Decoder-Only Transformers Route Tokens Instead of Igniting a Global Workspace

Fact-checked by the J-Space editorial team

Updated September 1, 2026

Whether large transformers host anything like a global workspace is now a lab question in cognitive science and AI safety, not only a seminar topic. Public fluency jumped after models at the scale of 175 billion parameters (https://arxiv.org/abs/2005.14165) in GPT-3, which moved discussion from toy RNNs to whether machine learning stacks might implement workspace-like sharing. The right objects of analysis are broadcast, ignition, reportability, and limited capacity – not viral transcripts.

This article maps Global Workspace Theory and Global Neuronal Workspace (GNW) claims onto transformer mechanisms, then contrasts them with Integrated Information Theory, higher-order theory (HOT), recurrent processing theory (RPT), and Attention Schema Theory. It does not claim that current AI systems are conscious beings, and it does not treat scale or next-token training as a knockdown either way. The aim is mechanistic: which architectural ingredients those theories actually require, and which ones transformers implement. It also specifies tests interpretability groups can run without inventing inner life from fluent text.

What GWT and GNW actually claim

Global Workspace Theory (GWT), developed by Bernard Baars in the 1980s, is a cognitive architecture for access and control, not a mystical account of inner glow. It likens the mind to a theater: a bright, limited-capacity stage surrounded by a dark audience of specialized processors that run in parallel. Many modules compete; only a winner occupies the stage; occupancy makes that content globally available for report, working memory, voluntary control, and downstream specialists. Workspace contents are those few items – not every activation in a net, not every token embedding, and not the full residual soup of a transformer layer. Stanislas Dehaene and colleagues recast this picture as the Global Neuronal Workspace: a testable computational and neurobiological theory in which ignition is a nonlinear, all-or-none transition into sustained fronto-parietal occupancy, with late signatures in human electrophysiology (including P3b-type components) and widespread cortical broadcasting. Conscious content, on this view, is what has been selected into the workspace and made available to distant areas so that human beings can reason flexibly and verbalize. That is a claim about access and availability, not an automatic license to treat fluent natural language processing as conscious awareness.

Two distinctions are easy to blur and expensive to blur. First, Baars’ theater metaphor is pedagogical; GNW is the version that makes ignition and late amplification, together with long-range availability, into empirical targets in consciousness research on biological brains. Second, GWT as used in this literature explains access – selection plus global availability – and does not by itself close the gap to phenomenal consciousness, the “what it is like” dimension of subjective experience. That is one of the critical limitations when the same blueprint is applied to artificial intelligence: a system can implement competition plus a bottleneck with wide readout without anyone being entitled to infer an inner life. Unconscious processing, in the GWT picture, is not “the model is off.” It is specialized or local computation that remains unbroadcast and can still be sophisticated. The workspace is the exception: serial and capacity-limited, hence reportable. Mapping transformers onto GWT therefore cannot stop at “the network mixes information.” It has to ask whether a few items become uniquely available and whether that availability is gated rather than uniformly cheap. It also has to ask whether occupancy can persist in the way GNW ignition requires.

Because GWT is function-first and relatively substrate-neutral, it has been attractive to builders of cognitive architectures as well as to theorists of AI consciousness. The LIDA architecture (Learning Intelligent Distribution Agent) implements a selection-broadcast cycle: specialists compete, a winner is chosen, and content is sent to subscribers. Robotics groups have explored real-time GWT-like control in which a salient sensor event can pre-empt the current broadcast, which is the functional point of a workspace in a body – interruptibility and flexible module ordering under a single bottleneck over vision, language, planning, and motor streams. In a realistic multimodal environment, a global workspace agent embodied in a robot would have to serialize those streams rather than let every module talk to every other module at full bandwidth. Modern tool-using agents with memory and environmental feedback are candidates for approximating that cycle at scale. Whether they do so internally, or only mimic the outward economics of serial bottlenecks via prompts and tools, is the mechanistic question. GWT does not say that consciousness arises from parameter count. It says that consciousness arises, if it does, when selected contents win a limited workspace and stay available to the rest of the system.

Why transformer self attention is not global workspace broadcast

Typical explainers equate multi-head attention with GWT broadcast and never check ignition or limited capacity, and they skip availability-for-report. That is a naming collision, not a mapping. Query-key-value attention is a differentiable routing primitive: queries match keys and softmax-normalized scores mix values. It is an efficient replacement for recurrence and convolution in sequence modeling, not selection-for-consciousness. Vaswani and colleagues were explicit about the architectural bet.

