How not to think about ChatGPT

Why we should dismiss the claims of those who dismiss the significance of large language models.

Until very recently, machines didn’t converse with humans, and neither did animals, beyond the odd snort or grunt. Our species, and only our species, used complex language. But we’ve now built a machine that can converse with us. Most of us, in most contexts, would find it difficult to distinguish its output from that of another human.

What on earth does this mean?

It’s hard for me—as it clearly is for everyone—to figure out how to make sense of this. How significant is this, really? What does it represent? How much will it change the world? How promising is it? How dangerous? What can we reasonably predict will happen next? How do we distinguish between reality and hype?

Having now spent a number of sleepless nights thinking about it, I’ve concluded it would be easier for me to say what isn’t true about ChatGPT, or what can’t be true, than to make affirmative statements about its capabilities and significance. And perhaps that’s good enough. Perhaps, through a process of elimination, we’ll narrow things down to a useful assessment. And if not, at least we’ll dispense with a few bad arguments.

Below are a series of claims I’ve recently seen made about ChatGPT. I can say (with some confidence) that they’re wrong.

---

“It’s not a big deal.”

This is a surprisingly common thing to say. The urge to deny the significance of ChatGPT—to claim that we’re not really seeing what we seem to be seeing; it’s all some kind of Clever-Hans sleight-of-hand—is so widespread that I’ve concluded most of us are in denial. Perhaps this is because it’s so astonishing it’s hard to take in. Perhaps this is because it shouldn’t be possible: Machines are not supposed to do this! Or maybe it’s a function of anxiety. This is clearly a revolutionary technology, and it makes the future both very hard to predict and, patently, extremely dangerous. People tend not to like big changes and uncertainty. They certainly don’t like existential risk. So maybe it’s easiest just to say, “Don’t believe your lying eyes. It’s not really doing that.”