Two consciousnesses and one argument

The distinction at the heart of this debate has a clear origin. In 1995, in On a confusion about a function of consciousness, Ned Block argued that consciousness is a mongrel concept. As his publication summary puts it, phenomenal consciousness is experience: the phenomenally conscious aspect of a state is what it is like to be in that state. Access consciousness is a functional notion. A state is access-conscious when its content is poised for use in reasoning, in the rational control of action and in report. Block's complaint was that scientists and philosophers kept sliding between the two, arguing from facts about access to conclusions about experience and back again.

Making the distinction doesn't show that the two come apart. Many theorists accept the conceptual difference and hold that, in fact, every phenomenal state is also an access state. Global workspace theory, the view that a content becomes conscious when it is broadcast to many specialised systems at once, makes access the mark of consciousness. If the two always travel together, the distinction is a piece of bookkeeping. Block's later work set out to show they don't always travel together. In the 2007 target article Consciousness, accessibility, and the mesh between psychology and neuroscience, and again in Perceptual consciousness overflows cognitive access in 2011, he argued that the capacity of phenomenal consciousness is greater than the capacity of cognitive access. In the 2011 abstract's words, "when observing a complex scene we are conscious of more than we can report or think about."

Why the question matters beyond philosophy

The overflow question decides what a science of consciousness is allowed to measure. Most experiments on the neural basis of experience compare trials on which people report seeing something with trials on which they report not seeing it. If experience regularly outruns report, those comparisons misclassify trials: some "unseen" trials were seen, and the neural difference found is a difference in access, not in experience. Block's argument, if it succeeds, implies that much of the field has been studying the neural basis of report. If it fails, report-based methods are on firmer ground than their critics say, and the main danger is the opposite one, mistaking the neural activity that produces a report for the activity that produces an experience. The same choice reappears in medicine, where clinicians must decide whether a patient who can't respond is aware, and in the study of animals and infants, who can't report in words at all. A clear answer about flashed letters won't settle those cases, but it fixes the method that will be used on them.

The overflow argument, step by step

The argument has more steps than its slogan suggests, and the steps matter, because each one is a place where a sceptic can push. Here is my reconstruction of its empirical version.

1. In Sperling-type displays, people can report only about four items from a brief array of twelve.
2. Cued after the array has gone, they can report most items from any row the cue picks out.
3. So, just after the array, information about most of the items is available to the visual system.
4. That information is held in a specific, item-level form: letters, not just "letter-like stuff".
5. The item-level information is phenomenally conscious at the time, before the cue selects some of it.
6. Only about four items can be accessed (held in working memory and reported).
7. So phenomenal consciousness has a higher capacity than access: it overflows.

Lines 1, 2 and 6 are measurements, and they have replicated many times. Line 3 follows from 1 and 2 with very little extra. The argument's weight rests on lines 4 and 5, and they are the lines this paper is about. Line 4 says the stored information is specific. Line 5 says it was experienced. A sceptic can grant line 3, that a lot of information is held, and deny either 4 or 5. The two main sceptical positions do exactly that. The "generic" reading denies 4: people experience the array as letters in general, a texture of letterhood, and the cue makes specific identities determinate. The "post-perceptual bottleneck" reading denies 5: the information is held, but unconsciously or pre-consciously, and becomes experienced only as it is accessed.

Kavya's hesitation, "I saw something, I'm not sure I saw letters", is a denial of line 4 delivered from the inside. Notice that her introspection can't settle the matter, because the question is what she experienced before she tried to report it, and every attempt to answer is itself a report.

The refrigerator analogy is common in this literature, and I use it because it states the difficulty sharply. It also overstates it, and the place it breaks is where the experimental progress has been made.

What Sperling found

The experiments Block relies on start with a doctoral thesis from the summer of 1959.

The design was simple. Observers saw an array of letters, typically three rows of four, flashed for between 5 and 500 milliseconds, according to a summary of the monograph. In the whole-report condition, they wrote down as many letters as they could. They averaged about 4.5 items, whatever the size of the display or the length of the exposure, which matched what earlier work had called the span of apprehension. In the partial-report condition, a tone after the display told them which row to report. They didn't know which row until the display had gone, so whatever they reported from the cued row must have been available for every row. Observers usually reported about 75 to 90 per cent of the cued row. On the summary's worked example, partial report implied that about 75 per cent of the whole display was available immediately after presentation, against 33 per cent indicated by whole report.

The third result was the time course. As the delay between the display and the tone grew, the partial-report advantage shrank, and by a delay of one second partial report was no better than whole report. Whatever held the extra information faded within a second. Sperling called this a visual information store; Ulric Neisser later gave it the name iconic memory.

