The greeting took me by surprise: “Hey, you.”
I was expecting a “hello” or “hi,” statements of straightforward recognition. “Hey, you” was more knowing, more familiar. Say it in a crowd, and you place yourself and the other person in a cozy kind of bubble. It’s personal. But it came from a chatbot.
“Another day, another chat, right?” it continued, offering a slight twist on an old expression. Complimented on its speech, it responded modestly, “What can I say?” You could almost imagine it holding out its hands with palms up and giving a shy smile.
The conversational tone is almost uncanny. But it’s not a sign that large language models are becoming human; it’s a sign that they have mastered pragmatics.
That’s the word linguists use for verbal gestures of this kind, the ones we use to convey attitude, to indicate shared experience, to distinguish old news from novelty. They’re pragmatic in the sense that they use language in a practical way, beyond literal definitions. When a computational linguist told me several years ago about the advances artificial intelligence was making in simulating human language, the first thing I thought was: The hot next step will be getting to pragmatics.
We are taught to think of language as words strung together with grammar. Words evoke specific referents: “apple” refers to the fruit, “yesterday” means the day before today. Then grammar provides the traffic rules: “walk” in the present, “walked” in the past, “-ly” for adverbs.
But it’s the pragmatics that make a sentence sound like thought. Like feeling. For example, words and grammar get you the neutral, factual “She had a horse.” But what about a slightly inflected, more conversational version of that sentence, such as “She even had a horse.” “Even” conveys that her possession of a horse was the most striking example of a scale of things distinguishing her, as in “She had a mansion and a Lamborghini, and on top of that — get this — she had a horse!” Used this way, it conveys that something is counterintuitive, novel.
The pragmatic aspect of language resists meat-and-potatoes description. Just imagine trying to explain the use of “even” in that sense to someone just beginning to learn English. And yet pragmatics have a huge effect on our understanding of context.
Without them, you barely get past the “My uncle is a lawyer, but my aunt has a spoon” level of language. I took in-person Mandarin lessons for a spell, and my teacher once remarked “You like ‘shenzhi,’ don’t you?” That’s the word for “even,” and I told her that without it, I felt I couldn’t say anything interesting.
In the marvelous Onion headline “Five or Six Dudes Jump Out of Nowhere and Just Start Whaling on This One Guy,” “five” and “jump” are words, and the “-ing” on “whaling” is grammar. But what makes the headline funny is that it is couched with pragmatic features that the formality of print has traditionally avoided. The “just” conveys surprise, and “this one guy” invites shared attention, as if the person were telling you the story in conversation and situating the characters almost cinematically.
Emojis convey much of what pragmatics do, lighting up the bare-bones terseness of texting with the spark of lively interaction. That’s why people love using them to color their communications. (It’s also why emojis will never replace language. The emoji version of “Moby Dick” is a good stunt. It’s not going to knock the original off the shelf.)
You might say that words are Language 1.0, grammar is Language 2.0 and pragmatics are Language 3.0, where we take the basic material and set it to use for the interpersonal purposes of real-life communication. But pragmatics are not just frills or seasoning. They are as crucial to the tool kit of any language as harmony is in much music, even if, like harmony, it is not immediately easy to perceive or describe.
This, then, is why it was inevitable that chatbots would find their way here. Nearly three years ago, when ChatGPT made its debut, the biggest shock was that machines had seemingly learned to speak. The more advanced that speech gets, the more it leans into subtle contextual constructions of meaning and mood, the more we can all learn about how language works. We are so often taught about language in terms of prohibitions — use “fewer apples,” not “less apples”; avoid vague pronoun references — rather than the revelations that pragmatics offer.
John McWhorter (@JohnHMcWhorter) is an associate professor of linguistics at Columbia University. He is the author of “Nine Nasty Words: English in the Gutter: Then, Now and Forever” and, most recently, “Woke Racism: How a New Religion Has Betrayed Black America.” @JohnHMcWhorter
The post The Single Word That Explains Why Chatbots Sound So Human appeared first on New York Times.
