Every Time I Fire a Linguist

There’s a line in AI folklore that every linguist knows and none of us enjoys. Fred Jelinek, who led speech recognition research at IBM in the 1980s, is said to have remarked that every time he fired a linguist, the performance of his system went up.

He may not have put it quite like that, and he later suggested it had grown in the retelling. But the point stuck because it was largely true.

Early attempts to get computers to handle language tried to teach them the rules: grammar, vocabulary, sentence structure, the sort of thing linguists spend careers mapping out. It was painstaking work, and it kept running into the same problem. Real language is messy. People break rules, invent words, leave sentences unfinished and rely on context nobody wrote down. Every rule needed an exception, and every exception needed another one.

Jelinek’s team took a different approach. Instead of telling the computer how language should work, they gave it huge amounts of real language and let statistics find the patterns. It worked better, and that idea, scaled up over the following decades with vastly more data and computing power, led more or less directly to the large language models we have today. Modern AI didn’t learn English from a grammar book. It learned it from reading a large part of the internet, which explains rather a lot.

So it’s fair to ask what a linguist is doing writing about AI. Wasn’t our side of the argument lost in about 1988?

I’d say the opposite has happened. The engineers solved the problem of making machines produce fluent language. That’s exactly what has made understanding language more important, not less, because now machines produce it at a scale and polish that nobody can keep up with.

The questions that matter today are mostly linguistic ones. What does a word like “understands” or “intelligent” actually mean when we use it about software? Why does a confident sentence feel trustworthy even when it’s wrong? Why does a vague instruction get a vague answer? What does a policy mean when it says “the system decided”? How do you explain a complex change so that people believe it and act on it?

None of those questions is about how the technology works inside. They’re about what happens between people and language, which is the territory linguists have always worked in, and the territory most organisations are now having to navigate in a hurry.

There’s still plenty of specialist linguistic work inside AI itself, from testing how well models handle less common languages, to studying bias in training data, to evaluating whether a response actually answers the question asked. But for most people, the AI question isn’t how to build it. It’s how to talk to it, talk about it, and tell when its words are doing more work than its thinking.

That’s my patch. Jelinek may have been right that linguists didn’t make the machines better at talking. Somebody now needs to help the people using them listen more carefully, and that, it turns out, is a job for someone who has spent a career paying close attention to words.