The current era of highly sophisticated chatbots known as Large Language Models, popularly marketed as “Artificial Intelligence,” has presented us with a society-scale version of the mirror test. Per Wikipedia:

The mirror test—sometimes called the mark test, mirror self-recognition (MSR) test, red spot technique, or rouge test—is a behavioral technique developed in 1970 by American psychologist Gordon Gallup Jr. to determine whether an animal possesses the ability of visual self-recognition.^[1] In this test, an animal is anesthetized and then marked (e.g. paint or sticker) on an area of the body the animal normally cannot see (e.g. forehead). When the animal recovers from the anesthetic, it is given access to a mirror. If it subsequently touches or examines the mark on its own body, this behavior is interpreted as evidence that the animal recognizes its reflection as an image of itself, rather than another animal.

While the mirror test is not a flawless test for self-awareness or consciousness (whatever we think that means), I still think it’s fun to watch animals interact with mirrors. Broadly, they do one of three things: (1) they don’t seem to notice the reflection at all, (2) they react to the image of “the other” animal behind the glass and react accordingly, or (3) they seem to recognize in some way that they’re looking at an image of themselves, not a whole other creature.

We humans have been at step 3 for, probably, the entire existence of Homo sapiens, when it comes to mirrors and our own reflections. We understand that there is no “other” living behind the glass of a mirror or below the surface of a reflecting pool. When we aren’t presenting the mirror with a face, the mirror does not show a face. It can produce a reflection which is obviously distorted, or a reflection so perfect it can fool a bystander into thinking it’s the real thing (a common camera trick in movies). We also understand that no matter how absolutely, atomically perfect a mirror might be, the reflection can’t cross the boundary from image to reality.

I’m bringing up all this talk of mirrors in a post about LLMs/”AI” because I am asserting that LLMs do for language and thought what mirrors do for light and faces. And I am also asserting that, on a societal level, we are grappling with step 2 of the “mirror test” that LLMs present to us.

If someone didn’t know what a reflection is, seeing an image of a human face that looks and moves exactly like a human face does could only mean one thing: “there’s a human being there, and I’m looking at their face!” Once they know that a reflection is only the image of a face, and that it’s your face, they can start using mirrors intelligently and with purpose. Mirrors are indeed very useful, from advanced technological applications to picking a bit of green vegetable out of one’s teeth. If you still labor under the mistaken belief that there’s another person living behind the glass, you’ll struggle to get worthwhile use out of any mirror.

LLMs work in language, not reflected photons. They respond to language inputs with statistically-likely, or statistically-preferred, language outputs. To be clear, machines have been outputting language for a long time, arguably well before computers. But a printing press doesn’t dynamically respond to inputs, and a Scrabble board doesn’t communicate with intelligible sentences, so we have little trouble understanding that those objects are only showing us language, not thoughts. In the case of a book (or a blog post), we understand easily that another person transferred their thoughts into the written language. The book itself is not having thoughts, the author had those thoughts. Perhaps we can liken them more to a drawing or photo of a face, rather than a mirror.

LLMs are far more tricky than a printed page. They respond with language that mimics, sometimes with incredible fidelity, the language that people use to communicate thoughts. Most of us have never seen anything that performs such mimicry, let alone so fast or with such fidelity. LLMs do not always give correct answers, but neither do people, so that’s hardly a point against it in terms of the power of the illusion. We give the machine our thoughts in the form of language, and it sends language back at us. We know that our own language is motivated by thoughts, and so our conclusion might well be that surely the language coming from the machine is also motivated by thoughts.

But when we take our language away, the machine produces nothing at all. There is no mind behind the glass of the screen, just as there is no face behind the glass of a mirror. An LLM will not proactively start speaking without some instruction to do so, and the mirror will not produce the image of your face unless your face is there in front of it.

Culturally, or perhaps as a species, we are facing a mirror for our thoughts in a way we’ve never been confronted with before. We are on step 2 of the mirror test, like a bird pecking at a mirror to figure out what that “other” bird is up to. Until we fully crest step 3, and stop bamboozling ourselves with the notion that the machine is an independently thinking being, we won’t be able to use the machine intelligently, and with purpose.