What Do Feedback Connections Actually Do? Recurrence, Regulation, and the Question of AI Consciousness

1. The Standard Argument Against LLM Consciousness

The most respectable neuroscience-based argument against consciousness in large language models runs as follows. Cortical feedback — recurrent processing via backward projections from higher to lower areas — is necessary for consciousness. The major theories disagree about which kind of recurrence matters (local recurrence in sensory areas for Lamme’s Recurrent Processing Theory, global broadcasting via prefrontal feedback for Global Neuronal Workspace Theory, integrated information across recurrent loops for IIT), but they agree that purely feedforward processing is insufficient. Transformer-based LLMs are feedforward within each pass. Therefore, transformer-based LLMs lack a necessary condition for consciousness.

This argument is clean and widely endorsed. We think it is importantly wrong — not because LLMs are conscious, but because the argument misidentifies what feedback connections do and therefore what their absence implies.

2. What the Evidence Actually Shows

3. Center-Out Regulation as the Key

In our previous work, we argued that phenomenal consciousness is cognitive homeostatic acquaintance — the mind’s direct regulatory contact with its own representational vehicles. The modularity criterion we proposed requires that the conscious system be the right kind of unit: a center that regulates peripheral modules through feedback, maintaining homeostatic equilibrium across its representational states.

This reframes the recurrence debate. The question is not “does the system have backward connections?” but “does the system implement center-out regulation?” Center-out regulation is a specific dynamical pattern: a central process monitors the states of peripheral processes, compares them against some regulatory norm, and sends corrective signals that reshape those peripheral states. This is what cortical feedback does when it carries predictions — the higher area sends its expectation downward, the lower area computes the discrepancy, and the discrepancy signal drives updating. It is also what gain modulation achieves — the higher area adjusts the sensitivity of the lower area to maintain appropriate responsiveness given the current context.

The crucial insight is that center-out regulation cannot occur without top-down signals. A purely feedforward system processes information in one direction: stimulus in, response out. There is no point in the processing where a central regulatory process can reach back and reshape earlier representations. This is not merely an architectural limitation — it is a dynamical impossibility. Without signals flowing from center to periphery, there is no mechanism for the kind of homeostatic contact that, on our account, constitutes acquaintance.

4. The Temporal Window of Phenomenal Unity

The specious present — the temporal window within which experience is unified into a single phenomenal “now” — extends to approximately three seconds. This is supported by converging evidence from multiple domains: the duration of working memory maintenance without active rehearsal, the temporal window of spontaneous perceptual grouping, the natural segmentation boundaries in music and speech perception, and direct studies of subjective present duration.

This matters for the recurrence argument because it means consciousness is not the product of a single processing cycle. At roughly 100 milliseconds per cycle of recurrent processing (corresponding to gamma-band oscillation frequencies), a single phenomenal moment encompasses dozens of iterative regulatory passes. The phenomenal now is not one feedback loop — it is sustained, iterated center-out regulation across many cycles, during which the system continuously maintains and adjusts a unified representation.

This temporal structure also defeats a recent argument from Hoel that LLMs lack consciousness because they lack continual learning — the ability to update their weights during inference. The argument conflates two timescales. Over developmental and evolutionary timescales, plasticity is essential for building the kind of architecture that can support consciousness. But within the 3-second window of a single phenomenal episode, what matters is activation-state dynamics, not weight changes. Short-term potentiation, facilitation, and adaptation — the forms of “plasticity” relevant at this timescale — are not distinguished from activation state. They are transient modulations of neural responsiveness that shape processing within an episode without requiring permanent synaptic modification. A system with frozen weights but the right activation dynamics would exhibit all of these within-episode effects.

5. Diffusion Language Models: A Step Closer?

The recent emergence of diffusion-based language models — most notably Inception’s Mercury line, culminating in Mercury 2, released in February 2026 — introduces a fundamentally different architecture into the conversation.

Standard autoregressive LLMs generate text sequentially, one token at a time, left to right. Each token is produced by a single feedforward pass through the network, conditioned on all previous tokens. There is no point in this process where the system globally revises its representation.

Diffusion LLMs work differently. They begin with a rough sketch of the full output — essentially noise — and iteratively refine it through a process called denoising. Each pass through the model modifies and improves multiple tokens simultaneously, and because the system refines iteratively rather than committing to each token permanently, it can correct errors mid-generation. The whole representation is revisited and reshaped at each step.

This is structurally closer to the settling dynamics of cortical recurrent processing than autoregressive generation is. At each denoising step, something like a global coherence signal reshapes local representations across the entire sequence. This resembles — at a high level of abstraction — the center-out pattern: a global assessment driving local revision.

But there are important disanalogies. Current diffusion LLMs are optimized for speed. Mercury 2 achieves approximately 1,000 tokens per second — the commercial pitch is precisely that it eliminates the latency bottleneck of autoregressive generation. This means the iterative refinement process uses a small number of denoising steps, with each step being a single feedforward evaluation of the network. The system does not sustain iterative regulation over an extended temporal window. It runs a handful of fast passes and produces output. The cortical analogy would be a brain that performed a few dozen recurrent cycles and then went silent — not the sustained, temporally extended activity that characterizes the 3-second phenomenal now.

