Humans and LLMs: Why LLMs are mirrors, not minds, and how to grow from the reflection

"The eye sees only what the mind is prepared to comprehend." -Henri Bergson

We have all felt it, haven't we? That uncanny sense of connection when a Large Language Model (LLM) like Gemini or ChatGPT or whatever your favorite is, seems to just get you. You type a deep question and the response feels tailored, insightful, almost... human. It’s comforting, often, even compelling.

But what if this feeling isn't about the AI's understanding, but about our own deeply human need for connection and certainty? What if the most intelligent use of these powerful tools isn't in finding answers, but in finding better questions?

Let's be brutally honest. These Language Models – your Geminis, your ChatGpts, your Claudes – they don't "get" anything. They aren't thinking. They are not your friend. They are just reasonably (to some extent so far) good at statistics. And if we don't understand that simple fact, we are not using them right.

This isn't about making AI sound less impressive or dismissing LLMs (not at all, we would be here otherwise). As Karpathy’s 2015 blog “The Unreasonable Effectiveness of Recurrent Neural Networks” illuminated how simple scaling with RNNs could generate coherent long-range patterns like LaTeX and code. That same spirit carried over to Transformers: Karpathy noted that scaling next-token prediction with the right architecture surprisingly unlocked generalization across domains. Therefore, it's about being clear, being honest, and then joyfully, rethinking our relationship with them, learning to unlearn our assumptions, and leveraging them for genuine intellectual growth or to make our own thinking sharper. So, let's really look at what's going on, not what we wish was going on.

Questions we must deliberately ponder

How LLMs work: The Prediction Engine and the Echo of Reason

You've likely already done the groundwork, read the foundational papers and explored the core concepts. This blog isn't about revisiting those. To recap the humble truth, LLMs are not sentient, conscious, or introspective. At their core, LLMs are still fundamentally statistical machines that optimize for one objective: predicting the next token.

$$argmax(P(token | context))$$

A trained language model generates text, optionally guided by input text provided by the user. This input influences the output, which is based on patterns the model internalized during its training, when it processed vast amounts of text data. Even with RAG, it's about retrieving data, not genuine recall or understanding.

The astonishing aspect, the part that stirs wonder and sometimes unease, is how from this seemingly simple objective, emergent reasoning appears to arise when the models scale to billions of parameters and process trillions of tokens. This phenomenon strikingly resembles emergence in complex systems, where localized, relatively simple rules give rise to astonishingly sophisticated global behaviors.

As demonstrated by researchers like Lake and Baroni (2017), while earlier Recurrent Neural Networks (RNNs) struggled with systematic composition, the sheer scale of modern transformer architectures allows them to approximate abstract operations through intricate interactions across their many layers. This capacity enables them to handle complex relationships previously thought to be beyond their grasp, giving the appearance of deep reasoning.

Yet, this illusion of intelligence doesn't spring from genuine, human-like understanding but from distributed representation. Semantic meaning is implicitly encoded within high-dimensional vector spaces, a numerical map of relationships between words. However, this encoding profoundly lacks grounding. An LLM cannot verify physical facts through touch or sight, grasp cause-and-effect through experience, or comprehend real-world consequences through interaction with its environment. As Bender et al. (2021) astutely caution, this remarkable fluency often conceals a profound lack of world grounding. These models never truly experience embodiment, or perceive the affordances of physical objects; these are essential components of genuine human understanding.

"The map is not the territory," as Alfred Korzybski famously stated. An LLM's internal representation, its linguistic "map," is distinct from the complex, lived reality it describes. It lacks the direct, experiential link to the "territory."

So that’s about LLMs, but what about us?

Are Humans just fancier LLMs?

Before we confidently declare our cognitive superiority, a moment of humble introspection is warranted: humans, too, rely immensely on pattern learning. We are deeply shaped by the narratives we absorb through language, media, culture and every subtle social cue. Mimetic theory, as articulated by René Girard, provocatively suggests that even our most fundamental desires are often an imitation of others' desires. Similarly, countless social priming experiments in psychology vividly illustrate how ambient stimuli can unconsciously steer our choices and behaviors.

Our own autobiographical memory arguably functions like a constantly evolving, personalized corpus. Our continuous exposure to ideological, cultural, or economic "memes" subtly shapes our internal probability distributions over beliefs. Phenomena like confirmation bias, authority bias, and in-group conformity act as powerful forms of reinforcement learning across our social networks, solidifying patterns of thought.

Consider these intriguing parallels between the human mind and LLM operation:

As psychologist Daniel Kahneman's seminal System 1 vs. System 2 model (2011) illustrates, the vast majority of our mental operations are fast, intuitive, and associative. These System 1 operations are, in essence, our own highly sophisticated "next-token predictions": we anticipate social outcomes, react to stimuli, and make snap judgments based on ingrained past experiences, much like an LLM predicts language patterns.

