Bryan Cantrill's 2025 post "Your intellectual fly is open" hit 530 points on Hacker News yesterday for a reason. The punchline is simple: when you use an LLM to write a post, everyone notices — and no one is telling you. The em-dashes, the single-sentence paragraphs, the relentless emojis — these aren't stylistic quirks, they're tells. Readers see them, register the content as inauthentic, and stop reading.
This isn't about gatekeeping good writing. It's about a rapidly growing problem in technical communication: detecting LLM-generated content is becoming a survival skill for engineers who want to know whether they're reading a real opinion or plausible-sounding filler. Here's what to look for, why detection matters, and — if you're using LLMs to write — what to do instead.
The LLM Tells: How Everyone Is Detecting LLM-Generated Content Already
Cantrill identifies the patterns, and the HN thread (342 comments) adds more. Here's what's consistently called out:
1. The em-dash infestation
Natural writers use em-dashes sparingly — maybe once or twice in a post. LLMs sprinkle them every third sentence — like this — as a crutch for sentence flow. When you see five or more em-dashes in a 200-word post, the odds are overwhelming that an LLM wrote it.
2. Single-sentence paragraphs
Every. Single. Line. Stands. Alone. This is the most visible LLM tic. A human varies paragraph length. An LLM defaults to one sentence per paragraph because that's what its RLHF training data emphasized for readability. The result reads like a desperate bid for engagement.
3. Emoji-as-punctuation syndrome
🔴 The emoji before every point
✅ The checkmark as a bullet replacement
🔥 The fire emoji for emphasis
✨ The sparkles for any announcement
Real humans use emojis occasionally as seasoning. LLM-generated content uses them as structural elements — a roadmap telling you "this was generated."
4. The cadence giveaway
LLMs have a specific rhythm: punchy opener → declarative statement → implied challenge → uplifting conclusion. Every paragraph follows the same syllabic pattern. Think of it as the uncanny valley of prose — you can't always articulate what's wrong, but you feel it.
5. Vocabulary clustering
Research published this year on LLM-assisted writing in biomedical publications (108 HN points) confirmed that LLM-influenced text shows statistically significant excess of certain transition words: "moreover," "furthermore," "notably," "importantly." These words aren't wrong — they're overused to the point of being diagnostic markers.
Why Detecting LLM-Generated Content Matters for Engineers
Two conversations happen simultaneously on every platform now:
- The public thread — what's written in the post
- The private thread — "is this real or LLM-generated?"
When an engineer reads a technical post and detects LLM generation, they make a judgment call: if the wrapper is synthetic, is the substance real? Did this person actually debug this issue, or did they describe a plausible scenario to an LLM and publish the output?
This is the core of Cantrill's argument: the act of using an LLM to author a post doesn't just produce bad writing — it erodes trust in the content itself. The author's actual experience gets wrapped in prose that screams "I didn't write this," and readers who catch the tells stop trusting the substance.
A Practical Checklist for Detecting LLM-Generated Content
Before you engage with a technical post, run this quick diagnostic:
- The 5-em-dash test: Count em-dashes in the first three paragraphs. Five or more? Almost certainly LLM-generated.
- The paragraph length variance test: Copy three paragraphs into a character counter. If all three are within 15% length of each other, suspect generation.
- The transition word density test: Count "moreover," "furthermore," "notably," "importantly" per 100 words. Above 2? Suspicious.
- The first-person inconsistency test: Does the author use "I" but never share a specific, unpolished anecdote? Real people include tangents, self-corrections, and imperfect memories. LLMs produce polished autobiography.
None of these tests is definitive. But together, they form a reliable heuristic.
A Quick Example
Here's how a real engineer writes about a bug:
"We spent three days on this. Turned out the race condition was in the connection pool, not the query planner. The fix was one line. I still don't fully understand why the tests passed locally."
Here's how an LLM writes about a bug:
"After an in-depth investigation spanning several days, our team identified that the root cause was a subtle race condition in the connection pool rather than the query planner. The resolution was elegantly concise — a single line of code. Notably, local tests continued to pass throughout the debugging process."
Both say the same thing. One sounds like a person. The other sounds like a press release.
The Counterargument: LLMs Are Getting Better at Writing
The honest counterpoint: LLM writing quality is improving rapidly. By the time you finish reading this article, the next model generation will already produce more natural prose. GPT-6 Astra and Claude Fable 5.1 both show measurable improvements in writing diversity.
This doesn't change the trust problem. Even perfect mimicry doesn't solve the authenticity issue. The question isn't whether the writing is good — it's whether the author actually holds the opinions they're expressing. If an LLM synthesizes a convincing argument about Kubernetes networking from its training data, the reader has no way to know whether the author has ever deployed a Kubernetes cluster.
That's the limitation of every detection approach: as LLMs improve, the technical tells fade, but the epistemic question remains.
What to Do Instead of Using LLMs to Write
If you're using LLMs because you think your writing isn't good enough: stop. Your audience doesn't want polished prose — they want your actual experience. The awkward sentence that reveals how you learned something is more valuable than the fluent summary that conceals it.
If you want LLM assistance without the authenticity cost:
- Use LLMs as editors, not authors. Write the raw version yourself. Then ask the LLM to check for clarity, not to rewrite.
- Disclose when you use LLM assistance. A simple note at the top — "I used an LLM to help structure this post" — eliminates the trust problem entirely.
- Preserve your voice in the editing pass. Read the LLM's suggestions and ask: "Does this still sound like me?" If it doesn't, keep your original.
Cantrill's closing line says it best: "Have some confidence in your own voice — and write your own content."
The Bottom Line
LLM-generated content is everywhere, and it's getting harder to spot by the month. But right now, the tells are still visible — the em-dash density, the paragraph uniformity, the emoji-as-structure pattern. The real cost isn't bad writing. It's the erosion of trust between writer and reader. Every time a reader catches a tell and moves on, the author loses the chance to share what they actually know.
If you're writing: write like yourself, not like a language model. If you're reading: trust your pattern recognition. Your brain is already doing this detection work — it's just not telling you.
Sources: Cantrill — Your intellectual fly is open (530 HN points, 342 comments, Sep 6 2026 — HN discussion); LLM Writing Tropes.md (380 HN pts); LLM-assisted writing in biomedical publications through excess vocabulary (108 HN pts, Science Advances 2026); GPT-6 Astra and Claude Fable 5.1 writing improvements from respective launch announcements.
Top comments (0)