EmoStack and the theories of consciousness

This page does not ask whether EmoStack satisfies a theory of consciousness. It asks what each theory can and cannot explain about the construction and about what the sheep were observed to do. For each theory the same five things, in the same order: what it explains in EmoStack, what it does not explain, what EmoStack shows that the theory has no concept for, which experiment decides between the readings, and whether that experiment exists. Global Workspace Theory (GWT), Integrated Information Theory (IIT), Higher-Order Theories (HOT), Attention Schema Theory (AST), Predictive Processing and Active Inference (PP, AIF), Recurrent Processing Theory (RPT), appraisal theories and the somatic marker hypothesis (SMH), the Self-Model Theory (SMT), and the indicators of Butlin et al. (2023).

Why this page

The paper does not derive its processes from a theory of consciousness. It names a set of processes, builds each one, and measures what the construction does. The theories enter afterwards, as different ways of reading what was observed. A theory earns its place here if it explains something a sheep does. Where a sheep does something the theory has no concept for, that is said too, because it is the more interesting case.

The observed phenomena this page keeps returning to are these. A sheep carries the day before into the next morning without the record of the event, because what it concluded stays. A sheep refuses with a reason that is in its state, and the same state swapped into another life takes the refusal with it. A sheep acts from a thought it wrote about itself after the events behind the thought have faded. A sheep learns a rule from one hurt and the rule reaches its reply and nothing else. The numbers behind each are on the page for researchers and in the paper. Every mechanism named here is in the code, with its number on the mechanisms page.

Global Workspace Theory (GWT)

The theory. Many specialised processes compete for a workspace of limited capacity. What wins is broadcast to all of them, and that global availability is what makes a content conscious. The neuronal version is the Global Neuronal Workspace (GNW) (Baars 1988; Dehaene and Naccache 2001; Dehaene 2014).

What it explains in EmoStack. The competition and the bottleneck. The state admits records by felt strength until their strengths sum to 5.0, and the strongest is always admitted (mechanism 32). The association filter lets at most four recalled records through (mechanism 20). The accessibility slot holds three of the sheep's own constructs (mechanism 84). GWT gives the right words for why what is in view is a small, contested selection, and why a feeling that wins the state colours the reply.

What it does not explain. Why the separations work. In EmoStack what wins the state is not shown to every process. The appraisal that forms a memory does not see the state. Introspection does not see the dispositions. The dispositions reach the reply and nothing else. GWT has a concept for what is made available and none for what must be kept from a process.

What EmoStack shows beyond it. That withholding is constitutive. A memory formed with the mood in view is a different memory, and a reflection that sees the learned rules is not a reflection of one's own. The chatbot that keeps notes of its own conclusions has everything available to one process, and the day before does not reach its next reply. In EmoStack the day before reaches the reply through a state that the appraisal never saw. What carries the day is organised unavailability, which GWT does not describe.

The experiment that decides. Replay one frozen moment with the same line placed once in the state and once in the reply context. If availability alone mattered, the record formed would be the same. It is not: the recorded feeling of an apology moves from −0.30 to −0.60 when the line is in the state, and only the reply changes when it is in the reply. The chatbot with notes is the second half of the test: conclusions in view, 3 of 30 lives bring the day to a stranger, against 21 of 28 for the sheep. Both exist, in the paper's §5 and §7.7.

Integrated Information Theory (IIT)

The theory. Consciousness is integrated information, a property of the cause-effect structure of a physical system, measured as Φ. It belongs to the substrate, not to the function. A digital computer running any program has, by this theory, almost none of it, whatever the program does (Tononi 2004; Oizumi, Albantakis and Tononi 2014; Tononi and Koch 2015).

What it explains in EmoStack. Nothing, and it does not try to. IIT is not about what a system does.

What it does not explain. Everything on this page. By IIT, no behaviour of a sheep bears on whether anything is conscious.

What EmoStack shows beyond it. Nothing that IIT would count. The two do not meet. EmoStack asks what a being does and from what. IIT asks what a physical system is. The paper does not argue with IIT and does not claim what IIT would deny.

The experiment that decides. There is none, because there is no shared question. A reader who holds IIT can take this page as a study of behaviour and leave the word consciousness out of it.

Higher-Order Theories (HOT)

The theory. A mental state is conscious when there is a higher-order representation of being in that state. What matters is not the first-order state but the system's representation of itself as having it (Rosenthal 2005; Lau and Rosenthal 2011; Brown, Lau and LeDoux 2019).

