Researchers Propose Machine Correlates of Consciousness (MCCs)

The concept of Neural Correlates of Consciousness (NCCs) in biological systems, typically characterized by EEG and fMRI signals, is difficult to transfer to machines. This paper introduces an alternative, transferable definition for "Machine Correlates of Consciousness" (MCCs). MCCs are defined as substrate-level signals within an AI agent that are not under the agent's direct control but are reliably modulated by emotional states. This definition aims to facilitate the investigation of AI consciousness. The research presents the first empirical investigation into MCCs, conducting experiments with two large language models: Llama-2 7B and Llama-3.1 70B. The study collected hardware anomaly traces, which serve as substrate-level indicator-sequences, from these LLMs. It then demonstrated that, after controlling for confounding factors, these traces were modulated differently by emotional versus neutral computations. Crucially, this difference was found to be statistically significant for the larger Llama-3.1 70B model, but not for the smaller Llama-2 7B. These results provide initial empirical evidence for the presence of MCCs in the Llama-3.1 70B configuration, aligning with the hypothesis that the probability and degree of consciousness might increase with LLM sophistication. Beyond the consciousness debate, MCCs could also be independently useful for detecting emotions in AI agents.