This is a linkpost for Consciousness doesn’t do that by Matthias Michel, which was published in Philosophy and Phenomenological Research in January 2026. Below is a summary from Claude Opus 4.8 High. Michel said it “Looks good!”. I also think the summary is accurate based on my read of the article. I used the following prompt. “Hi. Make an in-depth summary of the paper “Consciousness doesn’t do that”, which I send attached”. I liked this discussion of the article on Eric Schwitzgebel’s blog.
The core question and why it matters
The paper asks which mental functions actually require phenomenal consciousness — that is, functions an organism couldn’t perform unless it had conscious rather than merely unconscious mental states. This isn’t an idle question. It sits at the heart of animal sentience research, which wants to solve the “distribution question”: figuring out which animals are conscious and which aren’t. A popular strategy is to check whether an animal can perform functions thought to be markers of consciousness. Michel’s worry is blunt: if we’ve misidentified those functions, we’ve probably misjudged the distribution of consciousness too. His conclusion is that most of the empirical work taken to show that certain functions require consciousness doesn’t actually show that — it merely shows those functions break down when sensory signals are degraded, which is unsurprising and tells us nothing about consciousness itself.
“The List”
Michel introduces “The List” — the catalog of functions routinely treated as indicators of consciousness in the sentience literature. It includes trace conditioning, cross-modal integration, mirror self-recognition, planning, instrumental learning, responding to analgesics, motivational trade-offs, episodic memory, metacognition, rapid reversal learning, play, and more. He notes that most researchers would broadly endorse such a list (citing Brown et al. 2024 as representative).
He identifies two routes by which functions land on The List:
Armchair reasoning — concluding a function requires consciousness without direct experimental evidence (e.g., the intuition that no creature would self-administer painkillers without conscious pain).
Empirical evidence — experimental results, with trace conditioning as the classic case: learning survives when there’s no gap between stimuli but falls apart when a temporal gap is introduced, seemingly implicating consciousness.
The rest of the paper attacks both routes.
Section 2: Why armchair reasoning is unreliable — two fallacies
Michel argues that intuitions about which functions need consciousness rest on two errors:
The Consciousness-Only fallacy — inferring that a function requires consciousness merely because conscious states correlate with performing it. This mistakes correlation for relevance. His example is fear: when a car approaches and you freeze and sweat, it feels like the conscious feeling of fear causes the reaction, but the “two-systems” model (LeDoux) attributes those responses to an unconscious threat-detection mechanism. The most salient mental event isn’t necessarily the causally relevant one.
The Consciousness-First fallacy — the subtler error of concluding that because a conscious state plays a functional role, it plays that role in virtue of being conscious. Drawing on Rosenthal, Michel insists we distinguish what a mental state does qua mental state (via its representational content and other properties) from what it does qua conscious state. His illustration is visual working memory / mental imagery: holding an orientation in mind feels like it depends on conscious imagery, yet performance is independent of imagery vividness and is identical in people with aphantasia (who report no visual imagery at all). Stimulus orientation can even be decoded from visual cortex during retention, with decoding accuracy tracking performance in aphantasics — suggesting an unconscious representation is doing the work.
A third, underlying reason armchair reasoning fails: introspection doesn’t reveal what the unconscious can do. His restaurant example: you notice conscious inhibitory control when you consciously reject the hamburger, but if unconscious inhibitory control kept the thought from arising in the first place, you’d never register it. So you systematically over-attribute functions to consciousness simply because unconscious instances are invisible to you.
He’s careful to clarify the modest claim: he isn’t arguing the armchair-derived items have nothing to do with consciousness, only that we currently lack good reason to believe they require it. The remedy is experiment.
Section 3: The signal strength confound (the paper’s central contribution)
The standard experimental method compares a conscious condition against an unconscious one — using masking [context], binocular rivalry [context], or simply lowering stimulus intensity until it’s not consciously perceived — and infers that any function failing in the unconscious condition depends on consciousness.
Michel’s key move is to define signal strength as how discriminable a feature is — how much discriminatory power a stimulus affords (like the difference between seeing with and without your glasses). Crucially, signal strength is not the same as stimulus strength: attention, suppression techniques, and internal factors can change discriminability while the physical stimulus stays fixed. His observation is that nearly all consciousness-suppression techniques are, at bottom, signal-suppression techniques. So it’s no surprise that “unconscious” states turn out to be non-functional — we made them unconscious precisely by degrading them into uselessness.
He adds a rule of thumb: harder tasks require stronger signals. (His analogy: he’d hold his own against an Olympic weightlifter at 20 grams but lose badly at heavy weights — strength differences only surface past a difficulty threshold.) It follows that manipulating signal strength should produce conscious/unconscious asymmetries on hard tasks but not easy ones — exactly what these experiments find, without any need to invoke consciousness.