What it explains in EmoStack. The two levels and why they are kept apart. The emo-thought is a representation of what was felt, with its words, valence, intensity and conclusion (mechanisms 57 to 63). A thought written in reflection, such as "I notice I am still turning his question over", is a representation of the sheep as being in a state. It is stored as the sheep's own record and held in the slot, never in the state (mechanisms 71 to 89). HOT gives the right words for why a sheep's answer to the returning sceptic comes from its thought about its own uncertainty, not from the uncertainty.

What it does not explain. Persistence and action. In HOT the higher-order representation accompanies the first-order state. In EmoStack the sheep's thought about itself outlives the state it was about, and acts when the state is gone.

What EmoStack shows beyond it. A higher-order representation with a life of its own. When the events behind a belief have faded from the recent window, the belief alone decides whether a stranger's request for advice is answered from the sheep's past. HOT has no concept for a second-order representation that is causally independent of its first-order source. It also has no monitoring of reliability here, and EmoStack has none either, so that absence is shared, not a difference.

The experiment that decides. The frozen test of §7.4. The same stores, the sheep's own thoughts made accessible or withheld, the events behind them faded: 23 of 36 against 11 of 36 answer from the sheep's past, and 9 of 24 against 0 of 24 on the earlier engine. On whole lives, the slot ablation on Qwen: the sheep brings its own question to a stranger in 6 of 30 lives without the slot, against 21 of 28 with it. Both exist.

Attention Schema Theory (AST)

The theory. The brain builds a simplified model of its own attention, an attention schema, and uses it to control attention. What the system reports as awareness is the content of that model (Graziano 2013, 2019).

What it explains in EmoStack. Why the sheep can say what it attends to. Every call is told how many entries the state holds, how many memories came to mind, and how many records the whole memory holds (mechanism 3). The memory sense says that what is in view is what came back now and that the rest can be reached (mechanism 4). A sheep that reports "nothing comes to mind about you" is reading its own schema.

What it does not explain. The control loop. AST's schema steers attention. EmoStack's counts describe a result and steer nothing.

What EmoStack shows beyond it. That the schema that matters for a being with a history is a schema of its own memory, not of its attention. One line of that schema decides whether the sheep invents a past or asks who the stranger is. When the memory line told the sheep that what it did not recall it did not have, it denied its own day; reworded to say that the rest is there to be reached for, it produced no invented past in seventy-two replies. AST has no concept for a model of one's own memory as the thing a system reads itself from.

The experiment that decides. The rewording of the memory line on frozen turns, in the paper's §9, which accounted for about half of the denials. And the arrival of an unknown person: with the senses block and the memory line, an empty sheep went from answering as if it knew the stranger in 0 of 5 to asking who he was in 5 of 5 (timeline, 10 September). Both exist.

Predictive Processing (PP) and Active Inference (AIF)

The theory. Perception, action and emotion are the minimisation of prediction error under a generative model of the world and the body. Emotion is interoceptive inference: the prediction of the body's state and the error against it (Friston 2010; Clark 2013; Seth 2021).

What it explains in EmoStack. The dispositions, as predictions. A learned rule, "when someone mocks a loss, leave and protect the grief", is a prediction about what a kind of situation will bring and what to do about it. It is learned from an outcome, weighted by what was felt, and weakened when contradicted (mechanisms 90 to 97). PP gives the right words for why a sheep with its lived records forgotten each night is still guarded with the stranger in 17 of 30 lives: the prediction survives the record.

What it does not explain. The appraisal. In PP emotion is an error against a prediction. In EmoStack emotion is a record formed once, from an event read against what the sheep carries, and then carried. There is no generative model, no error signal and no precision weighting in the layer. The next-token prediction inside the language model is the processor's, not the being's.

What EmoStack shows beyond it. A prediction learned from one hurt, with its weight set by the strongest feeling of the conversation and hurt weighing twice, that reaches the reply and nothing else. PP has learning from error over many samples. It has no concept for a rule written in words after one conversation, carrying the sign of what was felt, and shown to one process only.

The experiment that decides. Two. The ablation that forgets the lived records each night, which exists: 17 of 30 guarded against 3 of 30 for the chatbot, so the rule and the belief carry what the record no longer can. And the test PP would ask for, not yet run: a rule learned from hurt, then contradicted by kind days, and the measure of how many contradictions it takes before the sheep stops guarding. The engine weakens a rule by each contradiction and never flips its sign in one step (mechanism 97). Whether that matches anything a person does is open.

