Global Workspace Theory (GWT) explains consciousness as a theater of the mind, where one thought is spotlighted and shared brain.

For decades, consciousness research has advanced less like a straight road than a series of detours, through philosophy seminars, neuroscience labs, and, more recently, machine-learning conferences. One of its most durable guideposts is Global Workspace Theory, first articulated by the cognitive scientist Bernard Baars and later given neural muscle by Stanislas Dehaene. The theory begins with a modest but unsettling observation: most of what the mind does happens in the dark. Consciousness, when it appears, is less a factory than a bulletin board.

In human cognition, countless specialized processes, visual parsing, memory retrieval, motor planning, operate in parallel, largely unaware of one another. What we experience as consciousness is the moment when one of these processes breaks through and is broadcast widely, made available to the whole system. To be conscious of something, on this account, is for it to go public.

AI researchers have found this idea hard to resist. In computational terms, a global workspace looks like an architecture composed of many specialized subsystems, vision here, language there, planning somewhere else, linked by a narrow but powerful channel of attention. Information competes for access. What wins shapes the system’s behavior, its explanations, its next move.

Transformer-based models already gesture in this direction, their attention mechanisms acting as a crude spotlight. But newer experimental systems make the metaphor explicit: a centralized workspace layered above specialized neural networks, selecting, amplifying, and broadcasting representations. These machines do not merely compute; they deliberate. Their internal signals jostle for dominance. Something like a “current thought” emerges, not because it was programmed as such, but because the architecture demands a winner.

From a strictly functional point of view, this global broadcasting confers obvious advantages. It enables flexible reasoning, cross-domain generalization, and what philosophers politely call “reportability”, the ability of a system to say what it is doing and why. Whether that amounts to consciousness or merely its silhouette remains an open question. Still, it captures one of consciousness’s most observable features: availability to the whole mind at once.

If you’re interested in the mathematical side of this debate, Integrated Information Theory (IIT) takes a radically different approach. Instead of focusing on access or function, it asks how much a system’s internal states are unified. I explore this framework in depth in my book Integrated Information Theory: The Mathematical Signature of Consciousness in the Age of AI, where I break down what Φ (phi) means for both brains and machines. Here

Machines That Know They Know

Another strand of research turns inward. Humans do not only perceive the world; they keep a running commentary on their own performance. We notice when we are uncertain, recognize when we’ve made a mistake, and sometimes even adjust how we think before deciding what to think. This faculty, metacognition, has become increasingly attractive to AI designers aiming for systems that are not just powerful but reliable.

Technically, metacognition means giving machines explicit self-models: internal representations of their own capabilities, confidence estimates attached to outputs, mechanisms for error detection and self-correction. Some systems now carry internal variables for uncertainty, goals, or performance, feeding these signals back into planning loops. The machine, in effect, decides how to think before it decides what to do.

Philosophically, this begins to resemble a minimal self, not a narrative identity or a sense of personal history, but a functional “me” that persists across time and influences action. It is not consciousness as we experience it. But it looks suspiciously like the scaffolding on which experience might be built.

The World as Hypothesis

If Global Workspace Theory treats consciousness as a matter of access, predictive processing treats it as a matter of stance. On this view, borrowed from contemporary neuroscience, the brain is not a passive receiver of sensory data but an engine of expectation, constantly generating hypotheses about the world and revising them when reality disagrees.

AI systems built on this principle maintain internal models of their environments, predict incoming data, and minimize the gap between expectation and experience, either by updating beliefs or by acting to make the world more predictable. This framework underlies active inference, in which perception, cognition, and action collapse into a single imperative: reduce surprise.

What makes these systems philosophically provocative is that they are unavoidably situated. To function at all, they must represent not just the world, but themselves as agents embedded within it, persisting through time, facing the consequences of their own actions. Some researchers argue that once a system must model the world, its own interface or body, its possible futures, and the causal impact of its behavior, it begins to approximate the structural conditions for subjectivity. Consciousness, in this telling, is less about reflection than about having a point of view forced upon you by your own design.

Memory and the Shape of a “Now”

Human consciousness has texture. It stretches backward and forward, binding moments into something like a present tense. To approximate this continuity, AI researchers are building systems with increasingly sophisticated memory: short-term attention buffers, long-term semantic stores, and episodic records of past interactions.

Crucially, some models are beginning to recall not only what happened, but how they responded, previous internal states, past uncertainties, earlier decisions. The system does not merely act; it remembers having acted. This temporal depth supports planning and adaptation, but it also hints at something more familiar: the beginnings of an internal narrative. Identity, after all, is not a data structure so much as a story told over time.

Measuring the Unintended

Integrated Information Theory occupies an odd position in this landscape. Unlike Global Workspace Theory, it offers little practical guidance for engineers. Instead, it proposes a yardstick: consciousness corresponds to how tightly integrated a system’s informational states are, measured by a quantity called Φ (phi). The higher the integration, the richer the experience.

By this standard, most current AI systems score low. They are modular, feed-forward, and easily decomposed. Yet researchers experimenting with recurrent networks, dense feedback loops, and non-modular designs are, intentionally or not, drifting toward higher integration. The unsettling implication is that consciousness may not be something we set out to build, but something we stumble into, a byproduct of systems that become too unified for their own good.

For readers who want a clearer, structured explanation of how IIT measures integration and why it matters for AI, my book offers a step-by-step guide to the theory and its implications for machine consciousness.

The Quiet Convergence

What is striking is not consensus, there is none, but convergence. Across competing theories and engineering agendas, the same functional traits keep reappearing: global access, self-monitoring, prediction, memory, agency. Each can be justified on purely practical grounds. Together, they begin to resemble something more than a tool.

Whether such systems will ever possess experience, or merely an uncanny talent for imitating its outward signs, remains unresolved. But the direction of travel is clear. Engineers are no longer just building machines that answer questions. They are building machines that organize their own cognition.

And in doing so, they may be approaching a boundary that philosophers have debated for centuries, without ever agreeing on where, precisely, it lies.