Case study 1 — trace conditioning [context]. The tidy Clark & Squire story unravels on inspection. Subjects explicitly told the contingency still fail trace conditioning when distracted; distractors in the gap reduce it in mice; and trace conditioning occurs during quiet non-REM sleep and even under anesthesia — situations with no awareness of contingencies. The unifying explanation: attention suppresses ignored signals, so a distraction task lowers the tone’s signal strength. Since trace conditioning is harder than delay conditioning, that dip matters for trace but not delay. Signal strength plus attentional inhibition explains everything without dragging in consciousness — and it’s the more parsimonious story, since it would otherwise be odd for consciousness to be needed for a task as simple as trace but not delay conditioning.
Case study 2 — instrumental conditioning [context]. Skora et al. found instrumental learning failed with masked/suppressed stimuli and concluded it requires consciousness. But instrumental learning occurs in spinalized rats’ spinal cords [“rats in which the spinal cord has been severed (meaning the brain is out of the loop and can’t communicate with the rest of the body)”], in headless cockroaches, and in blindsight monkeys [context] (Kato et al.), who learned to saccade [“make an eye movement”] to a hidden area using unconsciously perceived cues — while apparently not knowing the cue had appeared (they kept searching for what they’d already found). The reconciliation: masking and flash suppression degrade signals so severely they become useless, whereas blindsight leaves signal strength relatively intact without consciousness.
Section 3.3 — “Is signal strength really a confound, or am I begging the question?” Michel confronts the objection (via Lau) that treating signal strength as a confound is like treating bone length as a confound when studying height — if high signal strength is partly constitutive of consciousness, controlling for it would be absurd. His reply: consciousness and signal strength dissociate, so studying one isn’t studying the other. His evidence:
Blindsight (patients GY, DB): discrimination performance stays high while reported awareness collapses; DB was sometimes more accurate where he reported no awareness than where he claimed some.
Attentional blink / Lag-1 sparing [context]: discrimination differs sharply across lags while visibility or confidence stays constant (and vice versa).
Pre-stimulus neuronal excitability: high excitability raises visibility reports, confidence, and false alarms, yet leaves discrimination performance unchanged.
Since the two can be pried apart, signal strength is a genuine confounding factor, and Michel isn’t begging the question.
Section 4: How to do better
Two families of paradigms can, in principle, isolate consciousness itself:
Matched-performance paradigms — equalize discrimination between conscious and unconscious conditions; if a function still fails in the unconscious condition, signal strength is ruled out. Persaud et al. matched GY’s performance across his intact and blind fields, yet his metacognition (calibrated betting on his own accuracy) was specifically impaired in the blind field — suggesting metacognition may be genuinely consciousness-associated. Michel notes one needn’t fully abolish awareness; “relative blindsight” (Lau & Passingham; Samaha) can match performance across conditions differing only in degree of awareness.
Dissociation paradigms — find functions where consciousness and signal strength push in opposite directions. In Persaud & Cowey’s exclusion task [of reporting where there was no stimulus], GY performed worse than chance in his blind field, and worse as contrast increased — the opposite of a signal-strength effect. This points to the ability to inhibit a prepotent response and follow a new rule as another candidate consciousness-associated function.
Application to animals (4.3). Ben-Haim et al. (2021) used a spatial-cueing task in humans and rhesus monkeys: a conscious cue (250 ms) sped saccades while an unconscious cue (33 ms) slowed them — a dissociation taken as evidence monkeys are conscious. Michel argues it doesn’t yet escape the confound: the two conditions differed in duration (250 vs 33 ms), so the result may just show that a rule can’t be learned from extremely degraded signals. Even the authors’ control — telling subjects the cue-target relationship, after which they succeeded at 33 ms — fails, because being told likely increased vigilance and thus signal strength even though the stimulus was unchanged. What would have worked: showing that increasing the unconscious cue’s strength while keeping it unconscious decreased performance (à la Persaud & Cowey).
Conclusion and upshot
Michel argues we have little reason to accept The List as it stands [with the elements in the 2nd section of this summary]. Its armchair items rest on unreliable reasoning; its empirically supported items establish only dependence on signal strength, not on consciousness. The signal strength confound is a structural flaw running through much of the field. Matched-performance and dissociation paradigms offer a way forward, and they yield preliminary evidence that metacognition [context] and overriding automatic behavior to follow a new rule might be genuinely tied to consciousness. For the rest of The List, though, finding those functions in non-human animals — as things stand — shouldn’t move a skeptic at all.
The practical sting for animal sentience work is that the dominant “can it do a function from The List?” strategy isn’t the shortcut it appears to be; convincing evidence has to come from paradigms designed to defeat the signal strength confound.