What the result shows without interpretation

Three things follow from Sperling's data without any theory of consciousness. First, a brief display leaves more information in the visual system than a person can report from it. Second, that information is selectively accessible: a cue arriving after the display can pick out any part of it. Third, it is short-lived, lasting well under a second in its full form. None of these says anything about experience. They are statements about what is available to a cue. That is why the argument needs lines 4 and 5.

Sperling's observers also said something, and it is the part philosophers latched on to. They reported the impression of having seen more than they could report. That impression is itself a report, of course. It is a report about the overall experience, and its content is "more than I can name". Block treats it as evidence for overflow, alongside the performance data. A sceptic treats it as evidence that people experience a rich-seeming display, which is compatible with the experience being generic. The impression is real; what it is an impression of is the dispute.

Three readings of the same data

The philosophical literature since 2007 has produced three main readings of the partial-report data. Each accepts lines 1 to 3 and the measurements. They differ on lines 4 and 5.

The first reading is Block's. All or most of the twelve letters are experienced as specific letters during the brief period of iconic memory. The cue selects some of them for access. The rest fade, experienced but never accessed. On this reading, line 4 and line 5 are both true, and overflow follows.

The second is the generic reading, which Morten Overgaard set out clearly in Phenomenal consciousness and cognitive access. Observers experience, and access, unclear or "blurry" representations of the letters. When the cue directs attention to a row, those representations become clearer and more accessible. Overgaard's conclusion is that "existing data, in fact, do not demonstrate overflow", because the generic reading fits the findings as well as the overflow reading does. On this view, line 4 is false of experience: experience before the cue is of letter-like items, not of specific letters, and the specific identities are recovered from an unconscious store.

The third is the cue-sensitive reading, defended by Ian Phillips in Perception and iconic memory: what Sperling doesn't show. Phillips draws on the growing evidence for postdictive perception, in which later stimuli shape earlier experience, and offers an interpretation of Sperling's data "in terms of cue-sensitive experience which fails to support any such claims" of overflow. The paper won the William James Prize of the Association for the Scientific Study of Consciousness, which Phillips' CV records as the only time a philosopher has won it. On Phillips' reading, line 5 is false in a particular way: the experience of the cued row is fixed after the cue, so there was no pre-cue experience of specific letters to overflow anything.

Why the readings are hard to separate

All three readings predict the same performance in Sperling's task. That isn't a coincidence. Partial report measures what can be reported after a cue, and all three readings agree about what can be reported after a cue. They disagree about what was experienced before it, which partial report doesn't measure. The data are compatible with each reading because each reading was built to be compatible with the data.

What would separate them is a measure that reaches experience before the cue without going through the cue. Block's answer, in the 2011 paper, is to change the terms of comparison. He argues that the sceptics are committed to two things that need defending: "a peculiar kind of generic conscious representation that has no independent rationale", and "an unmotivated form of unconscious representation that in some cases conflicts with what we know about unconscious representation". That is a cost argument. It says that even if the data don't force overflow, the alternatives are more expensive. I'll come back to it, because I think the first half of it is weaker than it looks and the second half is stronger.

What the cue does

A detail of Sperling's procedure becomes important here. The cue is a tone, and a tone carries no information about the letters. It carries only a location. So whatever the observer reports from the cued row, they report from a store indexed by location, not by identity. That fits all three readings, but it shapes what "specific" can mean in line 4. The store holds information about which shape was where. It doesn't need to hold letter identities as such, since an observer could read the identities off the stored shapes once attention arrives. So the specificity that Block needs is at least the specificity of shape-at-location. Whether that is experienced as "an F, there" or as "a shape, there" is again the dispute.

Cohen, Dennett and Kanwisher's 2016 review What is the bandwidth of perceptual experience? puts the sceptical alternative in its most balanced form. Its argument, as summarised in its abstract, is that we see more than the handful of objects that prominent models of attention and working memory allow, but far less than we think. Its mechanism is ensemble statistics. The visual system computes summaries of groups, such as average size, average orientation, colour variety and the gist of a scene, over the whole visual field, cheaply and outside focal attention. On this view, the richness of the periphery is real but statistical. You experience a crowd's composition without experiencing each face.

Beyond Sperling: the fragile store

The strongest empirical case for overflow doesn't come from Sperling's paradigm. It comes from change-detection experiments run in Victor Lamme's group in Amsterdam, and Phillips himself noted that his postdiction reply "cannot be blocked in this way" for these experiments, though he argued they fail for different reasons.

