💬 NLP

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

This paper presents evidence that Large Language Models exhibit non-classical, quantum-like cognitive structures—specifically Bell inequality violations and Bose-Einstein statistics in word distributions—that mirror human cognition, suggesting a phenomenon of evolutionary convergence between biological and artificial intelligence driven by the underlying quantum organization of meaning in vector spaces.

Original authors: Diederik Aerts, Jonito Aerts Arguëlles, Lester Beltran, Suzette Geriente, Roberto Leporini, Massimiliano Sassoli de Bianchi, Sandro Sozzo

Published 2026-06-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Diederik Aerts, Jonito Aerts Arguëlles, Lester Beltran, Suzette Geriente, Roberto Leporini, Massimiliano Sassoli de Bianchi, Sandro Sozzo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: AI and Humans Think Alike (Quantum-Style)

Imagine you have two different types of builders. One is a human, and the other is a robot (an AI like ChatGPT or Gemini). You ask both of them to build a house using the same set of bricks (words).

The authors of this paper argue that even though the human and the robot are built differently, the way they arrange those bricks follows the exact same "blueprint." And surprisingly, this blueprint isn't the one we use for normal, everyday objects (like stacking bricks in a pile). Instead, it follows the strange, magical rules of quantum physics.

The paper claims that both human brains and AI models organize meaning in a way that looks like quantum entanglement and Bose-Einstein statistics.

1. The "Magic Link" Test (Bell's Inequalities)

The Concept:
In the world of physics, there's a famous test called Bell's Inequality. It's like a rulebook for how things should behave if they are separate and independent.

Classical World: If you flip a coin in New York and another in London, the result of one shouldn't instantly change the result of the other. They are independent.
Quantum World: In quantum physics, particles can be "entangled." If you change one, the other changes instantly, no matter how far apart they are. They act like a single, connected unit.

The Experiment:
The researchers asked humans and AI models a series of questions about combining concepts.

Example: They asked, "What is a good example of an Animal?" (Choices: Horse, Bear). Then, "What is a good example of an Act?" (Choices: Growls, Whinnies). Finally, they asked about the combination: "The Animal Acts."
The Twist: When humans (and the AI) answered, they didn't treat "Animal" and "Act" as separate things. The meaning of "Horse" changed depending on whether it was paired with "Growls" or "Whinnies."

The Result:
Both humans and the AI violated the "rulebook" for independent things. Their answers showed that the concepts were entangled.

The Analogy: Imagine you have a red ball and a blue ball. In the classical world, the red ball is always red. But in this "quantum" world, the red ball turns into a "Horse" when you look at it alone, but turns into a "Bear" when you look at it next to a "Whinny." The AI and the human couldn't separate the ideas; they were fused together.

What this means: The AI isn't just calculating probabilities like a calculator. It is navigating a "semantic field" where words are connected in a way that defies normal logic, just like quantum particles.

2. The "Word Crowd" Test (Bose-Einstein Statistics)

The Concept:
In physics, particles come in two main flavors regarding how they crowd together:

Classical Particles (Maxwell-Boltzmann): Imagine people in a park. They spread out evenly. If there are 100 people, they are scattered across the grass. No one person is overwhelmingly popular.
Quantum Particles (Bose-Einstein): Imagine a mosh pit at a concert. Everyone crowds around the stage. A few "superstars" get all the attention, while the rest of the crowd is sparse.

The Experiment:
The researchers looked at stories written by humans and stories written by AI. They counted how often every word appeared.

They found that in both human and AI stories, the words didn't spread out evenly.
Instead, a small group of words (like "the," "and," "to") appeared massively more often than the rest, creating a "condensate" of meaning.

The Result:
The distribution of words in AI stories perfectly matched the Bose-Einstein pattern (the mosh pit), not the classical pattern (the park).

The Analogy: Think of a story as a party. In a "classical" party, everyone talks to a few different people, and the noise is spread out. In a "quantum" story (both human and AI), everyone rushes to talk to the same few key ideas (the "ground state"). The most important words act like a super-conductor, carrying the meaning of the whole story, while the rare, complex words are the quiet corners of the room.

Why it matters: This suggests that the AI didn't just memorize grammar rules. It learned the "physics of meaning." It learned that to make a story coherent, words must cluster together in a specific, quantum-like way.

