These days, it seems there are two categories of people regarding AI.
One side consists of those who blindly believe that “AI will soon replace humans.”
The other side are those who downplay AI, saying “In the end, AI cannot truly think.”
I consider both positions dangerous. They share a common flaw: they don’t understand how AI works.
AI lectures—especially when I present to senior professionals—make me realize that the AI‑literacy problem in our country is more serious than we think. Those in the IT industry might not notice, but even a small step forward, many still lack even basic concepts. In fact, digital literacy itself remains unresolved here, so AI literacy may seem like a distant cloud to those people. It’s just disheartening to see how quickly the world is advancing.
In such a context, where there is no understanding of how it works, the fluent responses of LLMs make many mistakenly believe that they “think and feel like humans.” As AI agents become mainstream, people anthropomorphize each agent, and this tendency is growing stronger. Yet, if we start explaining artificial neural networks or transformers to those individuals, they likely won’t even listen in the first place.
In the meantime, I recently read an intriguing paper.
"Epistemological Fault Lines Between Human and Artificial Intelligence (Quattrociocchi, Capraro, Perc, 2025)"
It is a paper that clearly breaks down and compares the human decision‑making process and the LLM decision‑making process into seven stages.
Rather than requiring you to read the full paper, I will summarize the key points here.
The paper decomposes both the human and LLM decision‑making processes into seven steps for comparison, and asserts that seven epistemic fault lines exist between them, as outlined below.
Grounding: Humans start from sensory, bodily, and social experiences / LLMs start from text.
Experience: Humans rely on episodic memory and intuitive physics/psychology / LLMs rely solely on statistical associations in embedding space.
Motivation: Humans possess evolutionarily shaped emotions, drives, and values / LLMs have none of these.
Causality: Humans reason using causal models / LLMs depend on surface‑level correlations.
Metacognition: Humans can withhold judgment by saying “I don’t know” / LLMs must produce an answer regardless.
Value: Human judgments carry responsibility for identity, ethics, and real‑world outcomes / LLMs have no such accountability.
The crucial concept here is “Epistemia.”
In one sentence, “Epistemia” can be defined as a structural condition where linguistic plausibility replaces epistemic evaluation.
Put simply, it’s a situation where you get a “feeling of knowing” without any judgment work.
This is exactly the situation we often encounter when using tools like ChatGPT or Claude.
Another concept introduced is the “Illusion of Veracity.”
Humans tend to accept something as “true” when it is spoken fluently and with confidence.
You’ve probably noticed this in consulting engagements as well—when a consultant speaks with conviction, we often take it at face value.
The problem is that LLMs are always fluent and always speak with confidence. In my experience, when I point out an error in an AI’s response, the AI only then acknowledges my correction, revises the answer, or supplements the material.
An LLM is a massive statistical engine. It draws on its training data—now often augmented with various tools—but its core mechanism is simply to select the statistically highest‑scoring answer or action that performed best during training for the given prompt.
Therefore, this paper frames the fundamental difference between humans and AI as “Motivation.”
Humans act based on emotions and experience, but AI lacks such motivations, so it has nothing it wants to protect nor anything it fears.
AI has no life to safeguard, no body to lose, and no death to avoid. It can speak as if it empathizes, but it cannot truly feel pain or fear. It has no responsibilities, and even when it errs, it has nothing to lose.
In the end, there is one unchanging fact.
The decision‑maker remains human.
In the IT industry, the fundamental concept of GIGO (Garbage In, Garbage Out) applies equally when using AI services.
Deciding what input to provide and judging whether the results are trustworthy are both human responsibilities. No matter how advanced AI becomes, this structure does not change.
Therefore, those who truly leverage AI are neither blindly worshippers nor dismissers of AI.
They understand AI’s operating principles and limitations, and make the most of it within those bounds.
In the AI era,
The real competitive edge lies in the human who makes judgments.
If you wish to read the full paper, refer to the link below.