On September 16, Red Blob Games published a short blog post titled “English: a vs an.” As of this writing it sits at 306 points and 408 comments on Hacker News — a remarkable amount of argument for a rule that every one of those commenters follows flawlessly and almost none of them can state correctly.
The post is a programmer’s move: handed a grammar question, write code. The author parsed CMUdict, Carnegie Mellon’s pronouncing dictionary, built a trie, and computed the answer across the entire lexicon. The answer — the one your second-grade teacher gave you in a single sentence, and got slightly wrong — is that the article follows the sound, not the spelling. An hour. A union. An FBI agent. A one-year term. The AP Stylebook’s entry on “a, an” says the same thing, in nearly the same words.
So the thread is 408 people mocking English for being irrational while demonstrating, in chorus, that its most famous irregularity is a rule.
The Rule Nobody Teaches and Nobody Breaks
That is the part worth sitting with. The a/an split is not a memorized list; it is productive. Nobody was ever taught which article precedes “SaaS” or “NFT,” and everybody gets those right anyway. The rule has even rewritten the language underneath us: “a napron” became “an apron,” “an ewt” became “a newt,” and “an orange” arrived in English carrying an n filched from Spanish naranja. That isn’t a decorative quirk of a charmingly chaotic tongue. That’s an engine that reshapes the lexicon while nobody watches.
Which means the interesting fact here isn’t that English is weird. It’s that a rule this mechanical — one applied pre-consciously, in real time, to words coined last week — has no written authority behind it at all. There is no committee. There is no edition. There is just a few hundred million people running the same program without a manual, and a comment section loudly declaring that no manual could exist.
”An Historic” Is a Register, Not a Mistake
Here is where the written rulebook and the unwritten rule diverge, and it is instructive that they diverge so narrowly. The AP Stylebook prescribes “a historic event.” Merriam-Webster notes that both “a historic” and “an historic” are acceptable, that “a” is now more common, and that “historic” and “historical” spent more than 400 years drifting apart into separate jobs. The h in “historic” is pronounced by nearly everyone. So “an historic” has no phonetic justification whatsoever — which is precisely why it survives. It isn’t a grammar error. It’s a register, doing the same work as “whom” and “shall”: announcing where the writer went to school, or wishes they had.
Style guides cannot adjudicate that, because by the time a usage reaches “an historic,” it has stopped being a question about sound and become a question about class.
A product manager for a text-to-speech model put it well, over a hotel-bar table after a speech-tech conference this spring. “We stopped writing article rules years ago,” she said. “We ship a pronunciation lexicon and a model, and it just works — except for ‘historic,’ which we override by hand, because our customers are American and the older ones hear the rulebook in their heads.”
What the Rulebook Cannot Hold
Now notice what Red Blob’s author had to do to make a machine apply the rule the classical way: hand it the entire pronunciation lexicon as data. Parse the dictionary, build the structure, enumerate the exceptions. The written form of the a/an rule is not a sentence. It is a file.
And the post’s own footnote concedes the argument. “I did not use LLMs to write any of this code,” the author writes, “but in hindsight, I should have. This is one-off code to answer a question.” The man who built the rulebook says he should have let the model do it.
That is the whole dispute between written rules and learned ones, compressed into a blog post about articles. Stylebook committees, compliance manuals, grammar checkers, onboarding documents — every one of them patches exceptions and falls a vernacular behind. Meanwhile the thing that never got written down is the thing a model absorbs from raw text in the first week.
I’ll grant the obvious objection, because it’s the strong one: the a/an rule is also the closest thing English has to a checksum. It is deterministic enough that a violation registers instantly, which makes it a convenient scapegoat for people who want a tidy way to spot machine prose — and a convenient way to be wrong, since the tell is a failure of a rule the model has essentially never broken.
But that leads somewhere less comfortable than the standard AI argument, for both teams. For the crowd that insists these systems are stochastic parrots, here is a rule the average college graduate cannot articulate, executed by something that has never been handed a phoneme or a syllable boundary. And for those of us who like to believe that style is the irreducibly personal part of writing — voice, instinct, the sentence only you would have written — consider that a great deal of what we call voice is thousands of tiny involuntary rules of exactly this kind. None of them written down. All of them load-bearing. Writers who think they’re being original are, at the level of the article, mostly being grammatical.
The 408 commenters had a fine afternoon. They came to laugh at English for being chaos and spent the day proving that its most mocked rule is more reliable than any style guide that has ever tried to describe it. By all means, write that down.
Nobody will learn it from the writing.