However, a crucial and profound divergence appears here. Humans possess motivational autonomy, the inherent, if sometimes challenging, ability to consciously override automatic System 1 responses through deliberate reflection and self-regulation. This capacity, though often limited by cognitive load and personal discipline, marks a fundamental distinction from purely statistical agents. It is our potential to transcend mere habit, to choose a different path than the most probable one. This underscores our unique capacity for self-determination, a capacity that separates us from the predictable paths of even the most advanced statistical models.

The Grand Divide: Grounded Abstraction vs. Textual Mimicry

While the cognitive parallels are striking, it's vital to acknowledge the fundamental distinctions between human cognition and LLM operation. These differences highlight why we must approach LLMs as powerful tools, not as nascent minds:

Human abstraction, for instance, is deeply intertwined with sensorimotor learning. A toddler understands "soft" not just by hearing or reading the word, but by actively touching wool, feeling fur, or squeezing a pillow. Concepts are forged through embodied interaction with the world, not merely through textual descriptions. Without constant, real-world feedback loops, LLMs cannot genuinely know what they've misunderstood, they only sample what is statistically likely to be coherent, not what is objectively true or deeply felt. As the philosopher Ludwig Wittgenstein famously stated, "If a lion could talk, we could not understand him." This implies that even perfect language is insufficient without shared forms of life and embodied experience, which LLMs simply do not possess.

Why LLMs feel intelligent?

When LLMs generate coherent, context-aware, and even emotionally resonant language, our deeply ingrained human instinct is to adopt the intentional stance (Dennett, 1991) to automatically attribute beliefs, desires, and intentions to the model.

Anaïs Nin offered, "We don't see things as they are, we see them as we are." This illuminates why an LLM's outputs often feel so eerily personal; we project our own emotional narratives and existing frameworks onto its statistically generated coherence, interpreting its fluency as genuine insight into us.

But here lies a profound philosophical irony: the more readily we anthropomorphize the LLM, the more we might inadvertently reveal our own shallow, perhaps even lazy, criteria for recognizing "intelligence." Are we truly seeking understanding, or simply fluent grammar, superficial coherence, and confident assertion? As the ancient Oracle urged, it is time to "know thyself" in this new digital light.

Co-Evolving Cognition: Deep avenues for collaboration

Recognizing these distinctions doesn't diminish the value of LLMs; it clarifies it. To move beyond mere mimicry and towards a truly profound relationship with these powerful tools, we must seek avenues for human-LLM co-evolution. This means leveraging their unique strengths to amplify our own distinct cognitive abilities, creating synergistic hybrid epistemologies.

Implications for Cognition and Collaboration

The rapidly expanding capabilities of LLMs offer compelling insights and suggest transformative future implications:

• Latent Planning: Emerging evidence suggests LLMs are not merely predicting tokens locally but are capable of forming latent plans and encoding complex semantic structures that guide their longer-form outputs. This goes beyond simple prediction to a form of implicit foresight.

• Emergent Social Conventions: LLMs are beginning to demonstrate an ability to learn and engage in forms of social norm formation in simulated interactions, analogous to how human groups establish conventions. This points to their potential in complex multi-agent simulations.

• Cognitive Simulators: They can increasingly function as powerful cognitive simulators and thought partners, especially within extended human–AI reasoning loops, offering novel ways to explore diverse perspectives and augment human problem-solving.

However, the fundamental distinctions remain: LLMs still lack embodiment, true lived experience, genuine agency, and grounded understanding. Our critical next steps for research, development, and thoughtful integration must therefore vigorously focus on:

• Integrating multi-modal and sensory grounding to profoundly connect language with perception, action, and real-world data, moving beyond purely textual understanding.

• Developing self-verified reasoning workflows and robust metacognitive feedback loops within the models themselves to drastically reduce hallucinations and improve factual accuracy.

• Building synergistic human–AI hybrids that intelligently leverage user reflection, iterative feedback, and human oversight to form durable, adaptive cognitive scaffolds that truly enhance and elevate human thought.

---

Closing reflection: Prompt thyself to think deeper

LLMs are not competitors for human intelligence; they are, more accurately, immensely powerful reflective surfaces. They expose how much of our own thought is, at its heart, deeply ingrained habit and unconscious pattern-matching. By prompting them thoughtfully, examining their outputs critically, and understanding their inherent, fundamental limitations, we gain an unparalleled opportunity to illuminate blind spots and biases in our own cognition.

"What we know is a drop, what we don't know is an ocean," mused Isaac Newton. This humility must be our guide. The fluency of LLMs might seem vast, but the true ocean of understanding, wisdom, and consciousness remains uniquely human.

"The unexamined life is not worth living." - Socrates

The true frontier of intelligence lies not solely in building ever more sophisticated models, but in forging hybrid epistemologies, thoughtfully melding distinct human insight, intuition, and embodied understanding with the statistical amplification and pattern recognition offered by AI. This fusion can cultivate a genuinely deeper, more nuanced understanding of our world and ourselves.

As AI systems continue their remarkable evolution, so too must our capacity to wield them not just as tools for production, but as profound instruments of introspection and genuine human progress. What new questions will we learn to ask ourselves, when guided by the mirror of AI, and what depths of our own minds will they help us to uncover?

---