[ExI] My theory of consciousness

[Jason Resch]

On Mon, Sep 21, 2026 at 9:49 AM John Clark <johnkclark at gmail.com> wrote:

> On Mon, Sep 21, 2026 at 9:07 AM Jason Resch via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>
> > Computational functionalism treats cognition as implemented information
>>>> processing,
>>>>
>>>
>>> *Yes.*
>>>
>>> *> but leaves open why experience exists,*
>>>>
>>>
>>> *Yes.*
>>>
>>> > *and how it affects behavior.*
>>>>
>>>
>>> *No.*
>>>
>>
>> *> No that CTM leaves it open, or No that consciousness doesn't affect
>> behavior?*
>>
>
> *Computational functionalism can explain behavior while ignoring
> consciousness entirely.*
>

That everything can be explained in terms of third-person descriptions is
what gives the appearance that consciousness isn't a necessary part of the
process, and that is what leads to the idea of a hard problem. However, on
the understanding of consciousness my paper presents, consciousness is an
inseparable aspect of the sorts of systems that take in information,
discriminate upon it, and use those discriminative states to guide action. So,
while it is true you can ignore that consciousness is there, it would be
false to assume that you could replicate this intelligent reactive system
without consciousness (Dennett likens this to removing health from a body
while leaving all its bodily functions intact). The idea that you could
remove consciousness from a functionally equivalent system and change
nothing of its functioning, is what leads to the inherently
self-contradictory idea of philosophical zombies.


> * If that was not true then Darwinian Evolution would never have been able
> to produce a conscious being, and yet I know for a fact that it managed to
> do so at least once, and probably many billions of times. **Computational
> functionalism is certainly not claiming that consciousness doesn't exist,
> it's just saying it's not needed to explain behavior. *
>

I somewhat disagree here. If a robot responsively and reliably reacts to
catch a hockey puck as seen here:
https://www.youtube.com/watch?v=bOIxu2lv8sI if we seek an answer to the
question: how does this robot react in such a way to reliably intercept the
puck, there is no answer to that question that does not involve its
discriminative knowledge states its internal processing of visual
information produces within itself. Such discriminative knowledge states
used to guide action are (on my claim) conscious states. So a full account
of the behavior necessitates some reference to these states (though you can
try to obscure them away by restricting your account to the motions of
atoms, these states are still there).


>
> * >>> Consciousness is the inside of a difference that guides a process
>>>> from within. *
>>>
>>>
>>> *>> Consciousness is the way data feels when it is being processed
>>> intelligently.  *
>>>
>>
>> *> The way a process feels when it is informed by information.*
>>
>
> *That is circular. You are trying to explain consciousness but by using
> the word "informed" you are already assuming the existence of a conscious
> entity. **By contrast I am saying it's a brute fact that consciousness is
> the way that data feels when it is being processed intelligently, and that
> is NOT circular.  *
>

As Mike mentions below, the circularity above is in "feels" rather than
informed.
Most people don't assume any process that receives information and changes
its state is conscious, but that is essentially what my paper proposes. I
was hoping to show in my sentence was how close the adage you often repeat
(which comes from Chalmers and is repeated by Tegmark) comes to my own
view, but I was correcting what I see as an inherent problem in Chalmers's
framing of it, which is that information (absent the context of an
interpretive system/process) has no inherent meaning -- it can be
interpreted in an infinite number of ways. Information only acquires
meaning or "informative power" in the context of a receiving entity, and
the causal influence that information has on it. By viewing everything as
simply "raw information" he is trapped in third-person descriptions,
whereas the process view introduces inherent 1st-person / 3rd-person
distinctions. Two processes may share the same external information string,
but reach very different internal states by processing it, and moreoever
these states will be incommunicable and unsharable between the two
processes in ordinary circumstances. Thus they will come to reach states
that are private, and ineffable. That is why I corrected "information" to
"processes" above -- it is key to solving the problems Chalmers has been
stuck on for the past 30 years.



>
> *>>> We provide an empirical research program with explicit falsification
>>>> criteria. *
>>>
>>>
>>> *>> Tell me more about that! *
>>>
>>
>> *> See section 11.4 of the paper linked from the zenodo URL for details
>> on a series of empirical tests that could falsify the theory.*
>>
>
> *I am unable to access **that part, perhaps I could if I kept on trying
> but I don't think it would be worth the effort because **whatever you are
> suggesting I am quite certain it must be incorrect. Post the relevant parts
> here and prove me wrong. *
>

I am pasting the contents of that subsection here:


*11.4 Testing the centered-knowledge thesis*

The following proposals test the centered-knowledge thesis by varying
discriminative organization and recurrence, substitution boundaries, access
relations, judgment prove- nance, and phenomenal structure. Some outcomes
would revise a proposed access graph or substitution boundary rather than
the thesis itself. CKT would be challenged by a reproducible phenomenal
difference with the complete centered organization preserved, or by a
causally usable distinction within a center that could be shown to
contribute no phenomenal contrast. Together these proposals form an
empirical research program.

