Testing Consciousness Theories on Artificial Intelligence: Ablations and Functional Dissociations

Can artificial agents serve as testbeds for evaluating competing theories of consciousness? Can We Test Consciousness Theories on AI? Ablations, Markers, and Robustness, authored by Yin Jun Phua, demonstrates that synthetic neuro-phenomenology, constructing artificial agents that embody consciousness mechanisms, reveals that Global Workspace Theory, Integrated Information Theory, and Higher-Order Theories describe complementary functional layers rather than competing accounts.

Synthetic Neuro-Phenomenology: A Novel Experimental Approach

Yin Jun Phua adopts a synthetic neuro-phenomenology approach that addresses a fundamental challenge in consciousness research. Competing theoretical camps, including Global Workspace Theory (GWT), Integrated Information Theory (IIT), and Higher-Order Theories (HOT), each propose distinct neural signatures for consciousness.

Rather than relying on biological experiments that cannot precisely isolate individual mechanisms, Phua constructs artificial agents that embody these theoretical mechanisms. This approach enables precise architectural ablations, systematic manipulations of specific components, to test their functional consequences in a controlled manner.

The research reports dissociations across three experiments, suggesting that these theories describe complementary functional layers rather than mutually exclusive explanations of consciousness.

Experiment 1: Higher-Order Theories and Metacognitive Calibration

The first experiment tests Higher-Order Theories (HOT), which propose that consciousness requires meta-cognitive monitoring of first-order perceptual states.

Phua implements a no-rewire Self-Model lesion that ablates the metacognitive component while preserving first-order task performance. This lesion abolishes metacognitive calibration, the ability to accurately assess confidence in perceptual judgments, while maintaining successful object recognition and classification.

The result produces a synthetic blindsight analogue, a condition where agents perform visual tasks without metacognitive awareness. This dissociation confirms HOT predictions that consciousness depends on higher-order self-monitoring distinct from first-order sensory processing.

Experiment 2: Global Workspace Theory and Information Access

The second experiment evaluates Global Workspace Theory (GWT), which posits that consciousness arises from a broadcast mechanism that makes information globally accessible to multiple cognitive processes.

Phua manipulates workspace capacity, testing how complete and partial workspace lesions affect information access markers. A complete workspace lesion produces qualitative collapse in access-related markers, demonstrating that broadcast capacity is causally necessary for conscious information availability.

Partial workspace reductions show graded degradation, consistent with GWT’s ignition framework where sufficient broadcasting capacity enables global access while insufficient capacity leaves information locally confined.

Experiment 3: Broadcast Amplification and System Robustness

The third experiment uncovers a broadcast-amplification effect, revealing unexpected fragility in GWT-style architectures. GWT-style broadcasting amplifies internal noise, creating extreme sensitivity to perturbations.

The B2 agent family, which does not rely on global workspace broadcasting, demonstrates robustness to the same latent perturbation. This robustness persists even in a Self-Model-off workspace-read control condition, cautioning against attributing the effect solely to self-model compression.

Phua also reports an explicit negative result. Raw perturbational complexity (PCI-A), a metric derived from IIT-adjacent approaches, decreases under the workspace bottleneck. This finding cautions against naive transfer of IIT-adjacent proxies to engineered agents, highlighting the need for careful validation of consciousness markers in artificial systems.

Hierarchical Design Principle: GWT Provides Capacity, HOT Provides Quality Control

The dissociations across experiments suggest a hierarchical design principle. Global Workspace Theory provides broadcast capacity, enabling information to reach multiple cognitive processes. Higher-Order Theories provide quality control, monitoring and calibrating the reliability of that information.

This complementary relationship implies that consciousness may require both global access mechanisms and meta-cognitive self-monitoring. Neither mechanism alone fully captures conscious experience, but their integration creates the functional architecture necessary for human-like awareness.

Comparison to the ACM Project

The Artificial Consciousness Module (ACM) project develops layered simulations with meta-awareness and dynamic self-modeling. Phua’s synthetic neuro-phenomenology offers empirical methods for testing ACM’s architectural choices.

1. Testing ACM’s Workspace Implementation

ACM processes sensory inputs through multimodal units that feed into a consciousness core. Applying workspace capacity manipulations could determine whether ACM’s architecture suffers from broadcast-amplification effects observed in GWT-style systems.

2. Metacognitive Calibration in ACM

ACM includes meta-awareness modules that aggregate focus and intention data. Implementing no-rewire Self-Model lesions could isolate whether ACM’s meta-awareness genuinely supports metacognitive calibration or merely processes first-order task performance.

3. Robustness and Noise Amplification

Phua’s findings on broadcast-amplification effects raise questions about ACM’s robustness to internal perturbations. Evaluating ACM under controlled noise conditions could reveal whether its architecture exhibits similar fragility.

4. Avoiding Naive Metric Transfer

The negative result on PCI-A highlights the risks of applying consciousness metrics from biological systems to artificial agents. ACM development should validate any proposed consciousness markers through systematic ablation studies rather than assuming direct transferability.

Critical Methodological Insight: Agents Are Not Conscious

Phua emphasizes a critical point. The agents constructed in this research are not conscious. They serve as reference implementations for testing functional predictions of consciousness theories.

This distinction clarifies that demonstrating functional signatures predicted by consciousness theories does not entail that the agent experiences subjective awareness. The research provides a method for evaluating theoretical claims without conflating functional mechanisms with phenomenal consciousness.

For detailed analysis of the experiments and ablation methodologies, access the full paper here.