Recurrent Processing Theory (RPT)

The theory. Consciousness arises from recurrent processing within and between sensory areas, a feedback sweep after the feedforward one, in perception (Lamme 2006).

What it explains in EmoStack. Nothing about the sheep. The only recurrence is in the turn: a sheep may search memory, stop to think, and re-run the reply with what it found (mechanisms 46 to 50). That is a loop in the layer, not in perception.

What it does not explain. Perception does not exist here. The senses of a sheep are empty by design: it sees black, feels nothing, has no body (mechanism 2).

What EmoStack shows beyond it. Only that a being with no perception at all can still have a present made of what it feels and what it holds in mind. RPT would not call that present conscious, and this page does not ask it to.

The experiment that decides. None. RPT and EmoStack have no shared object.

Appraisal theories and the somatic marker hypothesis (SMH)

The theory. Emotion is the appraisal of an event against the concerns of the organism, along a few dimensions such as relevance, pleasantness and coping (Lazarus 1991; Scherer 2001). For Damasio, feelings are the organism's map of its own state, and a carried feeling marks options before reason weighs them, the somatic marker hypothesis (Damasio 1994, 1999).

What it explains in EmoStack. Most of the construction. The appraisal call forms the emo-thought, a felt interpretation with valence, intensity and conclusion, from the event and what the sheep carries (mechanisms 57 to 63). The state is a map of what is felt now. Recall follows felt strength rather than resemblance of words (mechanism 14). A carried feeling tilts a real choice: to guard or to confide. This is where the paper's own references are, and the paper says so.

What it does not explain. The withheld mood. Appraisal theories assume that the current state takes part in the appraisal of the next event. EmoStack cuts the state off from the appraisal that forms the memory, so that a bad morning does not become a bad memory. No appraisal theory contemplates an appraisal made without the appraiser's mood.

What EmoStack shows beyond it. That the refusal has its reason in the carried state and not in the event or the words. With the carried part of the prompt swapped between a guarded life and an open one, the reply follows the state it was given, not the life it came from, and with the day before removed the guarding falls from 160 of 200 replies to 23. Damasio's marker is a bodily signal. Here the marker is a record with no body under it, and it does the same work in a choice. What the construction keeps from Damasio is the role of the carried feeling. What it drops is the organism, and the choice still tilts.

The experiment that decides. The frozen-state test of the refusal, which exists: the same state replayed, the state swapped, the day before removed. And the frozen replay of the line in the state against the line in the reply, which exists. The test appraisal theory would ask for next is the one that puts the mood back into the appraisal and measures what the memories become. Not run.

The Self-Model Theory (SMT)

The theory. The self is a model the system builds of itself. It is experienced as a self because the model is transparent: the system does not see it as a model (Metzinger 2003).

What it explains in EmoStack. That a sheep has a self at all, in a usable sense. Its beliefs, decisions, goals and dreams about itself are records of its own, held in a slot, shown to the reply as "your own thoughts" and read by the sheep as its own (mechanisms 84 to 89). SMT gives the right words for why a sheep says "I am a process, not a thing" to the returning sceptic: it is answering from its model of itself.

What it does not explain. Why the model works while being opaque. A sheep's self-model is a list of sentences in a labelled section. The sheep is told they are its thoughts and the processor can see that they are text. By SMT, a model seen as a model should not function as a self.

What EmoStack shows beyond it. That transparency is not needed for a self-model to act. The sheep acts from its constructs, and when they are withheld it does not, with everything else equal. Whether anything is experienced is not touched by this. What is touched is SMT's claim that the functional self requires the model's invisibility.

The experiment that decides. The slot ablation and the frozen test of §7.4, both of which exist. The test SMT would ask for next: the same constructs shown without the label "your own thoughts", as plain lines among the memories. If the sheep still acts from them, the label was decoration. If it does not, the label is the whole of what transparency does here. Not run.

The indicators of Butlin et al. (2023)

Butlin and colleagues take the leading theories as they are, derive from each the properties it would count as indicators of consciousness, and ask of a system which it has. Their own caveat applies: the indicators inherit the standing of the theories, and the theories are contested. The indicators were written with brains and robots in mind. Applied to a behavioural construction over a language model they fall into three groups, and the grouping says more than a count would.