*The recurrence-necessity test*
Over any finite interval, a recurrent circuit can be unfolded into a
feed-forward circuit by replacing each return through the same component
with a new copy. The unfolded implementation preserves the same functional
states, causal dependencies, counterfactual transitions, and outputs; only
the physical reuse of components is removed (Doerig et al., 2019).
An autoregressive language model can generate 1,000 tokens either by
reusing one machine 1,000 times or by passing the complete inference state
through 1,000 machines with identical weights, each generating one token.
With identical decoding, both executions preserve every state transition
and produce the same sequence; reusing the hardware is economical, not
constitutive of the implemented computation.
The revealing test is substitution within a mind capable of comparison and
report. Replace a recurrent visual or other experience-related circuit in
an artificial mind—or eventually an uploaded human mind—with its
feed-forward unfolding while preserving every causal relation at its
interface. If the substitution changed color, shape, or any other
experience, the subject could notice, remember, compare, or report that
change only if some later transition also changed. Yet the unfolding
preserves those transitions. A theory that makes physical recurrence
constitutive must therefore posit a phenomenal change to which the subject
has no epistemic access. CKT predicts no change because the center
possesses and uses exactly the same distinctions before and after
substitution. Repeatedly switching between the two implementations turns
the case into a recurrence-specific version of dancing qualia (Chalmers,
1995a).
Modern large language models already weaken the motivation for treating
recurrence as the phenomenal threshold. Transformer architecture dispenses
with recurrent connections within each forward pass (Vaswani et al., 2017).
Autoregressive generation reapplies the model across successive tokens, so
an LLM is not feed-forward in every temporal sense; such iteration can
extend its reasoning across time. Nevertheless, systems built from these
feed-forward passes perform translation, programming, mathematical
reasoning, image interpretation, and many other intellectual tasks. Gurnee
and colleagues further find that a single pass contains a J-space
supporting verbal report, internal reasoning, flexible generalization, and
downstream control in a global-workspace-like organization (Gurnee et al.,
2026). This does not establish that LLMs are conscious. It shows that
recurrent neural circuitry is not required for rich discrimination,
workspace-like availability, or broad intellectual competence.
Large artificial networks therefore provide the practical starting point.
Recurrent components can be replaced with verified feed-forward unfoldings
while investigators trace whether every internal distinction and
report-producing path is preserved. Future digital minds with stable
memory, comparison, language, and self-report would sharpen the same
substitution challenge. The question is not whether their behavior would
diverge—functional equivalence entails that it cannot—but whether any
theory has grounds for positing a phenomenal divergence that the affected
mind could never know occurred.

*The neural-substitution test*
Make the prediction center-specific. Let vR ∈ V denote the
language-reporting machinery, treated as an anchoring component of a widely
distributed person-level center CR. A sufficiently detailed structural
connectome identifies possible causal paths into this coalition;
measurements of effective connectivity and direct interventions determine
which distinctions actually cross those paths and become available for
comparison, memory, report, and action. Tracing these dependencies
identifies both the other components participating in CR and the interfaces
through which they contribute to its knowledge state. Report anchors the
experiment but neither defines consciousness nor excludes local centers
elsewhere.
For a visual or auditory pathway, investigators would then locate the
coarsest interface whose exported distinctions, timing, and counterfactual
effects suffice to preserve KCR . This is CKT’s predicted substitution
level for that process. Machinery beneath the interface could then be
replaced with neural prostheses of progressively coarser grain. CKT
predicts that every replacement preserving the interface will leave the
participant’s perceptual comparisons and reports unchanged, despite
differences in its lower-level realization. Once coarsening omits a
constitutive relation, some distinction available to CR should change. The
predicted location of this transition makes the experiment a test of CKT
rather than of functionalism generally.
Within-subject cases provide the strongest comparison. A participant who
developed with sight or hearing and later became blind or deaf can compare
prosthetically restored states with previously occupied visual or auditory
states. Investigators could test color, pitch, timbre, location,
similarity, binding, memory, and use while progres- sively changing the
prosthesis’s implementation grain. A stable phenomenal difference produced
by a replacement below the predicted boundary would falsify that bound-
ary; if KCR were demonstrably preserved and no omitted relation could be
found, it would challenge CKT itself. Conversely, preserving every reported
distinction with a coarser replacement would show that the proposed map
included details that were not constitutive.