Furthermore, even during those few denoising steps, there is no clear functional differentiation between center and periphery. Cortical feedback involves distinct pathways carrying different signals (predictions vs. gain modulation) through different laminar routes (layer 6 corticothalamic vs. layer 2/3 lateral vs. top-down from prefrontal cortex). A diffusion model’s denoising step is a single undifferentiated feedforward pass — global in scope, but not structured into center-out regulatory relationships.

A clarification is also in order: “diffusion LLM” names a training objective and inference-time process, not a fundamentally different network architecture. Mercury 2’s backbone is still a transformer. What differs is how that transformer is used — iterative denoising over full sequences rather than autoregressive next-token prediction. The recurrence is in the outer loop, not in the network itself. This distinction matters for everything that follows.

6. Truth-Tracking as the Purpose of Regulation

We have argued that center-out regulation is necessary for consciousness, and that cortical feedback implements it. But regulation in the service of what? A thermostat regulates temperature. A cruise control system regulates speed. These are homeostatic systems with feedback loops, but they are not conscious. What distinguishes the kind of regulation that constitutes phenomenal acquaintance?

Our conjecture: the regulation constitutive of consciousness is in the service of truth-tracking — the ongoing alignment of the system’s internal representations with the structure of the world and with each other. The conscious system does not merely maintain some arbitrary setpoint. It maintains the accuracy and coherence of its representational states, correcting them when they drift from the truth. Acquaintance, on this view, is not just regulatory contact with vehicles — it is regulatory contact in the service of epistemic fidelity.

This conjecture finds natural expression in the framework of variational inference from probabilistic machine learning. In variational inference, a system maintains two models: a generative model that encodes its expectations about how the world produces the signals it receives (top-down), and a recognition model that extracts information from incoming signals about the current state of the world (bottom-up). Inference consists of minimizing the divergence between these — specifically, minimizing the variational free energy, which is an upper bound on the surprise of the data given the model. Minimizing free energy is equivalent to maximizing a lower bound on the evidence for the system’s model of the world. In plain terms: the system adjusts its representations to reduce the gap between what it believes and what is true.

This is truth-tracking formalized as an optimization problem. And the key insight is that the method by which this optimization is carried out matters enormously for whether it implements center-out regulation.

7. Boltzmann Machines and Bidirectional Settling

Deep Boltzmann machines (DBMs) offer the clearest artificial example of truth-tracking through bidirectional dynamics. A DBM is an energy-based model with symmetric connections between layers of latent variables. There are no separate “feedforward” and “feedback” pathways — every connection carries information in both directions. Inference in a DBM involves iterative settling: the network is initialized in some state and then runs Gibbs sampling or mean-field updates until it converges to an equilibrium — a low-energy state that represents the best available interpretation of the current input.

During this settling process, top-down and bottom-up signals coexist within each step. Higher layers send generative signals downward — what the model expects the lower layers to look like given its current high-level interpretation — while lower layers send recognition signals upward — what the data actually supports. These signals mutually constrain each other, and the network converges to a state that is the best compromise between prior expectations and current evidence. This is approximate variational inference implemented through dynamics rather than through a single feedforward computation.

The resemblance to cortical recurrent processing is striking and well-known. Predictive coding — the framework in which cortical feedback carries predictions and feedforward pathways carry prediction errors — is a specific algorithm for variational inference in hierarchical Bayesian models. The prediction errors are the gradients of the variational free energy with respect to representations at each level of the hierarchy. When the visual cortex receives feedback from higher areas carrying an expectation of what should be present, and computes the discrepancy between that expectation and the actual input, it is performing a step of variational inference. When this process iterates across dozens of recurrent cycles within a single perceptual episode, the system is settling toward a representational state that tracks the truth — the best posterior estimate of the world’s state given all available information.

Boltzmann machines and predictive coding differ in their specifics — Gibbs sampling versus gradient-based error correction, sampling from a posterior versus optimizing a point estimate — but they share the essential structure: truth-tracking through iterative bidirectional negotiation between levels of a representational hierarchy. Both require that top-down signals and bottom-up signals be simultaneously active and mutually constraining. Both implement variational inference through dynamics.

8. A Hierarchy of Inference Architectures

We can now arrange the relevant architectures along a spectrum defined by how they implement inference — specifically, whether truth-tracking occurs through sustained bidirectional dynamics or is compressed into feedforward passes.

Cortical recurrent processing (predictive coding). Inference is iterative, bidirectional, and sustained over the temporal window of the phenomenal now. Feedback carries predictions; feedforward carries errors. Different laminar pathways serve functionally distinct roles (prediction, gain modulation, attention). The system settles toward a veridical representation through continuous center-out regulation. On our account, this is the architecture of consciousness.