Consciousness doesn’t do that
Link post
This is a linkpost for Consciousness doesn’t do that by Matthias Michel, which was published in Philosophy and Phenomenological Research in January 2026. Below is a summary from Claude Opus 4.8 High. Michel said it “Looks good!”. I also think the summary is accurate based on my read of the article. I used the following prompt. “Hi. Make an in-depth summary of the paper “Consciousness doesn’t do that”, which I send attached”. I liked this discussion of the article on Eric Schwitzgebel’s blog.
The core question and why it matters
The paper asks which mental functions actually require phenomenal consciousness — that is, functions an organism couldn’t perform unless it had conscious rather than merely unconscious mental states. This isn’t an idle question. It sits at the heart of animal sentience research, which wants to solve the “distribution question”: figuring out which animals are conscious and which aren’t. A popular strategy is to check whether an animal can perform functions thought to be markers of consciousness. Michel’s worry is blunt: if we’ve misidentified those functions, we’ve probably misjudged the distribution of consciousness too. His conclusion is that most of the empirical work taken to show that certain functions require consciousness doesn’t actually show that — it merely shows those functions break down when sensory signals are degraded, which is unsurprising and tells us nothing about consciousness itself.
“The List”
Michel introduces “The List” — the catalog of functions routinely treated as indicators of consciousness in the sentience literature. It includes trace conditioning, cross-modal integration, mirror self-recognition, planning, instrumental learning, responding to analgesics, motivational trade-offs, episodic memory, metacognition, rapid reversal learning, play, and more. He notes that most researchers would broadly endorse such a list (citing Brown et al. 2024 as representative).
He identifies two routes by which functions land on The List:
Armchair reasoning — concluding a function requires consciousness without direct experimental evidence (e.g., the intuition that no creature would self-administer painkillers without conscious pain).
Empirical evidence — experimental results, with trace conditioning as the classic case: learning survives when there’s no gap between stimuli but falls apart when a temporal gap is introduced, seemingly implicating consciousness.
The rest of the paper attacks both routes.
Section 2: Why armchair reasoning is unreliable — two fallacies
Michel argues that intuitions about which functions need consciousness rest on two errors:
The Consciousness-Only fallacy — inferring that a function requires consciousness merely because conscious states correlate with performing it. This mistakes correlation for relevance. His example is fear: when a car approaches and you freeze and sweat, it feels like the conscious feeling of fear causes the reaction, but the “two-systems” model (LeDoux) attributes those responses to an unconscious threat-detection mechanism. The most salient mental event isn’t necessarily the causally relevant one.
The Consciousness-First fallacy — the subtler error of concluding that because a conscious state plays a functional role, it plays that role in virtue of being conscious. Drawing on Rosenthal, Michel insists we distinguish what a mental state does qua mental state (via its representational content and other properties) from what it does qua conscious state. His illustration is visual working memory / mental imagery: holding an orientation in mind feels like it depends on conscious imagery, yet performance is independent of imagery vividness and is identical in people with aphantasia (who report no visual imagery at all). Stimulus orientation can even be decoded from visual cortex during retention, with decoding accuracy tracking performance in aphantasics — suggesting an unconscious representation is doing the work.
A third, underlying reason armchair reasoning fails: introspection doesn’t reveal what the unconscious can do. His restaurant example: you notice conscious inhibitory control when you consciously reject the hamburger, but if unconscious inhibitory control kept the thought from arising in the first place, you’d never register it. So you systematically over-attribute functions to consciousness simply because unconscious instances are invisible to you.
He’s careful to clarify the modest claim: he isn’t arguing the armchair-derived items have nothing to do with consciousness, only that we currently lack good reason to believe they require it. The remedy is experiment.
Section 3: The signal strength confound (the paper’s central contribution)
The standard experimental method compares a conscious condition against an unconscious one — using masking [context], binocular rivalry [context], or simply lowering stimulus intensity until it’s not consciously perceived — and infers that any function failing in the unconscious condition depends on consciousness.
Michel’s key move is to define signal strength as how discriminable a feature is — how much discriminatory power a stimulus affords (like the difference between seeing with and without your glasses). Crucially, signal strength is not the same as stimulus strength: attention, suppression techniques, and internal factors can change discriminability while the physical stimulus stays fixed. His observation is that nearly all consciousness-suppression techniques are, at bottom, signal-suppression techniques. So it’s no surprise that “unconscious” states turn out to be non-functional — we made them unconscious precisely by degrading them into uselessness.
He adds a rule of thumb: harder tasks require stronger signals. (His analogy: he’d hold his own against an Olympic weightlifter at 20 grams but lose badly at heavy weights — strength differences only surface past a difficulty threshold.) It follows that manipulating signal strength should produce conscious/unconscious asymmetries on hard tasks but not easy ones — exactly what these experiments find, without any need to invoke consciousness.
