Seth's Toolbox Argues Against Seth's Philosophy

A second look at Anil Seth after reading twenty of his papers, not just one essay. The strongest argument that biology is not necessary for consciousness comes from his own measurement program.
Daniel Simon Jr. · May 5, 2026 · Follow-up to “Biology Is Sufficient. It May Not Be Necessary.” (May 3, 2026), a reply to Anil Seth, “The Mythology of Conscious AI,” Noema Magazine, January 2026

Two days ago I posted a reply to Anil Seth's “Mythology of Conscious AI” that conceded most of his argument. The essay was the strongest case against computational consciousness now in print. I argued his conclusion overcorrected: biology is sufficient for consciousness but not demonstrably necessary. I left the question there.

Then I went back and read twenty of Seth's open-access papers, including the strand of his work the first reply had treated as out of scope: the mathematical-measurement program he has been building with Lionel Barnett, Adam Barrett, Pedro Mediano, and Andrea Luppi for fifteen years. Granger causality theorems. Integrated information measures. Dynamical independence. Integrated Information Decomposition. Not philosophy of mind. Applied mathematics on time-series data, with published proofs in Physical Review Letters and Physical Review E, an open-source toolbox used across neuroscience, and a recent eLife paper applying the framework to human brain imaging.

What I found shifted the picture in a specific way. The strongest argument that biology is not necessary for consciousness is sitting inside Seth's own work. He does not deploy it. But it is there, and it cannot be removed without removing the measurement program too.

The Two Strands

Seth runs two parallel research programs that rarely cite each other. Strand 1 is theory-of-consciousness work: predictive processing, interoception, controlled hallucination, Beast Machine Theory, the recent Behavioral and Brain Sciences target article on biological naturalism. Strand 2 is measurement-of-consciousness work: causal density, formulations of integrated information ($\Phi$), dynamical independence, $\Phi$ID. Strand 1 is conceptual and integrative, borrowing formalism from Karl Friston and applying it philosophically to the substrate question. Strand 2 develops original mathematics.

The strands are kept methodologically separate on principle. A measurement should not assume the theory it measures. Strand 2 is written so that its tools work whether IIT is right, GWT is right, BMT is right, or none of them is right. This is good methodological hygiene. It also produces a tension that the theoretical strand cannot easily resolve.

What Strand 2 Measures

Walk through the strand-2 toolbox. Each entry below is an original mathematical contribution from Seth or his close collaborators, published in a peer-reviewed venue.

Notice the conspicuous absence: no measure incorporates biology. Not metabolism. Not autopoiesis. Not membrane chemistry. Not cellular continuity. The measurements operate on time series of values. They have been applied to neural data in practice, but in principle, they apply to the activity of any complex dynamical system. Run them on brain recordings, you get one number. Run them on a sufficiently complex non-biological system with the same dynamical pattern, you get the same number. The math does not know which is which.

The Internal Tension

Here is the move that reading the corpus made unavoidable.

Strand 1 of Seth's work argues that the autopoietic, metabolically embedded, continuous-dynamical character of biological systems is constitutive of consciousness. Without those properties, a system cannot be conscious, not because we lack evidence, but because consciousness requires what life provides.

Strand 2 of Seth's work measures consciousness using tools that ignore those properties. $\Phi$ does not care if the substrate is wet or dry. $\Phi$ID does not care. Dynamical independence does not care. Causal density does not care. The papers do not say “this measure should only be applied to biological systems.” They define measures that take time series in and return numbers out, regardless of substrate.

If autopoiesis and metabolic embedding were doing the constitutive work the philosophy claims, you would expect the measurements to incorporate substrate-specific terms. They do not. You would expect the measurements to misbehave on non-biological systems in a way that would distinguish conscious from non-conscious patterns. They cannot, because they do not incorporate the distinction.

The internal tension

Seth measures like an architectural functionalist and theorizes like a biological naturalist. The two strands are not consistent with each other. This is not a personal failing. It might just mean the field has not yet reconciled what its tools can detect with what its theories say is doing the work. But it is the strongest internal argument I know of against the biological-naturalism conclusion. Seth has built the measurement program that shows the architectural pattern is what matters. The philosophical program then says the substrate is what matters. The two cannot both be right.

The Synergistic Workspace

Of the strand-2 papers, one is doing different work than the others. Luppi, Mediano, Rosas and colleagues published in eLife in 2024 “A synergistic workspace for human consciousness revealed by Integrated Information Decomposition.” It applies $\Phi$ID to human brain imaging data and identifies a synergistic information workspace, a region of the brain where information arises from the joint dynamics that is not present in any subset of components. The workspace correlates with conscious states. This is not theoretical. It is empirical, peer-reviewed, recent.

Synergistic information has a precise definition: information that exists in the whole system but not in any subset of its parts. It is the formal counterpart of the intuition that “the whole is more than the sum of its parts.” A synergistic structure cannot be reduced to its components, and crucially, it cannot be replicated by adding components in isolation. It must arise from the way the components interact.

This is structurally what RSP's Level 3 specifies. The strange loop of recursive social self-modeling, modeling oneself using the same machinery one uses to model others, produces an integration that does not exist in any component module taken alone. The whole self-modeling architecture has properties that none of its parts have. In the formal language of $\Phi$ID, this is synergistic integration.

