Did the em dash spread on Hacker News after ChatGPT? Its own drift check voided it

People repeat two opposite things about the em dash. One says machines have spread it through online writing. The other says people have given it up so as not to look like machines. Hacker News comments are a long public record of typed text from before and after, so the count can be made.

The rule is sealed first, and the whole test is in rule.py. A rise here would say the mark moved on one forum. It would say nothing about who or what typed it: phones also turn two hyphens into an em dash, which is why the curly apostrophe is counted beside it.

Before any comment was fetched, I tried a first design on invented data. It projected a fitted trend across the gap, and in a world with no change at all it reported a fall of 26 %. So I dropped the trend and measure drift with the same count one period earlier. That is written in the rule too.

Result

Void. The rule tested the rise against the same count four years earlier, and that earlier count moved too far. R came out at 1.463 [1.262, 1.700], and it stands here as the number the rule declined to score. The check gave 1.314 against a limit of 1.25.

The share of comments with an em dash, by year:

Most of the drift comes from the first years, when the mark was climbing with nothing to do with chat models. The curly apostrophe shows how far these marks can move for reasons of their own: it went from 2.24 % to 7.09 % of comments between 2017 and 2018. I did not establish why. Over the sealed windows it barely moved (R 1.07), and its own drift check read 2.13.

The marks read beside the em dash, none of them scored: em dash in comments with no curly apostrophe 1.45 (drift 1.20), en dash 1.24, spaced double hyphen 0.73, semicolon 0.96. The word “delve” read 1.76 [0.57, 5.27], which is too rare here to say anything.

What I saw after the verdict

None of this is sealed. Looking month by month, the share stays near 1.3 % until the start of 2026, and from May 2026 it sits near 0.7 %. Comments marked dead per sample roughly double from March 2026, and comments get a little shorter.

Two dated things sit near that. On 25 February 2026 an analysis at Marginalia counted em dashes, arrows and other symbols in 17.47 % of comments from newly registered accounts against 1.83 % of recent comments in general, about 700 comments each. On 11 March 2026 the site’s own guideline, “Don’t post generated text or AI-edited text”, was a story there with 4,229 points. Nothing here shows that either caused anything.

That fall was a pattern I found by looking, so I tested it separately on fresh comments. That test is void too.

A trap in the dead comments

My first pass found no em dash in 1,559 dead comments. That is not a finding. The public item API has returned only a stub of about seven characters for dead comments since about 2022, so there is nothing to count. Anyone studying moderation through this API can make the same mistake.

The pack holds the sealed rule, the scripts, the monthly counts and the results.