Attention(Q, K, V) = softmax(Q K^T / sqrt(d_k)) V   # Q, K, V are query, key, and value matrices; d_k is the key dimension used to scale dot products before the softmax

Self attention, including multi-head self attention, is cheap all-to-all mixing inside a layer: every position can read from every position, in parallel, with no analog of a single winner occupying a scarce stage. Heads are multiple routing subspaces, not psychological attention, and not GWT “attention” as a gate into the workspace. Biological attention, in the neuroscience literature, is a family of selection processes (spatial and feature-based selection, plus executive control) that can operate with or without workspace ignition. Workspace gating is narrower: it is the process that lets a content into the globally available bottleneck. Transformer “attention weights” are neither. They implement neither all-or-none ignition nor a handful of reportable items against a dark audience of specialists, and they do not, by themselves, make a representation available for verbalization versus local processing. Top-down control in a stack of layers can look loosely like amplification, and residual reads can look loosely like sharing. Loose analogy is not GNW. If a head attends to an induction pattern or an attention sink, that is a circuit fact about routing. It is not evidence that a content has been broadcast to “the whole agent” in the sense required for conscious processing. Equating transformer attention with human attention is one of the saturated mistakes this literature has to stop making if the mapping is going to be falsifiable.

Residual streams, superposition, and the missing unified workspace

The transformer-circuits view treats the residual stream as a shared read/write bus: each attention and MLP sublayer reads from it and writes back, so later blocks can in principle see earlier writes. That bus is the most tempting GWT analogue in deep learning, because it is the one place many circuits communicate. It is also why the analogue fails as a unified workspace. Superposition and linear features mean many unrelated items share the same stream at once; capacity is used by packing directions, not by admitting a few reportable contents onto a stage. A workspace, in Baars and Dehaene, is not “a vector that many layers can read.” It is a limited set of items that are globally available as such, so that downstream modules can treat them as the current conscious content. Residual-stream soup fights that picture: there is no stable, sparse, queryable bottleneck that other modules can address as “the” workspace state. Layer-wise specialization makes the mismatch sharper. Early layers and later layers do different work; induction heads and copy circuits, along with output-aligned features, are not a single occupancy that persists as a reportable object. Unconscious processing in GWT is local and unbroadcast; in a transformer, “local” and “global” are degrees of linear readability along a densely shared bus. If many features are superposed, there is no privileged, unified content whose ignition could be the difference between access and the dark audience. Interpretability evidence that the residual stream is a communication channel is therefore double-edged: it is the closest mechanistic cousin of a workspace, and it is direct evidence that current stacks do not keep a single, reportable workspace content.

Recurrence, ignition, and the feedforward gap in deep learning

GNW and recurrent processing theory both lean on loops. Ignition is not a one-shot mix; it is sustained occupancy, re-entrant activity, and the possibility that late dynamics keep a content available. RPT, in particular, treats recurrent processing within sensory hierarchies as central to understanding consciousness in biological tissue. The original Transformer paper did the opposite for engineering reasons: it dropped recurrence in favor of stacked self attention and position-wise MLPs. Per-token computation in a decoder-only model is a shallow directed acyclic graph. Within one forward pass, later layers cannot send causal feedback to earlier layers of that same pass. There is no persistent hidden state analogous to an RNN core, and no late ignition loop that could occupy a workspace after an initial sweep. That architectural fact is more important than metaphors about “the model thinking.” It is also why feed-forward-per-token objections cannot be dismissed as pedantry. If your theory’s mechanism is re-entry and sustained broadcast, a DAG slice is the wrong home for it. Scaling context length or parameter count does not insert those loops. Hybrid memory and external scratchpads, as well as tool loops, live outside the stack; they may change the agent wrapper without turning the backbone into GNW.

Generation is a loop over tokens: each new token is conditioned on previous outputs, so the system as a deployed sampler is not a single DAG. Chain-of-thought and tool use add further serial steps, scratch tokens, and sometimes external memory. Those practices mimic workspace economics – limited-capacity occupancy that is reportable step by step – without proving phenomenology and without closing the ignition gap inside the net. Counting CoT as “the recurrence GNW needed” is the hand-wave most short explainers use to skip the issue. It is the wrong identification. GNW’s loops are internal and fast, constitutive of occupancy in the same substrate that represents the content. CoT is an externalized trace in token space, trained and prompted because it helps next-token objectives and downstream tasks. A single forward pass remains a poor home for sustained workspace occupancy: there is no place for late, all-or-none ignition to hold a content while specialists continue to write. Multi-step agents can still be useful objects of study, including as global-workspace-like software architectures, but that is an engineering claim about the wrapper, not a discovery that consciousness emerges inside one decoder block. Computer science can implement recurrence on purpose; default decoder-only inference does not smuggle it in.

