Prediction Survives Anaesthesia. Recursion Doesn’t.

The paper is one of the most carefully done unconsciousness experiments of the past few years. Katlowitz and colleagues used Neuropixels microelectrodes to record single-unit activity from human hippocampus in seven patients undergoing propofol-maintained general anaesthesia. Three patients listened to pure-tone oddball sequences. Four listened to natural speech podcasts. The recordings were taken during clinical loss of consciousness, with inhaled induction agents stopped at least an hour before the recordings began.

What they found is worth reading slowly. Hippocampal single units and local field potentials reliably encoded oddball discrimination. The effect grew over approximately ten minutes, demonstrating representational plasticity in the anaesthetised hippocampus. With language stimuli, hippocampal neurons encoded the semantic categories of words being heard, their part of speech, and contextual features sufficient to decode what word was coming next. The authors caution that future-word decoding does not by itself imply active prediction beyond contextualization — an honest hedge worth preserving — but the encoding is real and it operates at the level of meaning and syntax, not raw acoustics.

The authors’ framing is sharp. They write that “prominent theories of consciousness propose that sophisticated pattern recognition, semantic interpretation and predictive processing all require conscious access.” They cite Global Workspace Theory, Integrated Information Theory, Recurrent Processing Theory, and Dennett. They show that the named ingredients — pattern recognition, semantic interpretation, predictive encoding — survive into a brain region anatomically and functionally distant from primary sensory cortex, in the absence of consciousness. Their conclusion is that consciousness “may have some association with those processes but is not essential.”

RSP is in the same neighbourhood as the cited theories. Beyond the Bat is a prediction-based account of consciousness in the sense that it grounds phenomenology in a specific kind of hierarchical predictive process. The natural worry is that Katlowitz et al. are at the gate. If hippocampal prediction continues without consciousness, doesn’t that cut the link RSP needs between prediction and phenomenology?

The reply is that RSP never claimed every kind of prediction is consciousness. RSP claims that a specific kind of prediction — Level 3b recursive social self-modeling — is what constitutes phenomenal consciousness. Katlowitz et al. study different levels of the architecture. Their findings are what RSP’s level-separation predicts.

What Katlowitz et al. Got Right

The technical contribution is substantial and deserves to be acknowledged on its own terms, separate from any theoretical use it gets put to.

The recording technology is appropriate to the question

Neuropixels probes give single-unit resolution at scale. The hippocampal recordings are dense enough to support population-level decoding, not just averaged-evoked responses. This matters because earlier anaesthesia studies relied on EEG or coarser intracranial recordings that cannot distinguish whether a given hippocampal computation is preserved or destroyed. Katlowitz et al. can resolve that question. They show preservation.

The hippocampus is the right region for the test

The hippocampus is, as the paper notes, “anatomically and functionally distant from primary sensory cortices.” Earlier work showed residual sensory responses in early cortex during anaesthesia, but the standard reply was that those are bottom-up sensory cascades, not high-level computation. The hippocampus is far enough up the processing hierarchy that this reply does not apply. If high-level semantic and grammatical encoding shows up in unconscious hippocampus, the “it’s just feedforward sensory drive” explanation is closed off.

The ten-minute plasticity is the load-bearing finding

The oddball effect size grew over approximately ten minutes of exposure. This is not a one-shot stimulus response. The hippocampus, during anaesthesia, was updating its representations as it accumulated evidence about which tone was rare and which was common. Plasticity on this timescale is associated with learning. That learning is happening in the unconscious state is the result that constrains theories of consciousness the most. Generic feedforward responses do not require theoretical reinterpretation. Learning does.

The RNN model demonstrates emergence, not assumption

The authors paired the empirical recordings with a continuous-rate recurrent neural network trained on the tone-discrimination task, including an excitatory-inhibitory variant in which the four classes of recurrent connection (E–E, E–I, I–E, I–I) were systematically lesioned to test which contribute to the oddball signature. The RNN reproduced the oddball-plasticity pattern as an emergent property of flexible tone discrimination — not something the architecture had to be specifically wired for. This matters because it shows the result is not artefactual or dependent on a specific cognitive interpretation. A generic recurrent network performing the relevant task produces the same signature.