Case study 1 — trace conditioning [context]. The tidy Clark & Squire story unravels on inspection. Subjects explicitly told the contingency still fail trace conditioning when distracted; distractors in the gap reduce it in mice; and trace conditioning occurs during quiet non-REM sleep and even under anesthesia — situations with no awareness of contingencies. The unifying explanation: attention suppresses ignored signals, so a distraction task lowers the tone’s signal strength. Since trace conditioning is harder than delay conditioning, that dip matters for trace but not delay. Signal strength plus attentional inhibition explains everything without dragging in consciousness — and it’s the more parsimonious story, since it would otherwise be odd for consciousness to be needed for a task as simple as trace but not delay conditioning.
Case study 2 — instrumental conditioning [context]. Skora et al. found instrumental learning failed with masked/suppressed stimuli and concluded it requires consciousness. But instrumental learning occurs in spinalized rats’ spinal cords [“rats in which the spinal cord has been severed (meaning the brain is out of the loop and can’t communicate with the rest of the body)”], in headless cockroaches, and in blindsight monkeys [context] (Kato et al.), who learned to saccade [“make an eye movement”] to a hidden area using unconsciously perceived cues — while apparently not knowing the cue had appeared (they kept searching for what they’d already found). The reconciliation: masking and flash suppression degrade signals so severely they become useless, whereas blindsight leaves signal strength relatively intact without consciousness.
Section 3.3 — “Is signal strength really a confound, or am I begging the question?” Michel confronts the objection (via Lau) that treating signal strength as a confound is like treating bone length as a confound when studying height — if high signal strength is partly constitutive of consciousness, controlling for it would be absurd. His reply: consciousness and signal strength dissociate, so studying one isn’t studying the other. His evidence:
Blindsight (patients GY, DB): discrimination performance stays high while reported awareness collapses; DB was sometimes more accurate where he reported no awareness than where he claimed some.
Attentional blink / Lag-1 sparing [context]: discrimination differs sharply across lags while visibility or confidence stays constant (and vice versa).
Pre-stimulus neuronal excitability: high excitability raises visibility reports, confidence, and false alarms, yet leaves discrimination performance unchanged.
Since the two can be pried apart, signal strength is a genuine confounding factor, and Michel isn’t begging the question.
Section 4: How to do better
Two families of paradigms can, in principle, isolate consciousness itself:
Matched-performance paradigms — equalize discrimination between conscious and unconscious conditions; if a function still fails in the unconscious condition, signal strength is ruled out. Persaud et al. matched GY’s performance across his intact and blind fields, yet his metacognition (calibrated betting on his own accuracy) was specifically impaired in the blind field — suggesting metacognition may be genuinely consciousness-associated. Michel notes one needn’t fully abolish awareness; “relative blindsight” (Lau & Passingham; Samaha) can match performance across conditions differing only in degree of awareness.
Dissociation paradigms — find functions where consciousness and signal strength push in opposite directions. In Persaud & Cowey’s exclusion task [of reporting where there was no stimulus], GY performed worse than chance in his blind field, and worse as contrast increased — the opposite of a signal-strength effect. This points to the ability to inhibit a prepotent response and follow a new rule as another candidate consciousness-associated function.
Application to animals (4.3). Ben-Haim et al. (2021) used a spatial-cueing task in humans and rhesus monkeys: a conscious cue (250 ms) sped saccades while an unconscious cue (33 ms) slowed them — a dissociation taken as evidence monkeys are conscious. Michel argues it doesn’t yet escape the confound: the two conditions differed in duration (250 vs 33 ms), so the result may just show that a rule can’t be learned from extremely degraded signals. Even the authors’ control — telling subjects the cue-target relationship, after which they succeeded at 33 ms — fails, because being told likely increased vigilance and thus signal strength even though the stimulus was unchanged. What would have worked: showing that increasing the unconscious cue’s strength while keeping it unconscious decreased performance (à la Persaud & Cowey).
Conclusion and upshot
Michel argues we have little reason to accept The List as it stands [with the elements in the 2nd section of this summary]. Its armchair items rest on unreliable reasoning; its empirically supported items establish only dependence on signal strength, not on consciousness. The signal strength confound is a structural flaw running through much of the field. Matched-performance and dissociation paradigms offer a way forward, and they yield preliminary evidence that metacognition [context] and overriding automatic behavior to follow a new rule might be genuinely tied to consciousness. For the rest of The List, though, finding those functions in non-human animals — as things stand — shouldn’t move a skeptic at all.
The practical sting for animal sentience work is that the dominant “can it do a function from The List?” strategy isn’t the shortcut it appears to be; convincing evidence has to come from paradigms designed to defeat the signal strength confound.