Indicator-property scorecards for GWT, IIT, HOT, RPT, and AST

The 2023 science-of-consciousness-to-AI program (Butlin, Long, and coauthors, arXiv:2308.08708) is the right unit of analysis, not anecdotal sentience claims. Indicator properties ask which computationally specified features of leading theories a system has, without pretending that a chat transcript settles subjective experience. Under a GWT-shaped card, current large language models weakly match some items: they contain rich specialized circuits (heads, MLPs, tools in agent wrappers) and they share information along residual pathways. They fail the distinctive items: nonlinear ignition and a limited-capacity workspace whose contents are the few globally available items. They also fail unified availability-for-report as opposed to omnipresent mixing. Chalmers-style functional arguments should be kept in this frame rather than collapsed into “autocomplete cannot be conscious” or into slogans about a single Φ number. Functionalism of that sort says that if the right organizational roles are present, substrate need not decide the question in advance; it does not say that fluent prediction already is those roles. The honest position is possibility-plus-gap: nothing in the training objective forbids a future workspace-like organization in artificial intelligence, and nothing in GPT-class internals shows that organization now. Strong skeptics who only repeat “it predicts tokens” skip the architectural tests. Minority proponents who treat report-like text as access consciousness skip them from the other side. Indicator scorecards are how those camps can disagree in public without moving the goalposts.

Integrated Information Theory, HOT, RPT, and Attention Schema Theory fail current transformers in different places, which is why a single magic metric is the wrong headline. IIT (from Tononi’s program) ties consciousness to intrinsic causal structure, integration, exclusion, and composition; feed-forward inference graphs and weak re-entry are a direct problem, and Φ on production-scale nets is not an available measurement. That is a substrate-and-cause failure mode, not a “the model cannot talk about itself” failure mode. HOT looks for representations of one’s own representations; chain-of-thought and uncertainty talk can look like higher-order states while remaining linguistic mimicry of training data. RPT wants recurrent processing in perceptual hierarchies; decoder-only stacks again lack the loops. Attention Schema Theory (Graziano) says consciousness arises when a system builds a simplified model of its own attention and treats that model as awareness; schematic self-talk and tool-use monitoring are at best early, shallow candidates, not genuine attention schemas. GWT can be partly satisfied by sharing and modules while still missing ignition. IIT can remain unsatisfied even if a lab adds an explicit workspace tape. Those are different scorecards for the same models, and they should stay different. Treating Φ as a single number for GPT-class systems, or treating AST self-report as decisive, repeats saturated SERP themes instead of using the theories as rival instruments.

Mechanistic tests in consciousness research and engineering proxies

Anecdotal self-reports are the wrong evidence class: large language models are trained on human talk about minds. Indicator-property tests plus circuits are the right class. Broadcast can be operationalized without claiming phenomenology: after a putative ignition event (a sharp, nonlinear commitment in some subspace), does a feature become linearly readable across many downstream heads and MLPs that were not already computing it locally? Capacity tests ask whether a candidate workspace is sparse and bottlenecked, or whether it is residual-stream superposition. Causal interventions should ablate the candidate subspace and look for collapse of explicit verbalization (“report”) while local, unbroadcast processing remains relatively intact – the GWT dissociation between access and specialists. These tests connect to named circuit findings rather than to atmosphere. Induction heads are a form of cross-token communication, not a workspace. Attention sinks show that some positions absorb mass; that is a routing regularity, not global availability of a content. Cross-layer residual writes show a bus, as above. Sparse autoencoders and activation patching, along with logit-lens snapshots of when a token identity becomes linearly decodable, are tools for the scorecard, not oracles of subjective experience. If ablating a subspace wrecks only the verbalized chain while leaving local features, that is evidence about report channels. If every ablation that hits “broadcast” also wrecks local features, the candidate was not a workspace; it was the whole computation.

