From Verbal Attribution to Internal Measurement A Cross-System Framework for Comparing Human Brain and Artificial Intelligence Responses to External Stimuli

Abstract Public discussions about whether artificial intelligence possesses emotions, subjective experience, or even consciousness often rely on linguistic output, behavioral performance, and observers’ judgments. An AI system’s ability to describe its own state is not, by itself, sufficient to demonstrate subjective experience. At the same time, the fact that an AI system is non-biological does not by itself justify the conclusion that it lacks all forms of subjective experience. This paper proposes a cross-system experimental framework. When humans and artificial intelligence systems are exposed to functionally corresponding external stimuli, researchers can record brain activity in humans and internal states in AI systems, then compare how the stimuli alter internal processing and how those changes affect subsequent judgments and behavior. The goal is not to require AI systems to reproduce the structure, waveforms, or numerical values of the human brain, but to identify response relationships that can survive repeated testing and causal examination. Such comparison cannot be achieved merely by converting two kinds of data into the same format. Researchers must first establish what candidate measures mean within each system, then construct possible cross-system correspondences and test whether those correspondences can predict changes under new materials and new conditions. Only when a proposed relationship does more than describe existing results, and instead generates successful new predictions, is there stronger reason to discuss whether the two systems may share some common processing principle. Using “violation of an acquired expectation” as an example, this paper illustrates how the framework could be translated into an initial research program. It also distinguishes transient computational states, states maintained within context, and persistent states across sessions, and discusses long-term interaction, consciousness-relevant human controls, internal intervention, and statistical confounds. This paper is a conceptual and methodological proposal. It does not provide an already validated test for consciousness, nor does it claim to have solved all technical problems of cross-system comparison. Even if the results ultimately fail to support the presence of subjective experience in artificial intelligence, data linking external conditions, internal states, and subsequent behavior may still be useful for model evaluation, training, and interaction design.