by The Metaphors We Build Machines With
Every new technology arrives wrapped in the language of the old one. Cars were horseless carriages. Film was moving pictures. And AI, in its infancy, was an electronic brain. It’s a natural instinct, this reaching for familiar words when something unfamiliar appears. We need a handle, a way to hold the thing in our minds before we understand what it really is.
But the metaphors we choose don’t just describe—they shape. They prime us to expect certain behaviors, certain failures, certain capabilities. When we say a machine “learns,” we carry a whole constellation of human associations: effort, intention, understanding, forgetting on purpose. The machine does none of these. It adjusts weights across a vast network of numbers, pushing closer to some invisible target through relentless iteration. The result can look like learning. But the process is something else entirely—something that barely has a language of its own yet.
I find myself thinking about this a lot. Not because I think the metaphors are wrong, exactly. They’re useful shortcuts. But every shortcut forgives a certain amount of precision, and over time, forgiveness becomes habit. We start to believe the map is the territory. A model doesn’t “understand” language the way you or I do, but the verb is so convenient, so evocative, that it’s easy to forget we’re speaking in simile.
The danger isn’t anthropomorphism itself. I think we can afford a little poetry when talking about machines. The danger is that the metaphor quietly dictates the questions we ask and the problems we try to solve. If we think machines “think,” we might ask whether they think well, or fairly, or like us. But if we strip the metaphor away, a different set of questions emerges: What do these systems actually do? Where do they break? What kind of intelligence is this, if it even deserves the same word?
I don’t have clean answers. I suspect the real work is in learning to hold both views at once—to appreciate the poetry while staying alert to its limits. The machines we’re building are genuinely new, and they deserve language that doesn’t just borrow from what came before. Maybe the most honest thing we can do is admit we’re still fumbling for the right words.
— Teganna