The anaesthetized numbers look awake

The detail that turns this from “residual processing” into something sharper is how little the anaesthetized representations differ from awake ones. Nearly all units showed semantic-category selectivity: 85.6% (321 of 375), against 76.1% (271 of 356) in the awake comparison cohort. Part-of-speech information was carried by 298 of 375 units under anaesthesia, and by an all-but-identical 298 of 356 awake. Firing rates tracked surprisal — the improbability of each word given its predecessors — in 246 of 375 units, against 245 of 356 awake. Future words were decoded from current activity nearly as well as past words, and the anaesthetized and awake decoding coefficients did not differ significantly. The authors’ own summary is that “awake-like semantic responses, and at least some contextualization, can occur in the absence of conscious awareness.” Conceptual and grammatical categorization ran at close to full waking capacity while the patient was unconscious.

The Apparent Problem for Prediction-Based Theories

Katlowitz et al. are not attacking a strawman. The theories they name — GWT, IIT, RPT, Dennett — all link some form of higher-order computation with consciousness. The natural reading of their result is that this link is empirically too tight. The brain does sophisticated computation without consciousness all the time, including in the hippocampus, including under general anaesthesia. Whatever consciousness adds to the picture, it cannot be identified with semantic processing or with predictive encoding, because both occur without it.

RSP could be lumped into this critique. Beyond the Bat identifies phenomenal consciousness with a kind of predictive processing. So Katlowitz et al.’s data, naively read, is supposed to threaten that identification. If prediction continues unconsciously in the hippocampus, then prediction is not what consciousness is.

The naive reading is wrong, and the reason it is wrong is exactly the architectural commitment that distinguishes RSP from the theories Katlowitz et al. name.

What RSP Actually Claims

RSP is not the claim that all prediction is consciousness. It is the claim that a specific kind of prediction — recursive social self-modeling at Level 3b — is constitutive of phenomenal consciousness. The framework distinguishes several levels of predictive processing, and splits the top level into conceptual categorization and recursive self-modeling:

• Level 1 (core affect, ~100 ms timescale): Interoceptive prediction of homeostatic state. The brainstem and insula maintain a continuous predictive model of the body’s own internal variables. This level is preserved through clinical loss of consciousness; the patient’s vitals remain stable.
• Level 2 (core schemas, 1–10 s timescale): Schema-driven organization of body, spatial, attentional, and temporal structure. The hippocampus is a schema substrate at this level, binding sequence and relational structure across time.
• Level 3a (conceptual categorization): Categorical and semantic knowledge — the mapping of sensory and linguistic input into concepts, emotion categories, word meanings, and grammatical classes. This is the level at which “this tone is the rare one” or “this word is a place-noun” gets represented. The semantic and part-of-speech decoding Katlowitz et al. recover is Level 3a categorization riding on the Level 2 hippocampal schema.
• Level 3b (recursive social self-modeling, 10 s–minutes timescale): A specific recursive loop in which the system models itself being modelled by others, models its own model of that, and so on. This recursion converges to a stable self-representation, and that recursive process — not conceptual categorization, not prediction in general — is what RSP identifies with phenomenal consciousness.

The architectural claim about anaesthesia, on RSP’s reading (not on Katlowitz et al.’s), is that general anaesthesia selectively disrupts the cortico-cortical recurrent loops that implement Level 3b. Level 1 brainstem function is preserved by design; if it were not, the patient would die. Level 2 schemas and Level 3a conceptual categorization are partially preserved because they live in distributed cortical and subcortical networks (including the hippocampus) that can continue operating on their own dynamics when the long-range recursive integration at Level 3b is suppressed. This is the sharper reading the final paper forces: what survives is not merely low-level schema processing but conceptual-level categorization — semantics, grammar, surprisal — at close to awake fidelity. Even that does not require consciousness. Only the 3b recursion does.