In Large capacity storage of integrated objects before change blindness, Rogier Landman, Henk Spekreijse and Lamme showed people a circular array of rectangles, then a blank, then a second array in which one rectangle might have changed orientation. Without a cue, people could detect changes in about four items, the usual capacity of visual working memory. With a cue pointing to one location during the blank, after the first array had gone, their performance rose sharply, implying that much more of the first array was still held. The title of their paper states the conclusion: a large-capacity store of integrated objects, holding features bound together, exists before the point where change blindness sets in.

Ilja Sligte, Steven Scholte and Lamme then mapped the stages of this store in Are there multiple visual short-term memory stores?. They varied when the cue appeared: just after the memory array went (an iconic cue), a full second later (a retro-cue), or only when the probe array appeared (a post-cue). The results showed three stages.

First, iconic memory with a capacity they described as unlimited. It "seemed to depend on the strength of the positive after-image" and was "virtually absent under conditions of isoluminance or when intervening light masks were presented". Second, a fragile store lasting about four seconds, with a capacity "at least a factor of two higher" than working memory. It was unaffected by light masks but "completely overwritten by irrelevant pattern masks that spatially overlapped the memory array". Third, the durable, capacity-limited form of visual short-term memory, with a limit of about four objects. Their conclusion was that people "store more sustained representations than is evident from standard change detection tasks", and that these representations "can be accessed at will".

That worked case leaves the overflow argument in a definite place. The empirical premise that survives is a storage claim: a store with more capacity than working memory, lasting seconds, holding bound objects at locations, readable on cue. Call it the fragile store. Line 3 is now well supported, and line 4 is better supported than in Sperling's version, because change detection of specific orientations requires specific information. Line 5, that the store's contents are experienced, still needs a bridge from storage to experience. Block's 2007 paper builds that bridge from a "mesh" between psychology and neuroscience: if the neural signature of the fragile store matches the neural signature of experience found in other cases, we should infer experience. That is a respectable inference. It is also an inference, not an observation, and it inherits whatever doubts attach to the chosen neural signature.

The camera control isn't decisive against Block. A camera's buffer and a brain's fragile store differ in the respects a theory of consciousness might care about: the brain's store is held by activity in the same areas that support seeing, while the camera's buffer is passive memory chips. Lamme's whole point is that this activity is perception continuing, not storage in a separate place. Still, the control makes clear that the inference from storage to experience can't come from the capacity numbers alone.

What a pattern mask tells us

One feature of Sligte's result is easy to overlook and bears directly on the question. The fragile store was overwritten by a pattern mask in the same place. It survived a flash of light, which wipes out the after-image, but not a new set of shapes. That is how visual cortex behaves: new input at a location replaces the representation of old input at that location. It is also how experience behaves: if new shapes appear where old ones were, you see the new ones. The overwriting is evidence that the fragile store lives in the visual system's own format, close to perception, rather than in an abstract memory code. That evidence helps Block, because it puts the store close to where experience is generated. It doesn't settle line 5, because unconscious visual processing also lives in the visual system's own format.

Colour diversity and the 3,456 misses

If the trouble with partial report is that it measures only what a cue retrieves, the obvious move is to find a measure of unattended content that doesn't depend on a cue. Ensemble perception offered one.

In The representation of simple ensemble visual features outside the focus of attention, George Alvarez and Aude Oliva showed that people tracking a few moving objects could still report summary features of unattended objects, such as their average position, even when they couldn't report the individual unattended items. Ensemble statistics seem to be computed broadly and cheaply. This finding is double-edged. It supports the sceptical picture, where experience of the periphery is statistical, and it suggests a method: if people can report a statistic of the unattended items that depends on every item, perhaps the statistic reveals that every item was seen.

Zohar Bronfman, Noam Brezis, Hilla Jacobson and Marius Usher used that method in a 2014 paper with a title that states their conclusion: We see more than we can report: "cost free" color phenomenality outside focal attention. Observers saw a Sperling-like array of coloured letters, with one row precued. After reporting a letter from the cued row, they estimated the colour diversity of the non-cued rows. Colours were drawn either from a narrow part of the colour wheel or from the whole wheel. The abstract reports that "people could estimate the color diversity of the noncued array without a cost to letter report", which the authors took to suggest that colour diversity is "registered automatically, outside focal attention, and without consuming additional working memory resources". Their interpretation, as Overgaard summarises it, was that observers experienced the individual letter colours, though they could report them only as an overall diversity.

The argument is clever. Colour diversity is a statistic that depends on the individual colours. If you can estimate it accurately at no cost to the main task, the individual colours must have been registered. If the estimate is a judgement about how the rows looked, and observers say it is, then the colours were part of how the rows looked. The weak point is the last step. Registering a statistic that depends on every colour is different from experiencing every colour, because the statistic could be computed before experience, and experienced only as a statistic.