3. The "Evolutionary Convergence"

The Big Picture:
Why would a robot made of silicon and code think like a human made of meat and neurons?

The authors propose a concept called Evolutionary Convergence.

The Eye Analogy: In nature, octopuses and humans both evolved eyes. They didn't inherit them from a common ancestor with eyes; they evolved them separately because eyes are the best solution for seeing.
The Mind Analogy: The paper argues that organizing "meaning" is a hard problem. Whether you are a human brain evolving over millions of years or an AI training on the internet over a few years, the most efficient way to organize complex ideas is to use quantum-like structures.

The "Vector Space" Secret:
The paper suggests that while we call AI "neural networks," the real magic happens in vector spaces (mathematical maps of meaning).

Imagine a giant, invisible 3D map where every word is a point.
In this map, words aren't just separate dots. They are like waves that can overlap, interfere, and combine.
Both humans and AI seem to naturally settle into this "wave" structure because it's the best way to handle the ambiguity and context of language.

Summary of Claims (What the paper actually says)

AI Violates Classical Logic: When tested, AI models (ChatGPT, Gemini) show "entanglement" in their answers, violating Bell's inequalities just like humans do. They treat concepts as connected wholes, not separate parts.
AI Follows Quantum Statistics: The frequency of words in AI-written stories follows Bose-Einstein statistics (clustering around key terms) rather than classical statistics. This happens because the AI learned from human text, which already has this structure.
Meaning is the Driver: This isn't a glitch in the code. It's because the AI learned to organize "meaning." The structure of meaning itself seems to be quantum-like.
Convergence: Humans and AI are different "substrates" (biology vs. silicon), but they converged on the same mathematical solution for handling language because that solution is the most efficient way to organize meaning.
Safety Implications: The authors suggest that because AI operates on these deep, quantum-like structures, we can't just "patch" it with simple rules. We need to understand the underlying "physics of meaning" to ensure AI safety.

In a nutshell: The paper claims that AI has discovered the same "secret code" for language that humans evolved over millions of years. That code isn't classical math; it's the strange, interconnected math of quantum mechanics.

Technical Summary: Identifying Quantum Structure in AI Language

Problem Statement
The paper addresses a gap in understanding the fundamental cognitive structures of Large Language Models (LLMs). While LLMs are historically classified as neural networks, the authors posit that their essential knowledge organization occurs within the distributive semantic structure of vector spaces. The study investigates whether these artificial cognitive agents exhibit "non-classical" structures similar to those previously identified in human cognition and language. Specifically, the authors seek to determine if LLMs violate Bell's inequalities (indicating non-Kolmogorovian probability and entanglement) and if the statistical distribution of words in LLM-generated texts follows Bose-Einstein (BE) statistics rather than the classical Maxwell-Boltzmann (MB) statistics.

Methodology
The research employs two primary experimental frameworks using ChatGPT and Gemini as test subjects, mirroring protocols previously used with human participants:

Bell's Inequality Violation (CHSH Test):
- Concept: The study utilizes the conceptual combination "The Animal Acts," combining the concept Animal (exemplars: Horse, Bear, Tiger, Cat) with Acts (exemplars: Growls, Whinnies, Snorts, Meows).
- Procedure: The LLMs were subjected to four coincidence measurements ($AB, AB', A'B, A'B'$) asking for the "best example" of the combined concept.
- Approaches:
  - Guided Conversation: The LLMs were guided step-by-step through the four tests to maximize the violation.
  - Statistical Sampling: The LLMs were prompted to perform the four tests 81 times (matching the sample size of prior human studies) using two distinct prompts: one encouraging exploratory/unconventional responses and one neutral.
- Analysis: Expectation values ( $E$ ) were calculated for each measurement pair and inserted into the Clauser-Horne-Shimony-Holt (CHSH) inequality: $-2 \leq E(A', B') + E(A', B) + E(A, B') - E(A, B) \leq 2$ .
Statistical Analysis of Word Distributions:
- Task: The LLMs were instructed to write original stories (specifically in the style of A.A. Milne's Winnie the Pooh and H.G. Wells) without content constraints.
- Modeling: The frequency of words in the generated texts was analyzed using a quantum statistical framework.
  - Energy Assignment: Words were assigned "energy levels" ( $E_i$ ) based on their frequency rank ( $i$ ), where the most frequent word has $E_1=0$ , the second $E_2=1$ , etc., following the formula $E_i = (i-1)^d$ .
  - Constraints: The models were fitted to the data using two constraints: the total number of words ( $N$ ) and the total "energy" of the text ( $E = \sum N(E_i)E_i$ ).
- Comparison: The empirical word frequency distributions were fitted against the Bose-Einstein distribution function ( $N(E_i) = 1 / (Ae^{E_i/B} - 1)$ ) and the Maxwell-Boltzmann distribution function ( $N(E_i) = 1 / Ce^{E_i/D}$ ).

