The Em Dash Report — who started punctuating like a language model

The Em Dash Report

3 min read Original article ↗

80,913 posts · 59.7M words · 42 blogs

The em dash became the internet's favourite tell for machine-written prose. This counts them: every — in the body text of every post these 42 blogs publish, grouped by the month it went out. It makes no claim about who used a model. It is the punctuation, and the dates.

Company blogs moved together: cloud and infrastructure at 2.0×, AI labs at 1.8×, developer tools at 1.3× their 2018–2022 rate per 1,000 words. The newsrooms went the other way. Newsrooms still use the most at 3.42 per 1,000 words, as they always have.

Em dashes per post, 12-month trailing

6 of 38 shown

Click a name to show or hide it. Both charts share the selection and the address bar follows, so the URL always shares exactly what you are looking at. The dashed marker is Nov 2022, when ChatGPT launched — a date reference, not a claimed cause.

Em dashes per post counts what a reader meets in one sitting, but it moves when posts get longer as well as when punctuation changes, and across this chart the average post grew 2.0×, from about 540 words in 2014 to 1,059 today. Over the same span the rate per 1,000 words went down 4%. Most of the climb above is longer posts, not denser punctuation. Switch to per 1,000 words to take length out of it.

Other things that changed

A rise in one punctuation mark is thin evidence on its own. These are three other ways the prose moved over the same period. Sentence variation is the closest thing here to whether something reads like a person: human writing is bursty, a short sentence next to a long one, and a falling line means sentences are converging on one length.Marker words counts a narrow set that was rare in this corpus before 2023 — delve, tapestry, myriad, intricate and a few dozen others. None of these detects anything. Together they show whether a blog's writing changed shape, not who wrote it.

Words per post, 12-month trailing

The ranking

Ordered by em dashes per 1,000 words since 2024. Rates divide summed dashes by summed words over the window, so a prolific month counts for more than a quiet one. Blogs with fewer than 40 usable posts are left out; Method lists them and why.

BlogGroupPer 1k words, 2024–now2018–2022ChangePosts with onePosts
StripeDeveloper tools4.813.871.2×92%215 / 210kw
AnthropicAI labs4.590.5684%421 / 590kw
Microsoft AzureCloud and infrastructure4.181.323.2×86%3,425 / 3.1Mw
Y CombinatorOther4.180.3512×53%2,368 / 2.4Mw
WIREDMedia and commentary4.084.540.9×85%1,440 / 2.1Mw
TechCrunchMedia and commentary3.804.050.9×64%16,303 / 9.4Mw
The VergeMedia and commentary3.403.281.0×40%10,875 / 5.4Mw
Together AIAI labs3.211.8869%229 / 246kw
Grafana LabsDeveloper tools3.121.402.2×81%1,995 / 1.9Mw
Ars TechnicaMedia and commentary3.084.280.7×61%18,896 / 11.2Mw
ReplicateAI labs2.992.5745%124 / 95kw
Mistral AIAI labs2.9842%78 / 60kw
TailscaleDeveloper tools2.952.121.4×69%390 / 353kw
Netflix TechnologyCloud and infrastructure2.561.571.6×76%224 / 442kw
NetlifyDeveloper tools2.521.391.8×50%410 / 339kw
GitHubDeveloper tools2.341.691.4×56%4,733 / 3.0Mw
Google CloudCloud and infrastructure2.3071%324 / 414kw
Google DeepMindAI labs2.280.653.5×67%341 / 333kw
BunDeveloper tools2.182.0152%176 / 197kw
AWSCloud and infrastructure2.181.012.2×84%4,534 / 3.0Mw
Shopify EngineeringCloud and infrastructure2.001.161.7×64%423 / 771kw
DigitalOceanCloud and infrastructure1.890.822.3×54%807 / 648kw
CloudflareAI labs1.881.491.3×69%3,537 / 4.8Mw
SentryDeveloper tools1.702.370.7×50%909 / 926kw
DockerDeveloper tools1.561.880.8×39%930 / 808kw
VercelDeveloper tools1.551.810.9×35%460 / 405kw
DenoDeveloper tools1.540.265.9×53%248 / 279kw
NeonDeveloper tools1.500.3342%465 / 506kw
Meta EngineeringCloud and infrastructure1.451.650.9×36%1,085 / 1.3Mw
DiscordCloud and infrastructure1.2751%106 / 152kw
RailwayDeveloper tools1.270.1442%210 / 302kw
PlanetScaleDeveloper tools1.141.610.7×35%285 / 342kw
Fly.ioDeveloper tools0.941.010.9×44%410 / 507kw
Hugging FaceAI labs0.830.175.0×26%727 / 1.1Mw
SupabaseDeveloper tools0.720.501.5×18%412 / 350kw
NodecraftOther0.590.1110%211 / 104kw
JetBrainsDeveloper tools0.540.173.3×9%686 / 350kw
TechRadarMedia and commentary0.520.153.5×15%1,440 / 965kw