That replication is the most important single result in this debate, and it should be read carefully, including the part that doesn't favour the sceptic. Ward is explicit that the result doesn't absolutely rule out rich phenomenology. Observers might have experienced the colours and failed to encode them well enough to compare them with the colours a moment earlier. That would be a failure of memory, not of experience. Her reply is that missing all 3,456 changes makes any such phenomenology functionless and impoverished, so that a theory of rich awareness has to downgrade phenomenology to accommodate the finding. The richness that survives is a richness that does nothing and is noticed by no one, including the person having it.

The objection from memory

The rich model's defender has a reply, and it deserves more than a sentence. Change detection is a memory task. To notice that a colour has changed, you must remember the old colour and compare it with the new one. Change blindness, the well-established failure to notice large changes across a brief interruption, shows that people often don't hold that memory even for attended items. So the 3,456 misses could be 3,456 failures of memory, not of experience. If experience were rich and memory sparse, this is exactly what you would expect.

The reply is fair, and it brings out what the debate has been about all along. Any test of experience that isn't a test of current report is a test of memory, and any test of current report is a test of access. The rich model, by construction, puts experience where neither report nor memory reaches. It can always be saved by saying the experience happened and left no trace. What Ward's result does is put a price on that move. Before her study, the rich model could claim Bronfman's data as positive evidence. After it, the positive evidence is explained by ensemble coding, and what remains is the bare possibility of rich experience that leaves no trace.

Can we look without opening the door?

The refrigerator problem has a modern answer in method, if not yet in results. If report contaminates the measurement, remove the report. No-report paradigms show people stimuli without asking them to say anything about them, and infer what they experienced from some other signal: eye movements that track a percept, pupil size, or neural activity that has been calibrated against report in a separate session. The aim is to find the neural basis of experience without picking up the neural basis of reporting it.

Overgaard's assessment of this work is careful, and I adopt it. In his words, "no-report paradigms provide evidence, that may not be conclusive, that prefrontal activations sometimes found in experiments searching for neural correlates of consciousness relate to report rather than experience itself." He gives two reasons for caution. Neural activation alone makes it hard to infer function. And no-report studies still depend on earlier report-based sessions to establish which stimuli go with which experiences, so report hasn't been removed, only moved to the calibration step.

That second point is the general form of a difficulty this whole debate faces. Every route to experience runs, somewhere, through a report. The best we can do is put the report where it does least damage. In Sperling's task, the report comes after a cue that may itself shape the experience. In Ward's task, it comes after a comparison that depends on memory. In a no-report study, it comes in a calibration session where the stimulus is attended. Each placement leaves a different gap for the sceptic or the overflow theorist to use, and the research programme that might settle the question is one that triangulates: the same stimuli, several placements of report, and a model that predicts all of them.

Before I take that programme further, a third analogy, because the debate is partly about what kind of thing a store of unaccessed information is.

Bridging from storage to experience

Every version of the overflow argument has to cross from a claim about storage to a claim about experience, and there are only a few ways across. The diagram below sets them out.

The first bridge is introspective: observers say they saw more than they can report. It is real evidence of an impression of richness, and weak evidence about what the impression contains. The second is the neural mesh: the fragile store is held by activity of a kind that, in other cases, goes with experience. This is the strongest bridge, and it depends on a theory, recurrent processing theory or something like it, that identifies the relevant activity as sufficient for experience. The third is the cost argument: the alternatives to overflow are more expensive. The fourth is the behavioural signature of experience in tasks that aren't the report task: priming, after-effects, interference. That bridge is where future work could carry the most weight, and it has been little used for this question.

The point of laying these out is that the overflow debate isn't a single disagreement. It splits into separate ones about each bridge, and they have different shapes. The introspective bridge is a dispute about what impressions are evidence for. The mesh bridge is a dispute between theories of consciousness, and so it can be settled only as those theories are tested. The cost bridge is a dispute about theoretical economy. I take them in turn below, but my answer to the paper's question depends mainly on the second, because it is the only one that could deliver item-level overflow as a matter of evidence.

Dignāga's perception without names

The Indian philosophical traditions argued for centuries about whether a cognition is aware of itself, and about what perception is before concepts arrive. Both arguments bear on this debate more directly than parallels usually do.

Dignāga, the Buddhist epistemologist of roughly the sixth century, opens his Pramāṇasamuccaya by defining perception. The definition at verse 1.3c is three words: pratyakṣaṃ kalpanāpoḍham, perception is free from conceptual construction. The Austrian Academy of Sciences' edition of the first chapter gives the gloss that follows: "yasya jñānasya kalpanā nāsti, tat pratyakṣam", the cognition in which there is no conceptual construction is perception. A survey of Dignāga's epistemology explains that he immediately specifies "conceptual construction" as "association with names and classes and so forth". On this account, perception grasps the particular (svalakṣaṇa). Anything that sorts the particular under a kind, such as "letter" or, more narrowly, "F", is the work of concept, which grasps only universals (sāmānyalakṣaṇa).

