Unavoidable ambiguity

Why can no language avoid words that carry more than one meaning?

Ambiguity looks like a defect. Bank is a riverside and a place that holds money; set has dozens of senses; light is both a weight and a brightness. A carefully designed code would never do this — a barcode has one reading, a chemical formula has one reading. So why has no natural language, in thousands of years and thousands of independent traditions, managed to clean this up?

Let us not assume it is laziness. Let us ask instead what would have to be true for a language to succeed in avoiding it, and see whether those conditions can all hold at once.

A language that avoids ambiguity needs a distinct form for every meaning. So count the forms. Every language draws its words from a small inventory of contrasting sounds — commonly somewhere between about eleven and a hundred-odd phonemes — and constrains how they may combine, since not every string is pronounceable or permitted. That gives a supply of possible word-shapes.

Is the supply finite? Not in the strict sense: allow longer words and it grows without limit. So the constraint is not that forms run out. Notice where that leaves us — if there were no other pressure, a language could simply mint a new long word for every new meaning and be done.

So there must be a second constraint pulling the other way. What is it? Words have to be produced and understood in real time by a memory that is not unlimited. Long words cost breath, time, and recall. And the pressure is not spread evenly: George Zipf observed that the words we use most are systematically the shortest, which is exactly what you would expect if a language is trading form length against how often that form must be uttered. Short forms are the cheap slots, and there are only so many of them.

Now the two constraints meet. Meanings are open-ended — every new tool, plant, role, joke and abstraction wants one — while cheap forms are scarce and stay scarce however large the vocabulary grows. What happens when an unbounded demand meets a small stock of cheap resources? The cheap ones get reused. That is the pressure that produces set and run and bank, and the reason the words with the most senses are almost never the long technical ones.

But scarcity alone would only make reuse tempting, not tolerable. So ask the next question: what makes reuse cheap enough to accept? Reuse is only viable if the hearer can recover which sense was meant — and here a third fact enters. Words never arrive alone. I sat on the bank and I wired money to the bank do not need the word to disambiguate them, because the sentence, the topic, the speaker, and the room already have. Context is a resource the language does not have to pay for, because it is there anyway.

That reframes the whole question. Ambiguity is not a leak the system failed to plug; it is the system declining to pay for information that will be supplied for free. Piantadosi, Tily and Gibson argued exactly this in 2012: for any communication system where context is informative, reusing short, easy forms is more efficient than assigning each meaning its own unambiguous form, because the disambiguating work is already being done elsewhere.

Two honest qualifications. First, this efficiency argument is a well-supported account, not a settled proof — it is a claim about the shape of the pressure, and how much of a given language's polysemy it explains is still argued. Second, not every ambiguity is an efficiency: some are accidents of history. English bank is two unrelated borrowings that happened to collide in sound — true homonymy — whereas the many senses of run fan out from one word by extension, which is polysemy. The constraint argument explains why such collisions are tolerated and why they cluster in short, frequent words; it does not claim each one was chosen.

The misconception worth correcting is that ambiguity is a cost paid by the listener. Most of the time it is not paid at all: the resolution happens automatically, and speakers only notice ambiguity when context fails — in headlines, legal drafting, error messages, and jokes, which are all contexts deliberately stripped of the surrounding cues.

It also helps to see what the argument does not say. It does not say ambiguity is always harmless. Where context genuinely cannot help — an aircraft callsign, a drug name, a variable in a program — humans build artificial vocabularies that behave like the barcode we started with, and they pay the price in length and in the training required to use them. That is the exception proving the rule: unambiguous naming is achievable, and we reach for it exactly where the free resource of context is absent.

Finally, notice that the argument is about pressures, not about any individual word. No speaker decided that set should carry so much. The constraints act over generations, on a vocabulary shaped by reuse, borrowing, and drift, and their signature is statistical: short and frequent words carry many senses; long and rare words carry one.

• Zipf's law of abbreviation, and the evidence that frequency drives word length across languages.
• How homonymy and polysemy are told apart in practice, and why dictionaries disagree about sense counts.
• Garden-path sentences: what happens when the resolving context arrives too late.
• Designed vocabularies — aviation callsigns, drug naming, chemical nomenclature — and what they cost.

Every term the collection defines is gathered in the glossary.