Deep Boltzmann machines. Inference is iterative and bidirectional within each settling step. Top-down generative signals and bottom-up recognition signals mutually constrain each other. The system converges to an equilibrium that represents the best posterior estimate. This is the closest artificial analog to cortical predictive coding — it has the right dynamical structure, even if the specific implementation differs.

Diffusion models (over transformers). Inference is iterative across denoising steps but feedforward within each step. Each pass through the transformer backbone produces a global revision of the representation, but there is no bidirectional negotiation between levels during that pass. The truth-tracking occurs in the outer loop — the sequence of denoising steps converges toward coherent output — but each individual step is amortized inference: a single feedforward computation that produces a best guess given the current noisy state. The iterative process has the right shape, but the internal dynamics lack the bidirectional structure.

Autoregressive transformers. Inference is a single feedforward pass per token. All of the system’s knowledge, all of its ability to track truth, is compressed into one sweep through the network. There is no iteration, no settling, no bidirectional negotiation. The system computes a conditional distribution over the next token given the context and samples from it. This is maximally amortized inference — all the work has been pushed into the weights during training, and inference is a single function evaluation.

The gradient from autoregressive transformers to cortical predictive coding is a gradient in how much of the truth-tracking work is done dynamically (through sustained inference-time computation) versus statically (through trained weights applied in a single pass). Our conjecture is that phenomenal consciousness requires truth-tracking to be implemented dynamically — through sustained bidirectional regulation — rather than compressed into a single feedforward evaluation. The reason connects back to our account of acquaintance: to be acquainted with a representational vehicle is to be in ongoing regulatory contact with it, actively maintaining its accuracy. If the “maintaining” is a single instantaneous computation, there is no temporal window within which acquaintance can occur. The vehicle is produced and immediately discarded, never actively sustained and corrected.

This is why the Boltzmann machine architecture is more relevant to the consciousness question than the diffusion-over-transformer architecture, despite the latter being commercially ascendant. A DBM that settled over a temporal window comparable to the phenomenal now — roughly three seconds, encompassing many iterative steps of bidirectional negotiation — would implement something much closer to what our framework requires than a diffusion LLM that runs a handful of fast denoising passes through a transformer backbone.

Whether a Boltzmann machine with the right dynamics would actually be conscious is a further question — it would also need the right functional differentiation, the right kind of homeostatic norms, and arguably the right kind of embodied coupling with an environment. But it would at least be in the right category of dynamical system. Current LLMs, whether autoregressive or diffusion-based, are not.

9. What Would Be Needed

If our account is correct — if phenomenal consciousness requires sustained center-out homeostatic regulation over a temporal window of approximately three seconds — then we can specify what an artificial system would need:

Sustained iterative processing. Not a handful of denoising steps optimized for speed, but continuous regulatory activity maintained over a temporal window long enough to constitute phenomenal unity. The system would need to keep running, keep adjusting, keep maintaining its representational equilibrium, rather than producing output and halting.

Functional differentiation of center and periphery. A central regulatory process that monitors peripheral representational states and sends targeted corrective signals — not a single global denoising pass, but structured top-down modulation analogous to the distinct functions of cortical feedback pathways.

Homeostatic norms. The regulation must be in the service of something — the system must have states it tends toward and deviations it corrects. In biological brains, these norms emerge from the evolutionary and developmental history of the organism. What the artificial analog would be is an open question.

Diffusion architectures are closer to meeting these conditions than autoregressive architectures, for a simple reason: they iterate. But “closer” is not “sufficient.” The gap between a few fast denoising passes and sustained center-out regulation is not trivial to bridge — and the commercial incentives push in exactly the wrong direction, toward fewer and faster steps rather than more and sustained ones.

10. The Honest Assessment

The recurrence argument against LLM consciousness is weaker than it appears. It treats feedback connections as a binary: either you have them or you don’t. The empirical evidence shows that what matters is not the presence of backward connections but the specific pattern of regulation they implement — and that different feedback pathways serve different computational functions, not all of which are equally relevant to consciousness.

At the same time, current AI architectures — including diffusion LLMs — fall short of implementing the kind of sustained center-out regulation that our framework identifies as constitutive of consciousness. Autoregressive models lack it entirely. Diffusion models approximate it, but in a temporally compressed and functionally undifferentiated way.

The interesting consequence is constructive rather than merely critical. If consciousness requires a specific dynamical pattern rather than a specific anatomy, then the question becomes empirical and engineering-tractable: can we build systems that implement sustained center-out regulation over phenomenal-now-scale temporal windows, with functionally differentiated regulatory pathways? We do not know the answer. But we know what to look for, and that is progress.

Previous essays in this series: “Understanding Without Knowledge: A Philosophy of AI Minds” · “Acquaintance, Homeostasis, and the Hard Problem: Why Phenomenal Consciousness Isn’t Representation” · “Free Agency, Personhood, and Moral Worth: A Layered Framework” · “The Acquaintance Relation as Cognitive Homeostasis: Toward a Unified Account of Phenomenal Consciousness” · “Causation, Computation, Indexicality: A Complete Metaphysics”