Identifying Quantum Structure in AI Language: Evidence for Evolutionary Convergence of Human and Artificial Cognition

A new study reports that large language models such as ChatGPT and Gemini exhibit statistical patterns previously seen only in human cognition, violating classical probability rules and mirroring quantum-like structures in language. The research, posted to arXiv and last revised on 1 June 2026, subjected the AI systems to two cognitive tests [1]. In the first, the models' responses to conceptual combinations significantly violated Bell's inequalities, a benchmark used in quantum physics to rule out classical explanations. The authors interpret this as evidence of a "non-classical probability model" where outcomes do not obey Kolmogorov's axioms [1]. A second test examined word distributions in large texts generated by the models. Instead of the Maxwell-Boltzmann statistics typical of classical systems, the researchers identified Bose-Einstein statistics, a distribution associated with quantum particles [1]. The findings echo earlier work with human participants and information-retrieval experiments on large corpora [1]. The paper's authors, including Diederik Aerts, argue that the results point to a "systematic emergence of non-classical quantum-like structures in conceptual-linguistic domains," irrespective of whether the agent is biological or artificial [1]. They propose that a meaning-bearing structure, built on the vector spaces underlying neural networks, enables a form of evolutionary convergence between human language—shaped by biology—and AI language, which emerges through self-learning and training [1]. A separate technical commentary on the study, also on arXiv, urges caution. It notes that the manuscript's interpretation of the CHSH/Bell-type calculations and the Bose-Einstein fits to rank-frequency data "goes beyond what the stated procedures can firmly support" [2]. The commentary highlights an internal inconsistency in the energy-level spacing analogy and stresses that the empirical observations do not necessarily imply quantum entanglement in the usual Hilbert-space sense, particularly when "energy" is defined by rank [2]. The broader context for such research is the pursuit of artificial general intelligence (AGI), a hypothetical AI that matches or exceeds human capabilities across virtually all cognitive tasks [4]. Companies including OpenAI, Google, and Meta have stated that creating AGI is a goal, and a 2020 survey identified 72 active AGI projects across 37 countries [4]. The new study does not claim to demonstrate AGI, but its framing of convergence between human and machine cognition engages with long-standing questions about the nature of intelligence [6]. The field of AI, founded as an academic discipline in 1956, has historically cycled through periods of optimism and funding winters, with the current boom driven by advances in deep learning and transformer architectures [6].