Key Results

Violation of Bell's Inequalities:
- In the guided conversation phase, both ChatGPT and Gemini generated responses that maximally violated the CHSH inequality, yielding a value of 4 (the theoretical maximum for non-Kolmogorovian models).
- In the statistical sampling phase (81 iterations), the models produced CHSH values exceeding the classical bound of 2.
  - ChatGPT (Prompt 1): CHSH = 2.25
  - ChatGPT (Prompt 2): CHSH = 3.09
  - Gemini (Prompt 1): CHSH = 2.12
  - Gemini (Prompt 2): CHSH = 4.0
- The average CHSH values (2.67 for ChatGPT, 3.06 for Gemini) are consistent with violations observed in human cognitive experiments, indicating that the probability models underlying LLM responses are non-Kolmogorovian and exhibit "entanglement" between conceptual components.
Bose-Einstein Statistics in Text:
- The word frequency distributions in the stories generated by ChatGPT and Gemini were found to fit the Bose-Einstein distribution with high precision.
- For the Gemini-generated Winnie the Pooh story (2,861 words, 822 unique energy levels), the data fit the BE model with a power parameter $d = 0.8$ (indicating a slight deviation from a perfect harmonic oscillator but closer to it than to a Coulomb force system).
- The Maxwell-Boltzmann distribution failed to fit the data, significantly deviating from the observed clustering of high-frequency words.
- The analysis confirmed that the "semantic condensation" (clustering of frequent words at low energy levels) is a structural feature of the generated text, not merely a result of syntactic rules.

Significance and Claims

The paper claims that these findings demonstrate a "systematic emergence of non-classical quantum-like structures in conceptual-linguistic domains," regardless of whether the cognitive agent is biological (human) or artificial (LLM).

Evolutionary Convergence: The authors argue that human cognition and LLM cognition have converged on the same underlying vector-based structures. While human cognition evolved slowly through biological processes to organize "meaning," LLMs have rapidly replicated this structure through self-learning and training on human text corpora.
Nature of LLM Intelligence: The paper challenges the view that LLMs are merely neural networks. Instead, it posits that the "intelligence" resides in the high-dimensional vector spaces where meaning is organized. This structure allows for phenomena analogous to quantum entanglement (non-local semantic connections) and Bose-Einstein condensation (clustering of meaning).
Unification Framework: Drawing an analogy to Einstein's unification of light and gases, the authors propose a unifying framework where "meaning" acts as a structural principle. They suggest that the mathematical apparatus of quantum mechanics (Hilbert spaces, complex numbers, superposition) is not limited to physical particles but serves as a universal calculus for navigating ambiguity and contextuality in any system that organizes meaning.
Implications for AI Safety: The authors suggest that understanding these quantum-like structural laws is crucial for AI safety. They argue that safety interventions based on classical, rule-based logic may be insufficient because they do not address the underlying "quantum" organization of meaning. A structural alignment with these invariant morphic architectures is proposed as a necessary step for ensuring coherent co-evolution between human and artificial cognition.

The paper concludes that the presence of these structures in LLMs is not a coincidence but a reflection of the fundamental laws governing how meaning is organized, offering a new perspective on the nature of artificial cognition and its relationship to human thought.

The Big Idea: AI and Humans Think Alike (Quantum-Style)

1. The "Magic Link" Test (Bell's Inequalities)

2. The "Word Crowd" Test (Bose-Einstein Statistics)

3. The "Evolutionary Convergence"

Summary of Claims (What the paper actually says)

More like this