*The access-graph intervention test*
Anesthesia can be treated as a controlled deformation of the access graph
rather than a uniform reduction of neural activity. Cognitive-unbinding
theory proposes that anesthesia disrupts coordination among specialized
processes while local activity persists, and studies report weakened
directed connectivity during loss of responsiveness
(Ku et al., 2011; Boly et al., 2012; Mashour, 2013). CKT makes a sharper
prediction: as usable paths close, the normally integrated person-level
center may fragment before all local discrimination and use cease. Any
remaining subnetwork whose distinctions still guide its activity
constitutes a local center, even when its contents cannot reach language or
memory.
Test this prediction during graded induction and emergence by mapping
effective connectivity and determining which distinctions remain causally
usable. CKT predicts specific changes as particular paths close: if access
between sensory discrimination of pain and affective valuation is
disrupted, a subject may still locate and grade the pain without finding it
unpleasant (Rainville et al., 1997). If affective processing remains
locally active while isolated from language and contextual memory, it may
produce emotion on awakening without an accessible explanation. Under CKT,
this provides a plausible explanation of emergence distress, but the
prediction remains to be tested.
The Wada procedure supplies a reversible reporter subtraction. When the
language-dominant hemisphere is suppressed, the remaining active complement
can still discriminate stimuli and answer nonverbally (Loring et al.,
1992). One report describes every word as feeling on the tip of the tongue:
“It was amazing! I had no words”
(Anonymous, 2013). CKT predicts that removing verbal report leaves intact
any center whose distinctions continue to guide nonverbal activity.
A more targeted experiment would temporarily block communication through
the corpus callosum while leaving both hemispheres active. Zuboff imagines
such an intervention, with one hemisphere receiving a concert and the other
a lesson; once communication returns, the subject remembers both
experiences (Zuboff, 2025, pp. 12–13, 41–43). CKT predicts that during the
block, the two hemisphere-level coalitions constitute distinct centers if
each stimulus independently guides activity within its coalition. Immediate
hemisphere-specific responses would establish possession during the
separation; later integrated recall would show both preserved traces
becoming available to the reconstituted whole-brain center. Changing access
paths would thereby divide and recombine a person-level center without
eliminating the intervening perspectives.

*The paired-subtraction test*
Artificial systems permit interventions that would be impossible or
unethical in a brain. In a multimodal model, investigators can trace which
internal states causally affect report, memory, planning, and action, then
ablate or patch selected pathways, modality bridges, memory channels, or
representational directions. Gurnee and colleagues identify a small set of
verbalizable representations—the J-space—that supports directed modulation,
internal reasoning, and flexible downstream use within a transformer.
Swapping or suppressing those representations predictably changes the
model’s reasoning and output (Gurnee et al., 2026). This does not establish
consciousness, but it supplies an unusually explicit candidate access graph
on which CKT and rival theories can make different predictions.
Use this access graph in a paired subtraction test. In the first arm, leave
the reporting center intact but progressively remove all capacities and
contents not required for one target discrimination to enter and guide that
center. Intervene on the target state and verify that its alternatives
continue to produce accurate, counterfactually sensitive descriptions of
what is present in the reporter’s knowledge state. Because the reporter is
held constant, the experiment does not make the whole center minimal; it
isolates the target distinction as the only changing contribution to its
reported phenomenal content.
Then perform the complementary reporter subtraction. Close the path into
the reporting center, or remove the reporter, while preserving the target
discrimination and richer distinctions that continue to guide other local
activity. Causal tracing can verify that these states remain available to
the surviving processes even though they no longer generate verbal reports.
The first arm shows that a bare distinction can supply a reportable
phenomenal contrast without the removed enrichments; the second shows that
report is only one possible use of a distinction, not its constitutive
source. Together they support CKT’s claim that phenomenality begins with a
distinction possessed and used by a center, while report and further
capacities reveal or enrich that contrast.

