Consciousness is the most contested topic in cognitive science. Several serious scientific theories now exist, each making different predictions about which systems are conscious and why. None has been settled, but the debate has matured.
Attention, in the technical sense
Attention is the brain's mechanism for selecting which information to process deeply. It is not the same as consciousness, though the two are tightly coupled. Top-down attention is goal-driven (you look for your keys); bottom-up attention is stimulus-driven (a sudden noise pulls your gaze).
The neural mechanisms involve frontoparietal networks biasing sensory cortex. Transformer attention is named after the cognitive phenomenon but is mathematically a soft weighted lookup — a much narrower thing.
Global Workspace Theory
Proposed by Bernard Baars and developed by Stanislas Dehaene, Global Workspace Theory (GWT) says consciousness arises when information is broadcast across a network of brain regions, making it globally available for reasoning, reporting, and memory. Unconscious processing is local; conscious processing is broadcast.
GWT is one of the only consciousness theories with a plausible computational implementation, and several researchers have argued that suitably architected AI systems could satisfy its conditions.
Predictive processing and active inference
Predictive processing frames perception, action, and cognition as a single Bayesian inference: the brain maintains a hierarchical generative model of the world and updates it to minimise prediction error. Karl Friston's free energy principle generalises this further.
On this view, consciousness might be the high-level prediction the brain makes about itself. It is a beautiful theory but notoriously hard to falsify.
Integrated Information Theory
IIT, developed by Giulio Tononi, identifies consciousness with integrated information (Φ): the amount of cause-effect structure a system has that cannot be decomposed into independent parts. IIT is the most mathematically precise theory and the most philosophically aggressive — it implies many simple systems are minimally conscious, while feed-forward networks (including standard transformers) are not.
IIT and GWT made conflicting predictions in a 2023 adversarial collaboration; the results were mixed, and the debate continues.
The attention schema
Michael Graziano's attention schema theory says the brain builds a simplified model of its own attention, and that this self-model is what we call consciousness. The model is useful for controlling and predicting one's own behaviour; subjective experience is, on this view, the way the model presents itself.
Attention schema theory is one of the easier theories to imagine implementing in an AI, but most researchers think it under-explains the felt quality of experience.
Key terms
- Global Workspace
- A broadcast mechanism that makes information globally available across the brain.
- Prediction error
- The mismatch between expected and actual input; the brain's main teaching signal.
- Free energy principle
- Friston's unifying principle: minimise long-run prediction error.
- Φ (phi)
- IIT's measure of integrated information.
- Attention schema
- An internal model of one's own attention used for control and prediction.
Connects to AGI
Whether AGI systems are conscious — and whether we should care morally — depends on which of these theories is correct. The debate is not a sideshow; it directly affects how we think about AI welfare, deception, and rights.