correspond to variations in signal strength – which are themselves in turn good predictors of whether or not a subject will consciously experience a stimulus.

Aranyosi suggests that taking phenomenal consciousness seriously implies treating subjective reports as a confounding factor. Our view is that reports are simply indicators that can be used to reveal the contents of conscious experience in awake, attentive subjects. They might be imperfect under some conditions. But they are innocent until proven guilty: absent any theory-neutral reason to doubt the reports, we should believe that subjects are not wrong about what they think they experienced. This is especially the case when using measures such as discrimination capacity, which is less affected by concerns about response biases. Like Block, Aranyosi points out that ERPs obtained in no-report paradigms suggest that the P300 is not a correlate of consciousness. This might be the case, but this does not mean that consciousness takes less than 300 ms (again: just because we can predict from data obtained at time t whether a participant will experience a stimulus does not mean that consciousness occurs at t). Of note, Sergent et al. (Reference Sergent, Corazzol, Labouret, Stockart, Wexler, King, Meyniel and Pressnitzer2021) found that late (∼300 ms) bifurcation dynamics are still good markers of consciousness even without reports.

R9. The timing of consciousness

A question related, but distinct from, the speed of consciousness, is what our account means for the temporal structure of consciousness. One way to pose this question is to ask, as Mamassian does, “When is ‘now’?”

Mamassian notes that “it is not clear whether F&M believe that the organism lives in the past (because of sensation time) or in the future (thanks to the temporal sensory horizon).” Both options seem open to us – if the role of conscious vision is to create a world model for the service of planning, predictions are going to be just as useful as postdictions in updating and organizing temporally extended representations. However, as before, we caution against a strong predictive view of conscious experience. The main reason for this is that we often consciously experience violations of our expectations – which would not be possible if the content of conscious experience were entirely predictive in nature. Mamassian also asks about the duration of conscious experiences themselves – how long is the specious present. As Lee (Reference Lee2014) noted: “It’s easy to confuse the “specious present” … with (1) the objective length of the experience itself or (2) the minimal duration between two events required for them to be perceived as non-simultaneous” (p. 7). We take a broadly representationalist view on this issue. The duration one experiences corresponds to the duration an experience represents. Given a view of this kind, as Lee (Reference Lee2014) writes: “there is no requirement that the amount of time presented by an experience is the same as the length of the experience itself” (p. 7). The duration of the specious present in that sense is the duration an experience represents the present moment as lasting.

de Jong (like Mamassian) asks how to best conceive of the conscious “now.” The windows of integration that we refer to in the target article are both discrete and dynamic – they are characterized by windows of integration that represent dynamic properties. de Jong, like Peters, asks why we are committed to a discrete, serial model (first do model building and then fix the model for planning), rather than allowing for consciousness to fluctuate during the building of the model (e.g., continuous rather than discrete reality monitoring). We already partly address this in footnote 10 in the target article, and in our response to Peters above: the answer is likely that it only makes sense to compare courses of action if they are conditioned on the same perceptual evidence. This creates a computational imperative to engage in alternate sequences of perceptual inference and action.

R10. Anthropocentrism and biases

To close, we highlight commentaries that warn against evolutionary approaches in the science of consciousness falling prey to anthropocentric biases, and “just so” stories motivated by idiosyncratic scientific or moral agendas.

Allen suggests we fall prey to anthropocentric biases in our discussion of the speed of conscious processing, by focusing too narrowly on the human case. We think this is mistaken, for two reasons. First, humans must be the starting point for any research that is seeking functional dissociations between conscious and unconscious processing (Bayne et al., Reference Bayne, Seth, Massimini, Shepherd, Cleeremans, Fleming, Malach, Mattingley, Menon, Owen, Peters, Razi and Mudrik2024). Without the human case as a reference point, inferences about consciousness in animals are not possible. Second, our focus on consciousness being slow in humans enabled us to make claims about its function – revealing that it is ill-suited for immediate reactions to visual stimuli. We then assume that this latter claim can be generalized to other animals. This assumption is in fact anti-anthropocentric: it claims that consciousness in humans is unexceptional, and so should have a similar functional profile in animals.

Mazor argues that consciousness research offers a “marketplace of rationalizations.” Within this marketplace, the view that fish do not consciously experience vision might help rationalize behavior that is harmful for potentially sentient beings, if extended to consciousness in general. It is ironic that this commentary features among a large number of responses that disagree with our main claims – a testament to the fact that most researchers are perfectly capable of resisting the sirens of rationalization. Perhaps that is not so for actors who are not themselves involved in consciousness research, however. So, let us clearly state the obvious: there is much uncertainty about the question of the distribution of consciousness. We tried to reduce this uncertainty as objectively as we could. As the commentaries to our article indicate, the community is not close to a consensus on this question. Anyone who reports that the issue is definitely settled one way or the other is either ill-informed or intellectually dishonest. One should accordingly take this high level of uncertainty into consideration when it comes to relying on consciousness research for informing one’s decisions, for instance, by following a precautionary principle (Birch, Reference Birch2017).

Rationalization might have some influence over theory-selection beyond purely epistemic considerations. Other possible non-epistemic sources of influence might include virtue signaling or social-desirability bias (perhaps holding the view that all animals are phenomenally conscious might be a way of signaling that one is a caring person). While we agree with Mazor that considering non-epistemic influences on theory-selection is important, we nevertheless want to guard against two potential pitfalls. First, we should avoid appealing to (putative) motives in order to discard a view – a version of the genetic fallacy – for the simple reason that it is possible to defend a correct view with sound arguments for the wrong motives. Second, we should carefully distinguish the epistemic question of what we consider most likely to be true from the practical question of what we should do given our level of uncertainty and asymmetric costs of errors. Our responses to the practical questions should be considered on a case-by-case basis, because the costs of errors vary from case to case.