That gives Kavya a third option that the modern debate tends to skip. Block says she experienced specific letters. The generic reading says she experienced letterhood. Dignāga's frame permits something else: she perceived specific particulars, these exact shapes at these exact places, without perceiving them as letters at all. Their being letters, and which letters, is conceptual work that comes later, and on the Sperling design it comes after the cue. The particular is fully given; its name is not.

This isn't a neat solution, but it does pull apart two things the modern argument runs together in line 4. "Specific" can mean determinate in its sensory detail, or it can mean identified under a category. Block needs specific in the first sense for experience and gets evidence of specific in the second sense from report. The report is of letter identities, which are categories. If perception is specific in detail and silent on category, then the overflow argument's line 4 is true of experience in one sense and false in the other, and the generic reading is wrong about detail but right about identity.

The memory argument for self-awareness

Dignāga's second relevant idea is his argument that every cognition is aware of itself (svasaṃvedana, or svasaṃvitti). Its structure is worth setting beside Block's, because it is the same inference pointed at a different conclusion.

In the commentary on Pramāṇasamuccaya 1.11, Dignāga argues from memory. Some time after we perceive an object, we remember not only the object but our having perceived it. An article on svasaṃvedana in Bhartṛhari and the Pratyabhijñā quotes the Sanskrit: "yasmāc cānubhavottarakālaṃ viṣaya iva jñāne 'pi smṛtir utpadyate, tasmād asti dvirūpatā jñānasya svasaṃvedyatā ca", translated there as "Since memory arises in a moment that follows that of the perception, [and arises] for the content and the cognition as well, hence a cognition has two forms and a self-cognizing nature." The premise doing the work is that whatever is remembered was experienced before. If I remember having seen the blue lotus, my seeing it must have been experienced, and since a second cognition to experience the first would start a regress, the first cognition must have been aware of itself.

The parallel is exact enough to be useful. Block reasons from a later report, enabled by the cue, to an earlier experience of the item. Dignāga reasons from a later memory to an earlier self-experience of the cognition. Both rely on the principle that later access to something shows earlier experience of it. And both met the same objection from their own traditions. As a study of Dignāga and Candrakīrti records, Candrakīrti argued in his autocommentary on the Madhyamakāvatāra, at 6.74, that memory and the original experience are distinct events, so memory can't be used to read off the structure of the experience it follows. A chapter on Dignāga and later developments notes a narrower worry: the memory argument, strictly, proves self-awareness only of cognitions that are, or can be, remembered.

That narrower worry is Phillips' point about Sperling, stated about fifteen centuries earlier. A cue-enabled report proves experience, at most, of what the cue reaches, and only if the report doesn't help make what it reports. Candrakīrti's broader objection is the post-perceptual bottleneck: the later event is a different event, and its contents don't transfer backwards.

Where this sits on RTM's ladder

RTM's five layers give the overflow question a clean statement. Phenomenal consciousness, in Block's sense, corresponds to L2, prakāśa, manifestation: that by which anything appears. Access corresponds to the functions of L3 and L4, vimarśa and pratyabhijñā: reflexive grasp and recognition, the stages at which a content is taken up, named and made available for action. The overflow claim, on this mapping, is that L2 has a larger capacity than L3 and L4: more is manifest than is grasped.

The Pratyabhijñā tradition, which RTM follows on this point, holds that manifestation without reflexive awareness is not consciousness at all. In the last paper in this series I quoted Utpaladeva's image of the crystal: light without vimarśa is like a crystal coloured by the objects around it, and insentient. If that is right, then item-level manifestation without any vimarśa is the crystal, not experience. That doesn't mean RTM predicts that experience is as sparse as working memory, because vimarśa in the tradition isn't working memory. It is a reflexive self-grasp present in every cognition, closer to Dignāga's svasaṃvedana than to the four-item limit. So RTM's position is that what is manifest must also be reflexively grasped, though not necessarily reported, held or named. That puts it close to the reading I've reached from the data: experience includes whatever the reflexive grasp can hold at a glance, the gist, the layout and the statistics, and it doesn't include a per-item manifestation that no grasp touches. I state this plainly because it is a place where RTM's metaphysics and my reading of the evidence agree, and an honest reader should discount the agreement accordingly.

