Introduction

There is a line of code that looks like this:

if age > 65:
    flag()

It is a single comparison. It has been doing the same thing since the 1950s. And somewhere right now it is being sold as artificial intelligence.

At the other end of the same word sits a frontier model trained on a good fraction of everything humanity has written, costing tens of millions to run, capable of things its own builders can't fully predict. That is also "AI."

When one word stretches comfortably across both of those, it has stopped describing anything. The usual response to noticing this is a sigh — yes, everything's "AI" now, very tiresome. But the tiredness is the least interesting part. The interesting part is that this isn't laziness or hype in the ordinary sense. It's a specific, well-documented thing happening to a word, sped up by a market that pays for it to happen. Once you see the mechanism, you can see exactly where it does damage and where it doesn't — which matters, because the damage lands in the worst possible places.

A word informs by ruling things out

Start with what it means for any word to carry information at all. "Spaniel" tells you something because most animals are not spaniels. The word does its work by excluding — when I say spaniel, I've ruled out cats, herons, most other dogs. The narrower the exclusion, the more the word tells you.

Run that test on "AI." What does hearing it rule out? A hand-written conditional? No — that gets called AI. A spreadsheet formula? Increasingly no. A recommendation rule someone wrote by hand in 2009? It's being rebranded as we speak. By the time you've finished listing what the word doesn't exclude, you've listed almost the whole of computing. A term that rules nothing out tells you nothing about what it's attached to.

So "AI" hasn't become meaningless, exactly. It has changed jobs. It no longer describes a capability; it marks membership in a category — modern, advanced, worth investing in, not to be left behind. That residual function is real, and it's why the word refuses to die. Calling your product AI-powered says "we are the kind of company that does this," which is a useful thing to say even when it describes nothing about the product. But a badge cannot do the work of a description, and the trouble starts the moment someone asks it to.

Why this happened so fast

Words lose their edge all the time. "Smart" used to mean something; now it's on fridges. "Natural" survives on packaging as pure decoration. Linguists have a name for this — semantic bleaching — and they've studied it for decades. It's normally a slow erosion, generations of ordinary use each wearing the word down a fraction.

"AI" did it in about three years. That speed is the part worth explaining, and the explanation isn't linguistic. It's financial.

Most words bleach by accident. "AI" is being bleached on purpose, because the incentive only points one way. Attaching the word to your product adds perceived value — sophistication, investment interest, a premium price — and it costs nothing, because there's no penalty for overclaiming and no standard the claim has to meet. Nobody is fined for calling a regression model AI. So every actor in the market has a reason to reach for the word, and not one has a reason to be more precise. When everyone pushes a word outward and no one pulls it back, it hollows out fast. The decades "smart" took, "AI" gets through in a product cycle.

This also gives you a clean way to tell genuine spread from inflation, which otherwise look identical from outside — both just look like the word being everywhere. Ask four questions. Is the label decoupled from any actual capability, slapped on systems that differ by orders of magnitude with no qualifier? Does using it pay, with no cost for overreach? Is the claim basically uncontestable — is there any standard it has to clear? And after you hear it, do you know less about the thing than before? A word spreading because a real capability is genuinely diffusing scores low on all four. "AI" in public use scores high on all four. That's the signature of inflation, not diffusion.

Where it actually hurts

If this were confined to marketing decks it wouldn't be worth an essay. Bad packaging is survivable. The cost shows up somewhere worse: anywhere people have to reason about these systems and the trivial and the consequential are forced to share a name.

Take "we should regulate AI." At the level of the word this is almost incoherent, because the thing being regulated swings silently between a spreadsheet macro and an autonomous system with genuinely novel risks. You can't write a sensible rule that governs both. Here's the useful move: make the sentence name what it actually means. Most serious regulatory proposals, once you force them to be specific, turn out not to be about "AI" at all — they're about automated decisions that affect someone's legal rights, or about models above some capability threshold. They were never arguing about the umbrella. They just sounded like they were.

Be honest about what that buys you, though. Naming the referent doesn't tell you what the right rule is. Two people who agree they're discussing "frontier models with autonomous tool use" can still disagree completely about how to govern them. What precision kills is the fake argument — the one that existed only because two people were quietly talking about different things under the same word. The real disagreement survives, now visible and as hard as it ever was. That's progress: you've stopped wasting breath on a confusion and can start on the actual problem. It is not a solution, and anyone who tells you precision resolves these debates is overselling it the same way the word oversells itself.

The second cost is subtler and worse: claims migrate. Something true about one kind of system slides, under cover of the shared word, onto another. "The AI just flags transactions, it can't act on its own" — accurate about a narrow classifier, a flat lie once the same word carries it to an autonomous agent. Or in reverse: "AI might pursue goals we didn't intend" — a serious worry about a frontier system, pure scaremongering aimed at a regression model. In both directions the bleached word is the vehicle. Because it spans both systems, a true statement about one becomes a false statement about the other and nobody appears to have changed the subject. That's what makes it so hard to catch in real time. The sentence stays grammatical. Only the thing it points at has moved.

What to actually do

Not reclaim the word. That's a lost cause — bleached words almost never un-bleach, and the market that hollowed this one is still running. "AI" in the loose sense is going to keep meaning "the exciting computer stuff," and at a dinner party or in a headline that's completely fine. Nothing turns on the referent there. Insisting on "transformer-based language model" over drinks isn't rigour, it's just tedious.

The move is narrower: route around the word where the stakes are high. In regulation, in safety arguments, in any spec or claim where the difference between a trivial system and a consequential one is the whole point, don't say AI — say the specific thing, and watch whether your claim still stands once you do. Often it won't, and that's the word telling you it was doing your arguing for you.

This won't hold at the level of the whole language; the incentive guarantees that. But particular communities can keep their own vocabulary sharp for their own purposes, the way working scientists maintain technical terms the wider culture happily erodes. It's a local norm, not a reform.

So the question to carry into any high-stakes use of the word is just this: is it describing the system in front of you — or only marking it as the kind of thing we're currently excited about? If you can't tell, name the thing, and the answer usually arrives on its own.

---

Further reading: on semantic bleaching as a linguistic process, Eve Sweetser's work on grammaticalization and the broader account in Hopper and Traugott's Grammaticalization*. On how a dominant vocabulary shapes which questions even get asked, Kuhn's* The Structure of Scientific Revolutions*. The information-theoretic sense of "informative" — a signal informs by reducing uncertainty — traces to Shannon.*