In the film Close Encounters of the Third Kind, humans and aliens communicate through lights, sounds, and gestures, creating a shared language in the process. The interaction scene, which is the climax of the film, is not about words as much as it is about focus and connection—a kind of conversation where meaning emerges through mutual gestures. A chilling version of this same type of interaction can be found in A Quiet Place, where pointing becomes a survival tool. In a world where sound attracts deadly creatures, silent gestures are vital for directing attention to dangers or opportunities.
These scenes capture something essential about how we can understand AI systems like ChatGPT. Critics have called such systems “stochastic parrots,” suggesting they merely regurgitate patterns of language without understanding. Others have questioned whether LLMs can eventually function as logical reasoners, chaining together thoughts and propositions like a philosopher. Neither description is quite accurate. A better way to understand these systems is to think of them as partners in a pointing game: a collaborative process of navigating meaning. To do that, we exchange signs, as we would with any other person or even with an alien species. What we need to appreciate is that the signs we are exchanging are indexes, not symbols.
What Is the Pointing Game?
When I interact with an AI like ChatGPT, I initiate a kind of back-and-forth exploration. My input—a question, a phrase, or a single word—acts as a pointer. The AI responds by “pointing” back, highlighting concepts, patterns, or phrases it deems relevant based on its probabilistic training. This exchange continues until we’ve constructed something meaningful together.
The pointing game is neither mindless mimicry nor deliberate reasoning. It’s a dynamic process, grounded in the AI’s ability to navigate a vast mathematical space of relationships between words, ideas, and contexts.
Not a Stochastic Parrot
The “stochastic parrot” critique oversimplifies how AI works. While it’s true that these systems generate responses based on probabilities, the process is far more dynamic than simply parroting back data.
AI systems like ChatGPT operate in what’s called a high-dimensional vector space. In this space, words and ideas are represented as points, with their positions reflecting relationships and patterns learned from training data. For example:
– Words like “king” and “queen” might occupy nearby positions, with the difference between them encoding something like gender.
– Similarly, context shapes these relationships: the word “bank” might align with “money” in one context and with “river” in another.
When I provide input, the AI doesn’t simply repeat what it’s seen before. Instead, it dynamically interprets my words, adjusts its focus using attention mechanisms, and generates a response by pointing to the most relevant areas of this space.
This is a far cry from mimicry. It’s more like navigating a map, where each response represents a path chosen based on probabilities and context.
Not a Philosopher, Either
At the same time, calling AI a reasoner overstates its limitations. Philosophers construct arguments through logical chains, building conclusions step by step. Concepts shape the chain of reasoning, so in the classic example, Socrates is mortal because the concept of “man” implies mortality. AI doesn’t reason this way. Instead, it predicts the next word or phrase based on patterns in its training data.
For example, if I ask, “What is the meaning of life?” the AI doesn’t weigh premises and conclusions. It generates a response by drawing on patterns from texts where people have written about life’s meaning. The result may sound thoughtful, but it’s not the product of deliberation—it’s the product of probabilistic alignment.
How the Pointing Game Works
Here is a way we might think about the interaction with AI as a pointing game:
1. I Provide the First Pointer: My input directs attention to a part of the AI’s conceptual space, setting the context for our exchange.
2. The AI Navigates the Space: Using attention mechanisms, the AI analyzes my input and identifies the most relevant areas to focus on. It assigns probabilities to potential responses based on its training.
3. The AI Points Back: The response reflects the AI’s “best guess” at what’s relevant, drawing on patterns it has learned.
This back-and-forth continues, with each response refining the focus of our conversation. As the “conversation” unfolds, we are essentially pointing to different locations on the map, using words as the pointers. I may have reasons for pointing in the space, but the AI doesn’t really need to know them. It is treating my input as instructions for where it should point next for the “answer.” In a way, the interaction is a dance, where one party takes a step in a direction as the other follows.
The Dance of Attention
What makes this dance possible is the AI’s attention mechanisms. These allow it to weigh different parts of the input dynamically, determining what’s most relevant in context. For instance:
– In a sentence like, “The cat sat on the mat,” the word “cat” might direct attention to “sat” and “mat,” while ignoring less relevant words like “the.”
– These relationships are recalculated for every word, creating a dynamic, context-sensitive representation of meaning.
This mechanism is what allows the AI to handle ambiguity and context shifts, adapting its responses to the flow of the conversation.
Implications of the Pointing Game: Leading the Dance
The pointing game offers a way to think about AI that avoids two common pitfalls:
1. It acknowledges the system’s probabilistic nature without reducing it to a parrot.
2. It respects the collaborative nature of human-AI interaction without overstating the AI’s capacity for reasoning.
In this view, the value of AI lies not in its ability to think or reason like a human, but in its ability to navigate a shared space of meaning. The better we understand this dynamic, the more effectively we can use AI—not as a replacement for human thought, but as a partner in exploring ideas.
This reframing is useful now that AI has become a partner in exploring different types of information. By recognizing that our interactions are essentially a game of pointing, we realize that the goal of the game isn’t to get to the truth. A human has to want the truth in order to navigate there, and education is what increases the odds that we will distinguish truth from falsehood.
We need to keep this in mind in the current context where misinformation proliferates and AIs have the capacity to hallucinate or even fabricate untruths. We cannot take a hands-off approach to our interactions with AIs, because these machines are only looking for the next gesture that satisfies the game. As human partners in the dance, we need to lead instead of follow.