The same question in a machine

The overflow debate was framed for human vision, but it now has a machine counterpart, and setting the two side by side shows which parts of the debate are about evidence and which are about definitions.

A large language model processing a long document holds a great deal of information at once in its context. At any step, the computation that produces the next word draws on some of that information more than the rest. Researchers can probe the model's internal activations and recover information about parts of the context that the current output doesn't use. In functional terms, that is a store with more capacity than any single act of use, readable on demand. It is the fragile store's shape, built in silicon.

Nobody, as far as I know, argues from this that the model experiences its whole context. The camera in thought experiment 3 makes the same point, and the machine case sharpens it. What would make anyone say the model experiences more than it uses? Only one of the four bridges could carry the claim: a theory of consciousness saying that some kind of processing, present in the model, suffices for experience. The introspective bridge is unavailable in a useful form, because a model's reports about its own states are trained on human reports and can't be taken at face value. Recent work has begun to test whether models can report injected internal states accurately; Anthropic's study of emergent introspective awareness found such reports succeeded only some of the time, at particular layers and strengths. The cost bridge is unavailable too, since there is no rival theory of the model's experience whose costs could be compared. And the trace bridge, looking for effects of unused context on later behaviour, would show only that the information was processed, which nobody doubts.

That leaves the machine case exactly where the human fragile store was left in the second worked case: a measurement of storage plus a theory of which activity is experienced. The machine case simply makes the dependence obvious, because it strips away the impression of richness that carries so much weight in the human case. Kavya feels she saw the whole page. The model, so far as anyone can tell, feels nothing that we can check. Remove the impression and what remains of the overflow argument is the theory.

This matters for the human debate in a specific way. If the only bridge that can carry item-level overflow is a theory of consciousness, then the overflow question isn't a question that partial-report experiments can answer on their own, however refined they become. It is a downstream consequence of the choice between theories. Block saw this, and his 2007 argument is built as an inference to the best explanation that combines psychology and neuroscience. My disagreement with him is about the weight of the psychological part. I think the behavioural evidence, once Ward's result is added, leans towards sparse item experience, so that an overflow conclusion has to come from the neuroscience against the behaviour, not with it.

There is one more lesson in the comparison, and it runs the other way. If a theory such as recurrent processing theory is confirmed, and it implies that certain kinds of sustained, integrated, local activity suffice for experience, then the question of whether machines have their own fragile, unreported experiences becomes a question about whether their architecture contains that kind of activity. The overflow debate is, in that sense, a rehearsal for the machine debate. It teaches the same discipline: separate storage from experience, name the bridge, and ask what trace the claimed experience leaves.

The strongest case against my answer

The best case against my answer is Block's cost argument, and part of it is stronger than I have so far allowed.

Block argues that sparse views need "an unmotivated form of unconscious representation that in some cases conflicts with what we know about unconscious representation". The point is this. The fragile store holds bound objects, specific orientations at specific places, unaffected by light masks and lasting seconds. Unconscious visual representations, as we know them from masking studies, are usually weak, brief and unbound. So if the fragile store is unconscious, it is an unusual kind of unconscious representation: strong, lasting and integrated, exactly the properties that elsewhere go with consciousness. The sceptic has to posit a new kind of unconscious state to avoid admitting a familiar kind of conscious one. That is a cost, and I think it is the most serious one in the debate.

My answer is partial. The ensemble model doesn't need the fragile store to be unconscious in the way masked stimuli are. It can say the store is the substrate of a conscious experience whose content is gist and statistics, while the item-level detail in it is accessible but not itself manifest. That is a claim about which aspects of a representation are experienced, not about whether the representation is conscious. It is consistent with Ward's result: the rows were experienced, as diverse, and the individual colours weren't experienced in a comparable form. But I accept that this splits a single representation into experienced and unexperienced aspects, and that such splits have their own costs. Block's other cost charge, that the generic view posits "a peculiar kind of generic conscious representation that has no independent rationale", is weaker after 2008, because ensemble perception supplies the rationale. Generic, statistical experience isn't peculiar if the visual system computes ensembles everywhere.

A second objection says I am treating Ward's study as more decisive than one study can be. Twelve participants in each of two experiments is a small sample, and a null result, noticing no changes, is harder to interpret than a positive one. A different procedure might reveal noticing. That is fair, and it is why I hold the item-level claim at "likely" rather than "established". What makes 3,456 misses informative is not the sample size but the consistency: not one participant noticed even once, across trials in which every unattended colour changed.

