This is a fantastic point by Dagmar Monett.

AI is a subfield of study in computer science. If one were to pursue a PhD in computer science, one would have to take required classes in the algorithms subfield, the systems subfield, and the AI subfield.

AI is hence a subfield of study, which is extremely broad, and has, for many, many years, not really had much do to with the statistical methods it relies on now (there was a transition from rule-based to learning systems). The field of AI may aspire toward building AI. However, it’s never been clear how to even categorize something as AI. It’s a field, not a deliverable.

To label a product that an AI researcher builds as AI is therefore sort of misleading. A main (arbitrary) test for whether something is AI is called the Turing test, which suggests that something is AI if a human can’t tell it from a human.  However, this is surprisingly nuanced, especially when we begin to ask ourselves whether systems that depend on statistical models / machine learning, which is the core of today’s “AI,” are AI. These types of models are often very specific—sometimes even specific in being general.

In other words, it might seem like a language model is having a conversation and therefore has more general intelligence, but this is because it’s been trained to have a conversation. In a way, it’s a like a calculator that doesn’t add numbers, it has a conversation. However, you can’t use a calculator to have a conversation, and you can’t use a vanilla language model as a calculator.