Here's a deceptively simple argument that derives an empirically falsifiable conclusion from two uncontroversial premises. No logical leaps. No metaphysics or philosophy. Just premises, deduction, and a clear way to falsify.
I'll present the argument first, then defend each piece in turn. The full formal treatment is in the paper linked at the end.
By “manifest consciousness,” I mean those aspects of experience that can, in principle, make a difference to behavior, including self-report. Non-manifest aspects of consciousness are empirically unreachable, and their existence doesn't undermine the manifest case.
To avoid this conclusion, one must either reject Premise 1 (epiphenomenalism, handled below), or reject Premise 2, which can be falsified by demonstrating a way to alter intentional behavior without altering spike patterns.
Note: this argument relies heavily on self-reports. Assume the reports come from reasonably lucid, unimpaired, earnest subjects. The logic doesn’t require all subjects to fit that description, only that such subjects can exist in principle.
We treat self-reports as translations of experience. This is the gold standard across multiple scientific fields:
Even when we develop objective measures (e.g. EEG, fMRI), the subject's report is treated as ground truth. If a bright-eyed subject reports feeling awake and alert, while the machine says they're unconscious, we question the machine or the theory, not whether the person is actually conscious. For our purposes, we don't need self-reports to be perfectly accurate; we just need them reliable enough that entire scientific fields can be built on the data they provide.
We also do this in daily life:
When we communicate about felt states, we act as if the communication reflects the inner state better than random noise.
There is no consciousness detector to prove the flow of causation from experience to behaviour, so we must use evidence and causal/interventionist logic to make epiphenomenalism epistemically untenable.
First, we must establish experience as being somewhere in the causal chain. Our behaviour - specifically self-report - can function as a reliable translation of our experience (within the limits of language). Without both experience and behaviour sharing the same causal graph, that universal covariation would be just perpetual inexplicable coincidence, i.e. unscientific.
We'll keep this simple (formal treatment in Section 3 of the paper), but I think it's more legible to give ourselves a few symbols to work with:
This leaves us with only two reasonable options. Either:
Our premise 1 is that E causes U, so we will focus on the common cause hypothesis:
First let us define one last symbol (I promise):
This reporting policy might be a very coarse:
Or a more detailed:
Or it can even be a convoluted:
The fact that U is reliably a translation of E through any reporting policy K starts to make the common cause view a little shaky. If E is causally idle, then it should function like an exhaust fume/side effect of common cause Z, while the main purpose is to drive behaviour. The fact we can perform any intervention K and have U maintain the correct mapping to E is difficult to reconcile for a common cause framework.
The only reasonable move from there is to invoke a common cause Z rich enough to fully map experience to behavior over any K. However, also contained in E is the felt sense of translating experience into report; the experiential "what-it's-like-ness" of that translation process and its success. This means that Z must also contain it in order to feed it to both E and U.
This sort of "intentional" illusion is difficult to justify through any evolutionary argument where E can have no effect on behaviour. Set that aside, and we're still left with a Z that has enough information to fully define the shape and character of E, as well as the translation step from E to U. this leaves the epiphenomenalist one of two moves:
Option A accepts our conclusion.
Option B is an inexplicable perpetual coincidence.
Sherrington's "final common path" has been battle-tested as motor neuroscience 101 for over a century. It states that all movement (behaviour) must ultimately pass through lower motor neurons. It is treated as essentially fact among neuroscientists. No reproducible example has ever been documented of a behavior-changing manipulation that leaves the relevant spike pattern intact. PNM remains unfalsified.
With PCE we have established that consciousness can have some causal influence on behaviour, and with PNM, that the path to behaviour always eventually passes through neuron spike patterns. The only remaining move is to eliminate anything upstream of neuron spikes from being necessary for conscious experience. We won't need to go into detail about each, but for the neuroscience people, we're referring to glia, ephaptic fields, hypothesized quantum microtubules, etc, that have any ability (hypothetical or otherwise) through any route to eventually help resolve whether a neuron spikes or not.
The way we eliminate these is by screening them off causally. Spikes occupy a unique place in the brain as causal influences to behaviour for a few reasons. They are the only mechanism that contains (all in one package) the specificity, speed, long-range transmission, and density to encode complex stimuli in the way we experience and express it. But more importantly, every other factor eventually resolves to either a neuron spiking, or not spiking. If it has no causal effect on a neuron spike (or non-spike), then it is behaviorally idle, violating PCE. If it does affect spikes, then it's causally degenerate - multiple configurations of upstream factors can produce the same spike outcome. This means that upstream factors have no mechanism to distinguish their contribution through behaviour. Multiple paths lead to any given spiking outcome, but if consciousness cared which route you took, it would have no way to tell you (violating PCE).
Therefore, everything required for consciousness is encoded in neuron spiking patterns. To falsify this, show any manipulation that alters intentional behavior without altering spike patterns.
Interestingly, "neuron spiking patterns" can be defined very loosely; enough to establish substrate independence. If you replace any given one or more neurons (up to the entire brain) with any replacement, natural or artificial, and that replacement has the same downstream effects for any given set of upstream inputs, then you will replicate behaviour, including self-reporting behaviour where consciousness (per PCE) was part of the causal chain. This also holds for what I call "strong substrate independence", though I'm aware it's been called by other names. Essentially, the replacement "neuron" or node need not be a discrete physical object at all. If two or more functionally equivalent neurons (up to the entire brain) were implemented in software, and run on hardware that was connected to the same inputs and outputs, the same exact consciousness-dependent behaviour would result.
The full formal treatment is in the paper: https://zenodo.org/records/17851367