A third objection, from the sceptic's side, says I am too generous to overflow. Granting a "genuine overflow of storage and of gist" concedes the slogan while denying the substance, and the honest conclusion is simply that experience doesn't overflow access. I disagree, because gist and statistics are accessible as summaries while their basis isn't. The colour diversity of a row depends on twelve colours that can't be reported; the experienced diversity carries information from all of them. That is a real way in which experience contains more than can be itemised, and calling it "no overflow" hides something Block was right to point at.

A counterexample and a failure boundary

The counterexample that most troubles my reading is the after-image. Sligte's iconic stage depended on retinal after-images, and after-images are experienced. An after-image of a twelve-letter display carries the whole display. So for a short period, people do experience something that carries more items than they can report. That is item-level overflow of a sort, and I have to grant it. My reply is that it is overflow of the trivial kind that any continuing stimulus produces: when you look at a page, you experience more than you can name, because the page is in front of you. The after-image is the page still in front of you, by other means. The interesting question was always whether experience outruns access once the stimulus, in every form, has gone, and that is the question the fragile store raises.

The failure boundary of my answer is set by recurrent processing theory. If local recurrent activity in visual cortex is sufficient for experience, and if the fragile store is held by such activity, then the store is experienced, and my reading fails for the fragile stage. My answer is therefore conditional on that theory not being confirmed, or on the fragile store being held in some other way. The adversarial testing of theories of consciousness, which the next paper in this series examines, is the route by which that condition will be tested.

What would change my mind

What I don't know

I don't know whether postdiction, the shaping of earlier experience by later events, is large enough to account for the partial-report advantage, or only a part of it. I don't know whether the fragile store is held by the kind of activity recurrent processing theory picks out; the experiments that would show it are hard to do in humans. I don't know whether "experienced aspects of a representation" is a stable notion or a way of naming the problem. I don't know how to apply any of this to machines: a language model's context window is a store with more capacity than any single computation uses at once, and whether anything of the overflow question carries over to it is open. And I don't know whether Dignāga's distinction between particular and category can be made precise enough to test, though it is the conceptual tool from this paper I would most like to see put to work.

A tool for any overflow claim

Claims that "we experience more than we can report" are now made about people, infants, animals and machines. Here is a worksheet I use to locate what such a claim has shown.

Claim Ledger

Dated predictions

Glossary

Access consciousness

Block’s functional notion: a state whose content is poised for reasoning, control of action and report.

Ensemble statistics

Summary properties of a group, such as average size or colour diversity, computed by the visual system.

Fragile store

A visual memory lasting about four seconds, larger than working memory, overwritten by new patterns in the same place.

Iconic memory

A very brief visual store revealed by partial report; Sligte’s work links much of it to after-images.

Kalpanā (कल्पना)

Conceptual construction; for Dignāga, association with names, classes and so forth.

Kalpanāpoḍha (कल्पनापोढ)

Free from conceptual construction: Dignāga’s definition of perception.

No-report paradigm

A design that infers experience without asking subjects to report on the stimulus.

Overflow

The claim that phenomenal consciousness has a higher capacity than cognitive access.

Partial report

Sperling’s method: a cue after a brief display indicates which part to report.

Phenomenal consciousness

Experience: what it is like to be in a state.

Postdiction

The shaping of an experience at one moment by events a fraction of a second later.

Prakāśa (प्रकाश)

Light or manifestation; in Pratyabhijñā, that by which anything appears. RTM’s L2.

Sāmānyalakṣaṇa (सामान्यलक्षण)

The universal or general character, the object of conceptual cognition for Dignāga.

Svalakṣaṇa (स्वलक्षण)

The unique particular, the object of perception for Dignāga.

Svasaṃvedana (स्वसंवेदन)

Self-awareness: the reflexive awareness of a cognition by itself.

Vimarśa (विमर्श)

Reflective awareness; in Pratyabhijñā, the self-grasping that makes manifestation conscious. RTM’s L3.

Kavya's second look

Dev ran one more trial on his aunt's laptop. He flashed the grid, and on this trial gave no tone at all. "What did you see?" he asked. "Three rows of letters," Kavya said. "Black on grey. The top row looked crowded." He asked which letters. She laughed and said she would have known, if he had asked in time.

That answer is, I think, the right one, and it is the answer this paper defends. The Sperling family of experiments shows a store of visual information much larger than working memory, lasting up to a few seconds, from which a cue can fetch items. It shows that summary properties of unattended items are registered and experienced. It doesn't show that the individual items in the store are experienced before they are fetched, and the replication that added change detection suggests they aren't, at least not in any form that supports comparison. So the best reading is a post-perceptual bottleneck for items, with a real overflow of gist. "I would have known, if you had asked in time" is a precise description of a fragile store; "three rows, the top one crowded" is a precise description of ensemble experience.