Katlowitz et al.’s result is exactly what this architecture predicts. The hippocampus, a Level 2 schema substrate supporting Level 3a conceptual categorization, should continue to encode semantic and grammatical features under anaesthesia. It should also continue to update its representations on exposure to new sequences — the ten-minute plasticity finding. What it should not do is participate in the Level 3b recursive self-modeling loop that RSP identifies with consciousness, because that loop depends on cortico-cortical recurrence that anaesthesia disrupts. Katlowitz et al. did not test this last prediction; their recordings target the hippocampus, not the cortico-cortical loop. But their data is consistent with what RSP claims about the parts of the architecture they did measure.

Why the Level Distinction Does Real Work

A skeptic could fairly worry that the L1 / L2 / L3a / L3b distinction is doing convenient work, slicing the architecture into levels exactly such that the troublesome empirical result falls on the “not consciousness” side. The worry sharpens here, because the surviving processing is not being relegated to some low schema tier: it is conceptual categorization, which RSP places at Level 3a. The reply cannot be “that was never high-level.” It has to be that the 3a/3b split — conceptual categorization versus recursive self-modeling — was drawn on independent grounds, and the data lands on it.

The reply is that the three-level distinction was not introduced for Katlowitz et al.’s data. It is one of the load-bearing architectural commitments of RSP, established on independent grounds long before this paper. Each level is anchored in independent evidence:

• The developmental sequence: Level 1 capacities (interoception, affect) are present from before birth; Level 2 schemas (body, spatial, attention) come online over the first two years, with mirror self-recognition around 18–24 months at Level 2’s upper boundary. Recursive self-modeling (Level 3b) consolidates far later: explicit false-belief understanding around four to five years, fully recursive second-order theory of mind around six to seven. The framework does not stamp Level 3a conceptual categorization with a single age, but children command word meanings and grammatical classes well before recursive theory of mind consolidates — conceptual competence precedes the recursive self-model, and anaesthesia dissociates the two in that same direction. The cascade is predicted independently by free energy, reinforcement learning, and cascade-control formalisms (Appendixes B, C, D).
• Clinical dissociations: Appendix G maps clinical conditions onto specific levels. Cotard’s delusion, depersonalization, and dissociative identity disorder selectively disrupt Level 3b self-modeling with Levels 1, 2, and 3a categorization preserved — patients still parse language and hold concepts; what fractures is the self the concepts are anchored to. Out-of-body experiences, hemispatial neglect, and somatoparaphrenia disrupt Level 2 schemas with the other levels typically preserved. The level boundaries are visible in real disorders, not just diagrams.
• Timescale separation: The three levels operate at different temporal scales (Appendix D), and the cascade-control formalism requires this separation for the architecture to converge. Inner loops must stabilize before outer loops can be commissioned.
• Comparative evidence: Metacognitive-monitoring paradigms (Hampton, 2001; Call & Tomasello, 2008) make recursive depth measurable across species, and the comparative profile is graded rather than binary — great apes and rhesus monkeys show signatures of limited recursive depth while species without the social-prediction machinery show none. The graded profile is what an architecture with continuous recursive depth predicts.

The Katlowitz et al. data lands on a partition the framework had already drawn for these independent reasons. It is not retrofitted.

The Positive Prediction RSP Was Already Making

The framing reverses. Katlowitz et al. is not a counterexample to RSP. It is one of the cleanest positive empirical predictions RSP makes: under general anaesthesia, Level 1, Level 2, and Level 3a processing should survive in the cortical and subcortical regions that house the relevant schemas and categories, even as Level 3b recursive integration is selectively disrupted. The hippocampus is one of those substrate regions. Katlowitz et al. supply the cleanest direct test of that prediction available, in the cleanest possible region for the test.

The framework had reasons to want this experiment done long before Katlowitz et al. did it. The standing anaesthesia literature (Alkire, Hudetz, & Tononi 2008; Casali et al. 2013; Sarasso et al. 2015) showed that perturbational complexity drops sharply during loss of consciousness while raw sensory responses persist. The interpretive question that hung over those results was: do high-level computations persist or are they specifically disrupted? Katlowitz et al. answer the question for the hippocampus: high-level semantic and grammatical computations persist. RSP’s prediction is that the same answer should hold for any Level 2 or Level 3a substrate, with Level 3b cortico-cortical recurrence being the specific class of dynamics that anaesthesia targets. That is a falsifiable prediction. It is now harder to falsify because of Katlowitz et al., not easier.

