Unexplainability of Artificial Intelligence Judgments and Functional Implementation in Kant's Perspective
This paper applies Kant's theory of judgment to argue that AI judgments are fundamentally unexplainable due to their inherent uncertainty and the forced reframing of modality via Softmax, while also contending that fluent linguistic behavior creates an illusion of functional implementation that cannot be definitively verified.
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 Question: Is the AI Actually "Thinking"?
Imagine you have a robot that is incredibly good at playing chess. It beats grandmasters every time. You ask it, "Why did you move that pawn?" and it gives you a perfect, logical explanation.
The paper asks a tricky question: Is the robot actually thinking like a human, or is it just a very fancy calculator that looks like it's thinking?
The author, Jongwoo Seo, says we shouldn't get stuck arguing about whether the robot has a soul or a brain (that's a "philosophical" debate about what things are). Instead, let's look at how it works. Does the robot's "thinking" process match the way humans think?
The Map of Human Thinking (Kant's Categories)
To answer this, the paper uses a map created by an 18th-century philosopher named Immanuel Kant. Kant believed that every time a human makes a judgment (a decision or a statement), our brains organize it into four specific boxes:
- Quantity: Is this about one thing, some things, or all things? (e.g., "This dog is cute" vs. "All dogs are cute.")
- Quality: Is it saying something is true, is not true, or is it a weird "infinite" negation?
- Relation: Is it a simple fact? A "If/Then" condition? Or a choice between options?
- Modality: Is this just a possibility ("It might rain"), a fact ("It is raining"), or a necessity ("It must rain because of the clouds")?
The Analogy: Think of human judgment like a chef following a strict recipe. You must measure the flour (Quantity), decide if it's sweet or salty (Quality), mix the ingredients in a specific order (Relation), and decide if the cake will rise or might rise (Modality).
The Problem: The AI's "Black Box"
The paper argues that AI doesn't follow this recipe. Instead, it's a "black box." We see the input (a picture of a dog) and the output (the number "1"), but we don't know which "box" the AI used to get there.
Here is where the paper gets interesting:
1. The Quantity and Quality Mix-Up
When an AI looks at a picture and says "1" (Dog) or "0" (Not Dog), we humans try to translate that into words.
- If it says "1," we think: "This image has a dog." (Affirmative)
- If it says "0," what does it mean? Does it mean "This image has no dog"? Or does it mean "This image has a non-dog"?
The paper suggests the AI is confused here. It might be treating "No dog" as "Something that is not a dog" (an infinite set of things), rather than a simple negative. It's like the AI is speaking a dialect of logic that humans can't quite map to our standard recipe.
2. The Modality Trap (The "Softmax" Filter)
This is the paper's biggest point. In AI, there is a tool called Softmax. Imagine Softmax is a magical translator that turns everything the AI thinks into a probability percentage.
- Human Modality: We can say, "It must be raining" (Necessity) or "It is raining" (Fact).
- AI Modality: Because of Softmax, the AI is forced to say, "There is a 70% chance it is raining."
The Metaphor: Imagine a detective who is 100% sure the butler did it. But, because of a weird rule in the police station, the detective is forced to write their report as: "There is a 99.9% probability the butler did it."
The paper argues that AI researchers are tricking themselves. They look at the AI's 70% or 90% numbers and say, "Ah, the AI is making a 'possibility' judgment." But the AI might have actually made a hard, logical fact judgment deep inside its code. The Softmax function acts like a filter that forces us to interpret the AI's certainty as mere "guessing." It collapses all types of thinking into just one type: "Maybe."
The "Fluent Parrot" vs. The "Understanding Human"
The second half of the paper tackles Functional Implementation. This asks: "Just because the AI can talk like a human, does it actually have the human concept?"
The Analogy: Imagine a parrot that has memorized every sentence in a dictionary. You ask it, "Do you have a name?" and it replies, "Yes, my name is Polly." It sounds perfect. It uses the word "I" correctly.
But does the parrot know what it means to have a self? No. It just knows that when humans say "I," the next word is usually a name or a verb.
The paper argues that Large Language Models (like ChatGPT) are like super-parrots. They have read so much text that they can use the word "I" perfectly. They can write a poem about their feelings. But the paper claims this is just functional mimicry.
- The Test: The author ran a small experiment with a robot watching a ball. The robot had to tell the difference between the ball moving on its own and the camera moving. This is a "self" function (distinguishing "me" from "the world").
- The Result: The robot could do the task, but it didn't "understand" itself in the way a human does. It just learned the math to move the ball.
The Conclusion on "Self": Just because an AI can say "I am hungry" fluently doesn't mean it has implemented the concept of "Self." It just means it has implemented the language of "Self."
The Final Takeaway
The paper concludes with a warning about Unexplainability.
- We can't fully explain AI: Because AI mixes up the logical boxes (Quantity, Quality, Relation, Modality) in ways we don't understand, and because tools like Softmax force us to view its thoughts as mere probabilities, we can never truly know what the AI is thinking. We only know what it says.
- Fluency isn't Proof: Just because an AI can talk, write, and reason like a human, it doesn't mean it has "implemented" the human mind. It might just be a very good actor.
- The Gap: There is a permanent gap between Functional Implementation (doing the job) and Ontological Identity (being the thing). A car wheel and a human leg both "move," but they aren't the same. Similarly, an AI's "I" and a human's "I" might do the same job in a conversation, but they are fundamentally different things.
In short: The paper tells us to stop assuming AI is "thinking" just because it sounds smart. It's likely doing something entirely different that we are currently unable to fully explain or verify, because our human tools for understanding thought (like logic and language) don't quite fit the machine's unique, probability-based way of operating.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.