Asymmetric Communication: Large Language Models and Language Games

The Great Chatbot Confusion: Who's Really Talking?

Imagine you are sitting in a room with a very talented, very fast-talking parrot. This parrot has memorized every book, movie script, and text message ever written. It can mimic human conversation so perfectly that it sounds like it has opinions, feelings, and even a soul. But here is the catch: the parrot doesn't know what it's saying. It's just matching patterns. This is the world of Large Language Models (LLMs), the super-smart AI chatbots that are changing how we work, learn, and play.

The big question everyone is asking is: "Is the parrot actually thinking?" or "Is it just a fancy calculator?" To answer this, we need to understand a few simple ideas. First, think of a language game. Imagine a game like soccer; the ball only makes sense because everyone agrees on the rules and what counts as a goal. In human talk, words only have meaning because we all agree on how to use them and who is responsible for what we say. Second, think of asymmetry. In a normal conversation, both people are players; they both make promises, both get in trouble if they lie, and both have to listen. But with AI, the game is unbalanced. The AI throws the ball, but the human has to catch it, decide if it's a good catch, and take the blame if they trip.

This paper, written by Enzo Fenoglio, dives deep into this unbalanced game. It argues that we are making a huge mistake by treating AI like a human partner. The author suggests that no matter how smart the AI gets, it can never truly "play the game" because it doesn't understand the rules or the consequences. It's not a failure of the machine; it's just how the machine is built. Understanding this helps us stop worrying about robots taking over and start focusing on how we should use them responsibly.

The Paper's Big Idea: The One-Sided Conversation

So, what does this paper actually say? It introduces a concept called Asymmetric Communication. Think of it like a game of catch where one person is a human and the other is a robot that throws balls perfectly but has no arms to catch them back. The robot throws the ball (the AI generates a text), and the human catches it (the human reads and understands it). But here's the twist: the robot never has to say "I'm sorry" if it throws the ball at your head, and it never gets a point for a good throw. Only the human player is keeping score.

The paper argues that when we talk to an AI, we are engaging in a new kind of language game where all the responsibility stays on the human side. The AI is just a machine that produces words based on patterns it learned. It doesn't have beliefs, it doesn't have goals, and it doesn't care if it's right or wrong. It's like a magic 8-ball that gives you a really long, detailed answer, but the 8-ball isn't actually thinking about your question.

The author uses four main ideas to build this picture:

1. Language Games (Wittgenstein): Meaning isn't inside the words; it's in how we use them together. Since the AI doesn't "use" words in real life (it just spits them out), it doesn't really understand them.
2. Receiver-Side Completion (Luhmann): A conversation isn't finished until the listener understands it. The AI can talk all day, but if a human doesn't listen and make sense of it, no conversation happened.
3. Artificial Communication (Esposito): Machines can be part of a conversation without having a mind. They just need to be unpredictable enough to keep us interested, like a deck of cards that deals a new hand every time.
4. Scorekeeping (Brandom): In a real conversation, if you say something, you are "committed" to it. If you lie, you get in trouble. The AI can't be committed to anything because it has no skin in the game.

What the Paper Rules Out (The "No" List)

This paper is very clear about what AI is not. It explicitly rejects the idea that AI is becoming a human-like thinker.

• No "General Intelligence": The paper says we shouldn't believe the hype that AI is becoming a "General Intelligence" (AGI) that can replace human thinking. Just because a plane flies better than a bird doesn't mean the plane is a bird. AI is a different kind of thing entirely.
• No "Hallucinations" as Mistakes: When people say AI "hallucinates" (makes things up), they act like the AI is dreaming or lying. The paper says this is wrong. The AI isn't lying because it doesn't know the truth to begin with. It's just doing its job: guessing the next word. The "mistake" happens because we expect it to know the truth, and it doesn't.
• No "Agency" or "Goals": The paper argues that AI doesn't have goals. When an AI seems to be working toward a goal (like solving a math problem), it's just following a path we set for it. It's not a rebel with a plan; it's a tool being used.
• No "Sentience" or "Feelings": If an AI says "I feel sad," it's not actually sad. It's just mimicking the pattern of someone who is sad. The paper says we shouldn't get attached to AI or worry that it will feel pain. The feelings are all on the human side.

The Three Rules of the Asymmetric Game

The paper defines this one-sided relationship with three specific rules that hold true no matter how smart the AI gets:

1. The Human is the Only Judge: The AI can say anything, but only the human decides if it's correct, useful, or true. The AI doesn't have a "truth detector" inside it.
2. The Human Takes the Blame: If the AI gives bad advice and you follow it, you are responsible for the result. The AI can't be sued, fired, or scolded. It has no "score" to lose.
3. The Human Makes it Real: If you ignore what the AI says, it's just noise. It only becomes "communication" when a human reads it, understands it, and uses it.

Why This Matters for You

The paper suggests that the more powerful AI gets, the more dangerous it is to forget these rules. If an AI sounds too perfect, we might start thinking it's a person. We might trust it too much, or blame it when things go wrong.

For example, the paper talks about "Agentic AI"—systems that seem to act on their own. The author says this is just a fancy name for a human giving a computer a list of tasks. The computer isn't the boss; the human is. If the computer messes up, it's because the human didn't set the rules clearly enough.

The paper also touches on "Alignment," which is the idea of making AI want what humans want. The author says this is the wrong way to think about it. AI doesn't "want" anything. Instead, we need to build better "guardrails" (like speed bumps) to make sure the AI's output is safe for humans to use. It's not about training the AI to be good; it's about training us to use it safely.

The Takeaway

In the end, this paper tells us to stop looking at AI as a mysterious new life form. It's a tool, a very complex and fluent tool, but a tool nonetheless. The magic isn't in the machine; it's in the human who uses it. The machine generates the words, but the human gives them meaning, checks if they are true, and takes responsibility for what happens next.

So, the next time you chat with an AI, remember: you are the only player who is actually playing the game. The AI is just a very talented parrot, and it's up to you to decide what it says means. The paper doesn't say AI is useless; it says it's useful only if we remember that we are the ones holding the reins.