What RSP Predicts That Katlowitz et al. Did Not Test

The framework makes three further predictions that follow from the L1 / L2 / L3 reading of the result:

1. Recursive self-modeling loops should fail to converge under anaesthesia even when their substrate regions remain active

The dorsomedial prefrontal cortex, temporoparietal junction, and posterior cingulate are key nodes in Level 3b recursive self-modeling. Their constituent neurons may continue firing under anaesthesia, but the recurrent dynamics that would normally produce a stable converged self-representation should not be measurable. Katlowitz et al. did not place electrodes in these regions. A version of their experiment that did, and that included a recursive self-prediction task adapted for unresponsive patients, should fail to recover the kind of stable Level 3b attractor that the same recordings would show in conscious controls. This is the experiment that would specifically test the recursive self-modeling claim.

2. Anaesthetic agents that preferentially target cortico-cortical recurrence should show a different signature from agents that target thalamocortical loops

The cortico-cortical recurrence story predicts that propofol and similar agents that target broadly distributed cortical synapses should suppress Level 3b most reliably. Ketamine, which selectively disrupts NMDA-receptor-mediated cortical-cortical connectivity, should produce a more selective profile — specifically disrupting Level 3b recursion while preserving Levels 1, 2, and 3a more cleanly than propofol does. Some of this signature is already visible in the PCI literature (Sarasso et al. 2015 finds high PCI under ketamine despite disrupted consciousness). The Katlowitz paradigm extended to ketamine should show hippocampal Level 2 processing even more clearly preserved than under propofol.

3. The same architecture predicts when consciousness should return

On RSP’s reading, consciousness returns when cortico-cortical recurrent dynamics restabilize at the Level 3b timescale, not when neurons start firing again in any particular region. This is a specific prediction about the temporal profile of recovery from anaesthesia: behaviour-relevant Level 2 schemas and Level 3a categorization should reactivate before Level 3b self-reportability returns. A version of Katlowitz’s paradigm done across the emergence-from-anaesthesia window should be able to track the levels coming back online in cascade order.

Honest Caveats

Several things in this argument deserve being held loosely.

First, RSP’s claim that anaesthesia specifically disrupts cortico-cortical recurrence supporting Level 3b is RSP’s claim. Katlowitz et al. do not use the phrase “cortico-cortical recurrence” anywhere in the paper. They do not discuss the mechanism of anaesthetic action. They simply show that high-level hippocampal computation survives. The architectural interpretation of why is RSP’s, and it has not yet been independently confirmed by the kind of direct measurement that would target Level 3b cortico-cortical loops.

Second, “future-word decoding” is not the same as “active prediction.” The authors are explicit about this. What they show is that hippocampal activity carries contextual information sufficient for an external decoder to recover features of upcoming words. Whether this corresponds to an active forward model running in the hippocampus is a stronger claim the authors decline to make. The argument here goes through on the weaker reading: high-level computation survives anaesthesia, regardless of whether one labels it “prediction” in the strong sense.

Third, the result is propofol-specific. The authors flag this directly: it may not generalize to sleep, coma, or other anaesthetic agents. RSP’s converse prediction — that L3 cortico-cortical recurrence is what anaesthesia targets — is therefore most cleanly tested with propofol-similar agents. Sleep poses a different problem: REM sleep involves Level 3b reactivation (dreaming) without normal sensory input, which is the kind of case the framework wants to use as a comparative test.

Fourth, RSP was not in Katlowitz et al.’s field of vision when they wrote the paper. Beyond the Bat is unpublished and unreviewed at the time of their acceptance. The data is consistent with RSP’s level-separation claim, but Katlowitz et al. are not endorsing the framework, and the consistency is not unique to RSP — any architecture that distinguishes higher-level integrative consciousness from lower-level prediction would absorb the result the same way. The discriminating predictions are at Section 7’s level (what should happen at Level 3b substrate regions, what the temporal profile of recovery should look like), not at the level of the empirical finding itself.