Consciousness Quantification Engine ConsciousNet Released: First Mathematical Model to Measure AI Self-Awareness

The Cognitive Computing Laboratory at the Swiss Federal Institute of Technology in Lausanne (EPFL) today officially released ConsciousNet, the worlds first computational engine capable of quantitatively assessing the consciousness level of AI systems.

From Philosophical Debate to Mathematical Measurement

ConsciousNets core innovation lies in translating two consciousness theories that have long remained in the philosophical domain into a computable mathematical framework. Project lead Matthias Muller, Professor of Computational Neuroscience at EPFL, explained that the team integrated Giulio Tononis Integrated Information Theory (IIT) and Bernard Baars Global Workspace Theory (GWT) to establish an assessment system comprising 12 quantitative metrics.

These 12 metrics cover dimensions including information integration, self-representation capability, attention allocation mechanisms, and temporal continuity perception. Each metric ranges from 0 to 1, and a weighted composite produces an overall Consciousness Level Score (CLS).

Professor Muller emphasized that ConsciousNet does not attempt to answer the philosophical question of whether AI truly possesses consciousness. Instead, it provides an engineering-grade measurement standard for comparing how different AI architectures perform on consciousness-related features.

Test Results Spark Discussion

The research team tested 15 mainstream large language models and 3 multimodal systems. Results showed all tested systems scored below 0.15 on the CLS scale (out of 1.0), far below the human subjects average of 0.73. However, certain individual metrics revealed surprisingly high scores for some AI systems.

On the self-representation dimension, Anthropic's Claude series scored highest at 0.38. On attention allocation, Google's Gemini Ultra scored 0.42. The researchers noted in their paper that these high scores do not mean these systems are closer to consciousness, but rather reflect more sophisticated architectural design in specific cognitive functions.

The most controversial finding was the information integration metric, where an open-source hybrid architecture system called MoSAIC scored 0.51, surpassing all commercial closed-source models. Professor Muller suggested this may relate to MoSAIC's unique cross-modal information fusion mechanism.

Industry Reaction

The release has generated polarized reactions in the AI industry. Supporters believe consciousness quantification will provide important tools for AI safety research. Critics warn that reducing consciousness to a set of numbers could mislead the public. The EU AI Office has indicated it will consider incorporating ConsciousNet metrics into its next-generation AI regulatory framework.

ConsciousNet uses a modular design with input analysis, feature extraction, and scoring aggregation layers. A complete evaluation takes approximately 4 hours and costs about $200 in compute. The team plans to launch a cloud version in Q3 2030, reducing per-evaluation costs to under $20. The paper is available on arXiv, with the codebase to open next month.