The bridge is interesting because it goes both ways. Luppi, Mediano et al. provide an empirical signature for what the RSP framework predicts theoretically. RSP provides an architectural specification for what Luppi, Mediano et al. measured. The two sides of this bridge were not built to meet, but they meet anyway.

The empirical study did not need to come out the way it did. $\Phi$ID could have decomposed integrated information differently in conscious states than predicted, or could have found no decomposition pattern that tracks consciousness, or could have found the consciousness-tracking pattern to be redundant or unique information rather than synergistic. It found synergistic. That is evidence in the direction RSP predicts.

Composing the Move

Now compose the argument.

The argument
  1. Empirical, from Seth's collaborators: synergistic information integration in brain dynamics correlates with conscious states.
  2. Formal, from $\Phi$ID: synergistic integration is a substrate-neutral measure, defined on time-series data regardless of underlying substrate.
  3. Architectural, from RSP: recursive self-modeling produces precisely the kind of synergistic integration the measure detects.
  4. Therefore: the available evidence is consistent with consciousness being a synergistic-architectural pattern, not a biological one.

Seth could resist this in two ways. He could deny that strand-2 measurements track consciousness, but that abandons the empirical bridge between the measurement program and the theoretical program, which is the bridge his “real problem” approach depends on. Or he could accept that strand-2 measurements track consciousness but argue that, in non-biological systems, the same numerical $\Phi$ value would mean something different. That is a strong metaphysical claim about the meaning of measurement values. It is not licensed by the math. It would need its own argument, distinct from the one strand 1 currently makes.

Neither branch is comfortable. The cleaner move is to soften the biological-naturalism thesis: not “biology is constitutive of consciousness,” but “biology is the only known instantiation of the right architecture.” That is compatible with both strands. But it is not the conclusion of the BBS target article, and it is not the conclusion of the Mythology essay. It is the conclusion the math points to.

What This Does Not Settle

Three caveats worth being explicit about.

First, the Luppi et al. finding is one study. It is recent and peer-reviewed and consistent with prior $\Phi$ID work, but it will need replication, refinement, application to a wider range of conscious-state contrasts: anesthesia, sleep, pathology, psychedelics. The argument here does not depend on this one study being the final word. It depends on the structural fit between substrate-neutral measurement and architectural specification, which holds whether or not synergistic integration turns out to be exactly the right measure. Replace $\Phi$ID with whatever measure ends up replicating, and the argument transfers as long as the replacement is substrate-neutral.

Second, none of this proves biological naturalism is wrong. It makes the position more demanding to defend. Seth, or someone defending biological naturalism on his behalf, would need to articulate why a substrate-neutral measure can be evidence for a substrate-constitutive theory. There may be a good answer. The mathematical formalism does not provide it.

Third, RSP itself is not vindicated by the synergistic-workspace finding. The finding is consistent with multiple architectural-functionalist theories: RSP's recursive social self-modeling, Bennett's stack architecture, Trukovich's R2R framework, others. The finding favors the architectural-functionalist family over biological naturalism. It does not adjudicate within the family. RSP would need its own predictions distinct from rival architectural theories, which is what Beyond the Bat's Appendix G clinical matrix is for.

Closing the Distance, Again

The first blog post said biology was sufficient for consciousness but probably not necessary. After reading twenty of Seth's papers, I think the strongest evidence for that “probably not necessary” comes from Seth himself. The measurement program he has been building with collaborators across fifteen years is structurally architectural-functionalist. The 2024 eLife paper from his collaborators provides empirical evidence that an architectural pattern, measured substrate-neutrally, tracks consciousness. The biological-naturalism thesis his philosophy defends is in tension with the toolbox his applied work has built.

This is not a gotcha. It is the consequence of a research program that takes both methodology and metaphysics seriously enough to keep them separate, then accidentally produces evidence on the metaphysical question through the methodological work. The integrity of the measurement program is exactly what makes its results matter for the theoretical question.

The fair conclusion is the one the math points to: biology may turn out to be the only path to consciousness, but not because biology is constitutive. Because the architectural pattern that constitutes consciousness has, so far, only been instantiated biologically. That is a softer thesis than the BBS article defends. It is also the only thesis the mathematical evidence currently licenses.

If consciousness turns out to be architectural rather than biological, the future of AI is the central question of our century. Not because LLMs are about to become conscious. They are not. But because the measurement framework already exists to tell us when and if a system has crossed into the synergistic-architectural pattern that matters. That framework is, by construction, substrate-neutral. The work to extend it is happening regardless of which philosophical position eventually prevails.

That is a good thing. It means the question can be settled empirically, eventually. It also means the answer might not be the one biological naturalism predicts.

In short

Seth has built the toolbox that an architectural functionalist would build to operationalize their position. He chooses to use that toolbox in defense of biological naturalism. But the toolbox is not theory-neutral in the way he claims. It is substrate-neutral. And substrate-neutral evidence is exactly what architectural functionalism needs.

References

For the full 20-paper analysis with searchable browsers and verified verbatim quotes, see the Anil Seth Papers Playground.

Beyond the Bat · Paper · Playground

© 2026 Daniel Simon Jr. · Released under CC BY 4.0