*The phenomenal-judgment provenance test*
Argonov defines phenomenal judgments as observable speech or text about
con- sciousness and proposes testing whether a deterministic learning
system can originate them without preloaded philosophical knowledge,
philosophical instruction during learning, or models of other creatures
from which such knowledge could be inherited
(Argonov, 2014, pp. 53, 59–61). The system must not be designed to pass the
test, since that design would itself import knowledge about consciousness.
If a general-purpose learner nevertheless develops judgments about qualia,
unity, and other problematic properties, Argonov argues that the result
supports informational materialism and machine consciousness. A positive
result would therefore provide strong empirical evi- dence of
consciousness; a negative result would provide no corresponding evidence of
its absence.
The test derives its force from informational provenance. Once relevant
knowledge has been excluded from the system’s design and training, detailed
and coherent phenomenal judgments must arise from its own organization or
from systematic cognitive error. CKT predicts a distinctive family of such
judgments. A system may indepen- dently discover that describing a state is
not the same as occupying it, that another process’s qualitative state
cannot be determined from outside that process’s epistemic horizon, that
contents jointly available to one process form a unified perspective, and
that inaccessible contents can be known only inferentially. If these
judgments arise without instruction and change appropriately when the
underlying distinctions and access relations are altered, an error account
must explain why the supposed errors track the system’s actual
organization. CKT provides the direct explanation: the process is
recognizing and describing the causal-access relations that constitute its
perspective.
After a system passes the test, investigators can reproduce its
judgment-producing organization while varying the mechanisms that rival
theories treat as necessary. First, implement the same discrimination,
reasoning, and report in a genuinely feed-forward architecture with no
recurrent cause–effect loops. If it continues to pass, the result would
challenge recurrence-necessity theories and, given the evidential premise
of the test, falsify IIT’s denial of system-level consciousness to a purely
feed-forward whole
(Albantakis et al., 2023). Second, embed an independently functioning
center within a much larger architecture while preventing its
judgment-producing contents from being broadcast beyond that center and its
local reporter. If it still passes, the result would falsify the claim that
brain-wide or system-wide broadcast is necessary for consciousness.
The report remains intact in both experiments; recurrence and broadcast are
the variables being removed. CKT predicts that both implementations remain
conscious because the relevant distinctions continue to be possessed and
used within the judgment- producing center. The test thereby converts
independently originated phenomenal judgments from generic evidence for
computationalism into competing predictions among theories of consciousness.
One limitation is that a positive Argonov result cannot distinguish CKT
from the actualist HOT view compared here. Any system capable of the
relevant judgment must possess a first-order distinction, and the judgment
itself supplies a higher-order representation that HOT can treat as
consciousness-conferring. CKT attributes the experience to the possessed
and used first-order distinction, whereas HOT attributes it only once the
higher-order representation occurs; the positive result contains both.
Discriminating the theories therefore requires preserving the first-order
organization while removing or delaying the higher-order state, as in the
paired-subtraction and toggle cases.

*The sensory-expansion and phenomenal-structure test*
Add a controlled sensory distinction and measure how it enters
discrimination, similarity, memory, valuation, and action. Mancuso and
colleagues added a third cone opsin to adult red–green color-deficient
monkeys. The animals subsequently acquired trichromatic
color-discrimination behavior, showing that a new retinal distinction could
become usable by the adult visual system without intervention during early
development (Mancuso et al., 2009). The study cannot reveal whether the
monkeys experienced merely new shades of familiar colors or colors unlike
anything in their previous experience.
A comparable intervention in a color-deficient person would permit
within-subject phenomenal comparison. Investigators could ask whether the
newly discriminable states appear as finer gradations within the subject’s
former color space or as colors unlike any previously experienced. Such
reports would be evidence rather than proof: Locke’s violet–marigold case
shows that qualitative vocabulary does not uniquely determine qualitative
identity (Locke, 1689/1975).
Stronger evidence would combine these reports with psychophysics and neural
mapping. Color-matching and similarity tasks could determine whether the
added input merely increases resolution within the former perceptual space
or supplies an independently varying chromatic dimension. Neural
measurements could then test whether a corresponding dimension appears in
the activity available to and used by the center. Comparisons among
dichromats, trichromats, and functional tetrachromats would be informative,
although a before-and-after intervention in the same person would provide
the clearest test. CKT predicts new shades when an intervention refines
existing relations, but genuinely novel phenomenal structure when it adds
an independent distinction to the center’s usable perceptual organization.
A new receptor signal that remains outside that organization should produce
neither result.



> *If it's really as good as you claim then you will have made the largest
> philosophical advance in a 1000 years, if not 10,000 years. *
>

Thanks, I do consider it an advance, but I am not sure it is as great as
you say.

Jason
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20260921/3b045af1/attachment.htm>