The practical step is this. The next time someone tells you that a person, an animal or a machine experiences more than it can report, ask three questions before agreeing. What is the evidence of storage, and how long does it last? What bridge has been built from storage to experience, and does that bridge rest on a theory of consciousness that is itself in dispute? And has anyone checked, in the way Ward checked, whether the claimed rich experience leaves any trace at all? Score the claim on the worksheet. A claim that fails the third question isn't refuted. It has moved into the place where no experiment reaches, and you are entitled to ask why it should be believed there.

Sources

Primary research

• Sperling, G. (1960). The information available in brief visual presentations. Psychological Monographs 74(11), Whole No. 498, 1–29. dn760100.eu.archive.org/0/items/sim_psychological-monographs_1960_74_1
• Sperling, G. Measuring the time course of the information available in brief visual presentations (retrospective). Journal of Vision. jov.arvojournals.org/article.aspx?articleid=2143757
• Block, N. (1995). On a confusion about a function of consciousness. Behavioral and Brain Sciences 18(2), 227–247. www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/
• Block, N. (2007). Consciousness, accessibility, and the mesh between psychology and neuroscience. Behavioral and Brain Sciences 30(5–6), 481–499. pubmed.ncbi.nlm.nih.gov/18366828/
• Block, N. (2011). Perceptual consciousness overflows cognitive access. Trends in Cognitive Sciences 15(12), 567–575. pubmed.ncbi.nlm.nih.gov/22078929/
• Landman, R., Spekreijse, H. and Lamme, V. A. F. (2003). Large capacity storage of integrated objects before change blindness. Vision Research 43(2), 149–164. pubmed.ncbi.nlm.nih.gov/12536137/
• Sligte, I. G., Scholte, H. S. and Lamme, V. A. F. (2008). Are there multiple visual short-term memory stores? PLoS ONE 3(2), e1699. pubmed.ncbi.nlm.nih.gov/18301775/
• Alvarez, G. A. and Oliva, A. (2008). The representation of simple ensemble visual features outside the focus of attention. Psychological Science 19(4), 392–398. pubmed.ncbi.nlm.nih.gov/18399893/
• Bronfman, Z. Z., Brezis, N., Jacobson, H. and Usher, M. (2014). We see more than we can report: “cost free” color phenomenality outside focal attention. Psychological Science 25(7), 1394–1403. pubmed.ncbi.nlm.nih.gov/24815608/
• Ward, E. J. (2018). Downgraded phenomenology: how conscious overflow lost its richest source of evidence. Philosophical Transactions of the Royal Society B 373, 20170355 (reporting Ward, Bear and Scholl 2016, Cognition 152, 78–86). pmc.ncbi.nlm.nih.gov/articles/PMC6074092/
• Overgaard, M. (2018). Phenomenal consciousness and cognitive access. Philosophical Transactions of the Royal Society B 373. pmc.ncbi.nlm.nih.gov/articles/PMC6074085/
• Phillips, I. B. (2011). Perception and iconic memory: what Sperling doesn’t show. Mind & Language 26(4), 381–411. onlinelibrary.wiley.com/doi/10.1111/j.1468-0017.2011.01422.x
• Cohen, M. A., Dennett, D. C. and Kanwisher, N. (2016). What is the bandwidth of perceptual experience? Trends in Cognitive Sciences 20(5), 324–335. pubmed.ncbi.nlm.nih.gov/27105668/

Primary texts and translations

• Dignāga, Pramāṇasamuccaya chapter 1, Sanskrit reconstruction (Steinkellner), Austrian Academy of Sciences.
• Pramāṇasamuccayavṛtti on 1.11ab, Sanskrit and translation as quoted in Studies on Bhartṛhari and the Pratyabhijñā: the case of svasaṃvedana, Religions 8(8), 145. www.mdpi.com/2077-1444/8/8/145

Secondary or leads

• The philosophical works and influence of Dignāga and Dharmakīrti: epistemology, nonconceptual perception and its idealist implications. Oxford Handbook chapter. academic.oup.com/edited-volume/62249/chapter/551357371
• Dignāga and later developments. In Self-Awareness in Tibetan Buddhism, Edizioni Ca’ Foscari.
• The concept of svasaṃvedana in Dignāga and Candrakīrti. Contemporary Buddhism. www.tandfonline.com/doi/epdf/10.1080/09552367.2022.2119674
• Summary of Sperling (1960) by D. Z. Hambrick. www.jimdavies.org/summaries/sperling1960.html
• Ned Block, publications list (PhilPeople). philpeople.org/profiles/ned-block/publications
• Ian Phillips, curriculum vitae (William James Prize note). www.ianbphillips.com/uploads/2/2/9/4/22946642/cv_--_ian_phillips_july_