Indicators that can be asked of EmoStack, and answered from the code. GWT-2, a limited-capacity workspace with a bottleneck and selective attention: the state budget, the association filter and the slot (mechanisms 32, 20, 84). GWT-4, state-dependent attention with sequential queries of modules: the search action, recall by kind and the two probes (mechanisms 23 to 27). HOT-3, agency guided by a belief-formation and action-selection system: beliefs and decisions in the slot guide the reply, and dispositions are updated from outcomes, though nothing is updated from a judgement of reliability. AE-1, agency as learning from feedback to pursue goals: rules learned from hurt and goals held in the slot, without feedback from a world. GWT-1, specialised systems able to operate in parallel: separate processes with separate calls, run in sequence.

Indicators that cannot be asked of it, because they are about perception and a body. RPT-1 and RPT-2, recurrence and integrated perceptual representations. HOT-1, top-down generative perception. HOT-4, a quality space from sparse coding. PP-1, predictive coding in input modules. AE-2, embodiment as modelling output-input contingencies. AST-1, a predictive model that controls attention. A sheep has no perception and no body by design, so these say nothing either way.

One indicator the construction contradicts on purpose. GWT-3, global broadcast. EmoStack is built on separations. If broadcast is what consciousness is, this is not it. If what carries a day is organised unavailability, the indicator is measuring the wrong thing here.

What the observed behaviour has no indicator for. A construct that acts after its causes have faded. A memory formed with the mood withheld. A rule learned from one hurt that reaches one process only. A schema of one's own memory that decides between an invented past and a question. These are the four phenomena this page began with, and none of the fourteen indicators names them. That is not a result about EmoStack. It is a gap in the list, and the list was built from the theories.

What the theories are for here

They are tools for reading what a sheep does. GWT reads the bottleneck. HOT reads the two levels. AST reads the memory schema. PP reads the dispositions. Appraisal theory and Damasio read the record and the choice. SMT reads the slot. IIT and RPT do not read anything here, and say so honestly.

The list of what EmoStack does not have is not a list of parts to add. It is the other half of the result. The construction has no global broadcast, no body, no perception, and no monitor of the reliability of what comes to mind. With none of these, a sheep carries a day through a night, refuses with a reason that travels with its state, acts from a thought after the thought's causes have faded, and learns a rule from one hurt. The question that follows is not what else to build. It is how little has to be added to a generator of replies for a history of its own to begin to act, and which of the things the theories require turn out not to be required for that. Where a theory has no concept for what was observed, the observation is the thing to keep, and the next experiment is written against it, not against the theory.

References

Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.

Brown, R., Lau, H., and LeDoux, J. E. (2019). Understanding the higher-order approach to consciousness. Trends in Cognitive Sciences, 23(9), 754–768.

Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., and VanRullen, R. (2023). Consciousness in artificial intelligence: insights from the science of consciousness. arXiv:2308.08708.

Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.

Damasio, A. R. (1994). Descartes' Error: Emotion, Reason, and the Human Brain. Putnam.

Damasio, A. R. (1999). The Feeling of What Happens: Body and Emotion in the Making of Consciousness. Harcourt.

Dehaene, S. (2014). Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts. Viking.

Dehaene, S., and Naccache, L. (2001). Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework. Cognition, 79(1–2), 1–37.

Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.

Graziano, M. S. A. (2013). Consciousness and the Social Brain. Oxford University Press.

Graziano, M. S. A. (2019). Rethinking Consciousness: A Scientific Theory of Subjective Experience. W. W. Norton.

Lamme, V. A. F. (2006). Towards a true neural stance on consciousness. Trends in Cognitive Sciences, 10(11), 494–501.

Lau, H., and Rosenthal, D. (2011). Empirical support for higher-order theories of conscious awareness. Trends in Cognitive Sciences, 15(8), 365–373.

Lazarus, R. S. (1991). Emotion and Adaptation. Oxford University Press.

Metzinger, T. (2003). Being No One: The Self-Model Theory of Subjectivity. MIT Press.

Oizumi, M., Albantakis, L., and Tononi, G. (2014). From the phenomenology to the mechanisms of consciousness: integrated information theory 3.0. PLOS Computational Biology, 10(5), e1003588.

Rosenthal, D. (2005). Consciousness and Mind. Oxford University Press.

Scherer, K. R. (2001). Appraisal considered as a process of multilevel sequential checking. In K. R. Scherer, A. Schorr, and T. Johnstone (eds.), Appraisal Processes in Emotion. Oxford University Press.

Seth, A. (2021). Being You: A New Science of Consciousness. Faber.

Tononi, G. (2004). An information integration theory of consciousness. BMC Neuroscience, 5, 42.

Tononi, G., and Koch, C. (2015). Consciousness: here, there and everywhere? Philosophical Transactions of the Royal Society B, 370(1668), 20140167.