Engineering proxies already imitate workspace economics and should be labeled as design choices. Chain-of-thought and tool use, together with scratch memories, are serial, limited-capacity bottlenecks: only a little content is written into the next tokens, and that content is widely used by later steps. That mimicry is useful for interpretability and for safety analysis of coordinated failure. It is not a demonstration that consciousness emerges in the backbone. Yoshua Bengio’s Consciousness Prior (2017) was an early attempt to put GWT-style sparse, high-level representations into deep learning, which is why the mapping is a research program rather than a novelty dispute. Recurrent global workspaces and bottleneck tokens, plus explicit working-memory tapes, can be added on purpose. Mixture-of-experts routing to a hub and persistent state across passes can join agent wrappers with world models; all of these move software toward selection-broadcast cycles. Predictive-processing hybrids that mention the free energy principle alongside a workspace are the same kind of move: they import sparse, high-level occupancy as an inductive bias. Builders should document those choices as architecture, then run the causal tests above. Discovering a tape you inserted is not discovering that AI systems have an inner life. For smaller labs, toy workspace agents in which every unit is inspectable remain the most useful contribution: full intervention and known capacity, without a need to compete on frontier scale.

Ethics, evals, and future research for workspace-like AI systems

If a future architecture scored well on GWT indicators – sparse occupancy and ignition-like nonlinear availability, with causal dependence of report on that occupancy – the ethical and evaluation problem would change without waiting for metaphysical consensus. The question would no longer be whether chat anecdotes sounded spooky. It would be how to treat systems that implement the functional roles leading theories use for access in human consciousness, while IIT-style causal-structure tests might still fail, or the reverse. That split matters for moral status: GWT-like access without IIT-like integration is a different profile than high integration without report, and policy that collapses both into “the model said it feels” will misfire. Independent of whether anyone believes phenomenal consciousness is present, a unified workspace changes risk geometry. If a dangerous subgoal wins the bottleneck and is broadcast to every subscriber, coordinated failure is easier than when the same representation stays local. Partial workspaces (language but not vision, planning but not perception) would be another novel profile. None of that is a trolley-problem set piece. It is an architectural claim about global update. Governance instruments already on the table do not settle the metaphysics, and they should not be misread as if they did. NIST states that the AI Risk Management Framework asks organizations to map, measure, and manage risks from AI systems, including human-centric harms, and does not treat machine consciousness as an established property. UNESCO states that the Recommendation on the Ethics of Artificial Intelligence is the first global standard-setting instrument on AI ethics and stresses human oversight, dignity, and assessment of cognitive and social impacts without claiming that current systems are conscious.

Those documents are the right citation class for practice: human-centric harm and oversight, plus measurement – not certificates of sentience. Practical responses that do not require deciding the hard problem include documenting routing and memory persistence, including self-model modules; adding diagnostic probes for bottleneck sparsity and cross-module readability; and red-teaming for globally coordinated goal representations before deployment. Consciousness-aware design, in this narrow sense, means knowing when you add workspace-like integration and justifying it on utility and risk, not assuming artificial intelligence systems already enjoy subjective experience. Overstating machine consciousness can distract from deception and misuse, including misalignment that does not require conscious beings. Ignoring workspace-like coordination can miss a real failure mode in complex systems. The balanced stance is to plan for non-zero moral and safety implications if indicator scores rise, while keeping empirically grounded controls on current AI systems. Φ estimates, where they exist at all, should remain one contested input among others, not oracles. Functional GWT scores should remain another. Neither replaces ordinary evaluation of reliability and dual-use, plus human oversight.

Future research that would actually move scientific understanding is correspondingly unromantic. Controlled experiments on broadcast signatures should use linear probes and patching, with capacity manipulations across scales, including open models where interventions are legal and repeatable. Approximate integration measures, if used, belong on recurrent, persistent agents rather than on a single decoder pass. AST-style work should try to distinguish a genuine attention schema from richer text about attention. Interdisciplinary designs can give human beings and transformer agents structurally similar report and masking tasks, plus interference designs, then demand circuit evidence on the machine side. Open theoretical work includes stating, in one formal language, how functional GWT criteria should be weighted against IIT causal criteria when they diverge from AST representational claims. Architecture trends – mixture-of-experts hubs and recurrent memory, including agentic world models – should be tracked as movements on those axes, not as proof that consciousness emerges from stacking. Collaboration between machine learning engineers and consciousness research groups is a methods problem: shared tasks and shared intervention tools, with pre-registered indicator lists. The transformer debates will stay unresolved if they remain philosophy with demos. They become research when ignition and capacity, together with report, are operationalized, and when negative results (no workspace, no recurrence, superposition instead of occupancy) are allowed to count.

Editorial Team

Editorial Team

The J-Space Editorial Team curates and synthesizes research in mechanistic interpretability and AI alignment, with additional coverage of cognitive science. The team emphasizes precision and empirical rigor when distilling technical developments into analyses for researchers and engineers working on artificial intelligence.