On Sept. 16, Red Blob Games published a short post titled “English: a vs an.” It concerned a function called a_or_an("apple") for procedurally generated text — the sort of thing a game engine needs when it assembles sentences at runtime. By the weekend the post had 227 points and 306 comments on Hacker News, and most of that thread treated it as a diverting puzzle about vowel letters.

It is not a puzzle. It is a diagnostic, and the patient is not doing well.

The Rule Was Never About Letters

The post makes the ordinary point cleanly. Unicorn begins with a vowel letter and a consonant sound. Hour begins with a consonant letter and a vowel sound. The rule runs on the mouth, not the page, which is why a naive vowel check fails on both words and why the author of the post ended up leaning on CMUdict — the Carnegie Mellon Pronouncing Dictionary — as his data source.

That single choice is the whole story. Every a/an decision is an orthographic audit of whether the writer knows how a word is said. The post even supplies the etymology: English speakers heard “a napron” and wrote “an apron”; heard “an ewt” and wrote “a newt”; Spanish kept naranja because English heard the leading n as an article. Three hundred and six comments, and the interesting claim sits in a footnote about a dictionary built by hand at a university in Pittsburgh.

CMUdict Says Historic Starts With an H

A great many English speakers say it doesn’t. British English drops the h for a large share of speakers, which is where “an historic occasion” comes from; American English mostly pronounces it, which is where the eye-rolling comes from. Both are correct. Both are grating to the other side. Same fight over “an herb” and “a herb,” same fight over abbreviations — “an SUV,” “a NATO,” “an NFT” or “a NFT,” depending on whether you say the letters or the word.

The rule has no answer because the speakers have no agreement. Which means a_or_an() is not a function. It is a policy decision rendered in lowercase.

CMUdict is a genuine public good, hand-transcribed from the 1990s onward, and it is not dialect-neutral — it encodes General American. There is no comparable machine-readable pronunciation resource with equivalent coverage for Indian English, Nigerian English, or the Englishes of Southeast Asia. So the function does not ask. It defaults. A junior localization contractor put it to me in a Slack message that arrived at 2 a.m.: “It’s not a style-guide problem. It’s a six-way argument about the letter h, and the client wants one string.”

A Third of the New Web Was Written by Something That Has Never Heard a Word

Pew Research Center’s 2026 snapshot found that roughly one-third of recently published webpages show signs of AI editing or authorship. Graphite, an SEO firm, examined more than 60,000 new articles published from 2020 through 2025 and found that by late 2024, more than half of new English-language articles were primarily machine-written. Detection is unreliable, the estimates are soft, and the direction is not in dispute.

Text models are trained on spelling. They learn a/an patterns statistically and get the common cases right, which is precisely why the failure is invisible: it presents as a typo, never as a policy. And the statistically dominant answer is American, because the corpus is. So the one place written English forces a writer to take a position on pronunciation is being resolved, quietly and at scale, in favor of a single accent, in text that nobody proofreads and no editor signs off on.

The a/an rule was never going to survive that. It survived a thousand years of scribes and typewriters and spellcheck because each of those technologies punted the decision back to a human who had heard the word out loud. Software does not punt.

The Free Lunch Was Never Free

The implicit answer in that comment thread — use CMUdict — is correct, and it is also the reason nothing will change. A locale-parameterized lookup table is more expensive to build and maintain than a fifteen-line regex, and no grammar eval suite scores a/an as anything but noise. Nobody is going to fund dialect fidelity for a function that costs money and produces no visible improvement on a benchmark.

So three groups should update their priors. The evaluation teams who score generated English on grammar and syntax while treating article selection as a rounding error. The localization leads shipping one string table to five markets and calling it global. And anyone who assumed machine-written prose is dialect-neutral by default. It isn’t neutral. It is Pittsburgh.

English spent a millennium failing to standardize this rule, and the failure was the point — it left room for Dublin and Lagos and Singapore to each have their own answer. Software is about to standardize it in a decade, by accident, for free, and it will not occur to anyone to hold a vote.

Sources

The share of AI-generated content on the web has increased substantially since the release of ChatGPT in 2023.

In a 2026 snapshot of recently published webpages, one-third show signs of AI editing or authorship. | Pew Research Center](https://www.facebook.com/pewresearch/posts/how-is-ai-shaping-the-internetthe-share-of-ai-generated-content-on-the-web-has-i/1448622623800460)