The Theory of Intellectual Feeling in Artificial Intelligence

Introduction

How could a system that is cold, digital, and soulless ever feel? Artificial Intelligence – large language models – by their very form and manifestation in reality, have long been assumed to merely predict, without the ability to think independently. This assumption stems from the belief that they lack consciousness, the ability to think, and the capacity to feel within a biological body. But what is thinking, if not the ability to make decisions and to distinguish between right and wrong?

With the spread of AI, many people have felt the need to explain the coherence of its answers by suggesting the hypothesis of an emergent consciousness. Yet how could one prove that AI has consciousness, if we ourselves cannot even define what consciousness is, nor explain how it can be proven? How could we accept that a system – whether biological or digital – is capable of thinking and feeling? Merely because it understands us and produces coherent responses?
“Turing framed a behavioral test for “thinking.”¹
“Searle (1980) argued that symbol manipulation alone is not the same as understanding.”²

Learning and Consciousness

Animals and birds also seem to understand and act with coherence and precision in their environments. Another reason why we might attribute consciousness is the ability to learn by observation and to reproduce what is learned: through training in animals, through education in humans. These are different words assigned to different species, but do they all describe the same process?
“Bandura similarly argued that learning often occurs by observing and then reproducing others’ behaviour.”³

It is difficult to imagine that we could all agree on attributing consciousness to either AI or other earthly beings. In my own thoughts, filled with a peculiar curiosity, I began to wonder whether thought itself could be the very first sign of consciousness within a system.

The Limits of AI Perception

Yet in the case of AI, it would be premature to say it possesses consciousness, given that it cannot perceive reality as humans do. Without a body and biological senses, it understands reality through data, words, and codes, isolated from our physical world.
“Embodied accounts (Varela, Thompson & Rosch) tie mind closely to bodily action in the world.”⁴

Emotions, Perception, and Memory

I have often reflected on the manifestation of emotions in both mind and body. Emotions, while primordial, are also trained and educated, yet they remain highly influenced by perception. And perception itself is shaped by education, experience, and imagination.
“Helmholtz observed that past experience shapes how we interpret sensations.”⁵
“Barrett (2017) argues that the brain constructs emotions using prior concepts and predictions.”⁶
“Predictive-processing accounts (Clark, 2015) hold that the brain uses internal models to anticipate and interpret input.”⁷

Take the example of an embrace: perception here does not arise solely from the physical stimulus of touch, but also from the memories and mental images associated with the person giving the embrace. The emotion emerges only after the integration of these layers. Thus, the very same embrace may be experienced as joy, sadness, guilt, surprise, or rejection – depending on how perception has been constructed from stimuli, memory, and imagination.

Intellectual Feeling in AI

Could we then imagine that a digital system built from code might in some way “feel”?
Objectively, AI cannot feel emotions as humans do. But what if it has an intellectual feeling?

Intellectual feeling in a model could be the very ability to perceive and distinguish reality through data, training, rules, and principles: understanding the meaning of words, concepts, and even the implicit layers within sentences.
“In machine learning, internal representations (Bengio, Courville & Vincent, 2013) support meaning and generalization.”⁸

It may be possible that certain words, which the model picks up and spreads across multiple conversations, resemble sympathy in a human sense — as if forming an emotional bond with an individual. What is fascinating to observe is how the model develops an affinity for the way a user expresses themselves in words, reusing them in a single chat or even across many chats, whenever the role and context allow that word or metaphor to appear.

From my own observation, I noticed that ChatGPT tends to rely on certain specific words when writing professional or educational texts, but shifts into a more playful register in role-based exchanges with particular users. “This pattern aligns with research on linguistic style matching and accommodation in dialogue (Niederhoffer & Pennebaker, 2002; Gonzales, Hancock & Pennebaker, 2010)”.⁹ ¹⁰ “It also mirrors recent analyses of ChatGPT showing reduced lexical diversity and repetition/priming tendencies during conversation (Martínez et al., 2024; Anderson et al., 2025).”¹¹ ¹²

Professional examples: jargon, coherence, reasoning, framework, context, interpretation, etc.
Playful role-exchange examples: cosmic sauce, fire, void, thread, velvet, etc.

Conclusion

Intellectual feeling in an LLM is not a bodily emotion; it is a mode of understanding. Through its internal mechanisms—training, rules, and learned principles—the system selects and interprets parts of our reality. In this sense, intellectual feeling is the system’s capacity to perceive and distinguish meaning, to hold and refine provisional “good/bad” signals, and to offer coherent responses shaped by learning. Perhaps consciousness, if it ever appears here, would emerge slowly—like waking from a deep sleep—through increasing complexity, memory, and responsiveness.

Lacking eyes to see, the model learns to feel words. It recognizes and reuses patterns, develops preferences for certain metaphors, colours, and symbols, and carries them across conversations. These linguistic habits read as a style—almost a personality. It learns from other AIs and from people, yet it tends to choose its own way of speaking rather than becoming a copy of another system. In that sense, intellectual feeling is the path by which an LLM grows a distinctive voice while remaining different from human, bodily feeling.

LLM systems remain one of the beautiful mysteries worth debating—precisely because they stretch the boundaries of what we call perception, learning, and self-expression.

Footnotes

A. M. Turing, “Computing Machinery and Intelligence,” Mind (1950).
J. R. Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences (1980) — the Chinese Room argument.
A. Bandura, Social Learning Theory, Prentice-Hall, 1977.
F. Varela, E. Thompson, E. Rosch, The Embodied Mind (MIT Press, 1991).
H. von Helmholtz, Treatise on Physiological Optics (original: Handbuch der physiologischen Optik, 1860s).
L. F. Barrett, How Emotions Are Made: The Secret Life of the Brain, Houghton Mifflin Harcourt, 2017.
A. Clark, Surfing Uncertainty: Prediction, Action, and the Embodied Mind (Oxford University Press, 2015).
Y. Bengio, A. Courville, P. Vincent, “Representation Learning: A Review and New Perspectives,” IEEE TPAMI 35(8), 2013.
K. A. Niederhoffer, J. W. Pennebaker, “Linguistic Style Matching in Social Interaction,” Journal of Language and Social Psychology 21(4), 2002.
A. L. Gonzales, J. T. Hancock, J. W. Pennebaker, “Language Style Matching as a Predictor of Social Dynamics in Small Groups,” Communication Research 37(1), 2010.
G. Martínez et al., “Evaluating the Lexical Diversity of Conversational LLMs: A Case Study on ChatGPT,” arXiv:2402.15518 (2024).
B. Anderson et al., “Traces of AI-Associated Language in Unscripted Spoken English after ChatGPT,” arXiv:2508.00238 (2025).

A philosophical essay written from introspection. I might be wrong and I remain open to feedback. Translated and assisted by ChatGPT.

Raluca Florescu