QWE AI Academy Most people scrolling LinkedIn can flag an LLM-authored post in under three seconds. Bryan Cantrill put a name on it in mid-November 2025 (blog version December 5): when you let a model author the post, your intellectual fly is open. The HN thread racked up hundreds of points. Nobody acted shocked.
Public tells, as of late 2025: emoji confetti, one-sentence paragraphs stacked with blank lines, “it’s not just X… but also Y,” em-dashes used like stage lighting. LinkedIn’s own “rewrite it with AI” button spits out that exact sludge – so the platform is training readers to distrust the format.
Readers stop. They don’t know what’s real anymore. Your actual insight gets treated as generated fanfic.
Why the usual fixes leave the fly open
Ban-lists and “add a personal anecdote” tips help a little. They don’t fix authorship.
Cantrill’s line is blunt: models are fine for brainstorming, for chewing dense text, and as editors you can ignore. They are lousy original writers. They are not you.
Ownership is the real casualty. In the MIT Media Lab study Your Brain on ChatGPT (arXiv:2506.08872, 2025), people who let the model author essays showed the weakest brain connectivity of the groups tested. Session 1: about 83% struggled to recall or quote their own text. Teachers called the essays soulless. That is cognitive debt, measured.
Community spotting guides through 2025-2026 keep landing on the same skeleton: negative parallelism, em-dash clusters, corporate vocab, predictable hooks. Em-dash use on LinkedIn rose sharply after mainstream LLM adoption (as of 2025 analyses) – one dash is weak signal; a cluster next to even cadence still flags. Word bans don’t break that skeleton. Human readers catch structure first. Detectors lag.
The only workflow that keeps the fly closed
Editor and sparring partner. Never author. Every shipped sentence is yours.
- Dump raw thinking first. Dictate a messy 5-10 minute voice note or type bullet chaos. No polish. Voice and real insight only show up here.
- Own the outline yourself. Sketch 4-6 points on paper or in a plain doc. You pick the hook, the one concrete story or number, and the ending. The model never chooses structure.
- Write the first full draft by hand (or from the transcript). Ugly is fine. Friction is the point – same “writing is thinking” line the HN thread kept repeating.
- One controlled LLM pass only. Paste the draft. Strict instructions: “Act as a cynical editor. List every cliché, every ‘not just X but Y’, every dramatic em-dash, every single-sentence paragraph cluster, every corporate word. Do not rewrite the piece. Suggest tighter wording on specific sentences only. Keep my exact voice and opinions.” You accept or reject line by line. Copy nothing wholesale.
Optional: drop in 2-3 of your old real posts so it knows your rhythm. Still never generate from a blank prompt.
Pro tip: After the edit pass, read the whole thing out loud once. Any place you stumble or sound like a press release – you cut or rewrite. That ~90-second check beats most detectors on residual machine cadence.
Real-world before/after on a LinkedIn draft
Raw human dump (me, after a failed deploy): “Spent 4 hours chasing a race condition that only hit under load. Turns out the lock was fine – the metrics dashboard was lying about queue depth. Classic. Next time I’m wiring a simple counter before I trust any fancy panel.”
Typical LinkedIn “rewrite with AI” version (fly open): “Currently, fast-paced engineering landscape, it’s not just about writing code – it’s about building resilient systems. I recently encountered a challenging race condition… Here’s what I learned: 1. Always validate your observability… 🚀 What strategies do you use?”
After the workflow (human draft + one editor pass I controlled): “Four hours on a race that only showed under load. Lock was clean. The dashboard was just wrong about queue depth. I’m putting a dumb counter in before I trust any pretty panel again.”
Same facts. Zero tells. Sounds like someone who was actually there.
Pro tips that actually change the output
- Ban the structure in the editor prompt, not only the vocabulary – forbid rule-of-three lists and one-sentence-per-line formatting too.
- Keep one slightly awkward sentence or a number only you would know. Models sand those off; leaving one is a human watermark.
- If you touch LinkedIn’s AI button for grammar, paste into plain text and gut every short paragraph and emoji before it goes back.
- Longer pieces: write the core argument with the chat closed. Open the model only after you already know what you think – the MIT pattern, inverted on purpose.
Have you noticed how some of the best engineers on HN refuse to let models touch their prose even while they let them touch code? There’s a reason.
FAQ
Is it ever okay to let an LLM write the first draft of a post?
No. Not if you care whether people trust you. Own the first draft; use the model only after words exist.
What if English isn’t my first language – doesn’t the model help more?
Say you draft in Polish, then need English clarity. Write the substance first in the language you think in. Hand the model a clarity-only pass with the same “do not rewrite voice or structure” rules. Grammar help is real. Authorship outsourcing still buys the same ownership hole the MIT participants hit.
Won’t better models make the tells disappear soon?
They get smoother. The statistical average still leans on high-probability constructions – negative parallelism, even cadence, tidy CTA shapes. Readers improved at spotting the pattern faster than models hid it. Detectors are secondary. The quiet failure mode is your audience stopping. Cantrill’s point still holds as of late 2025: the fly is open, and people notice even when they don’t comment.
Blank note. Dump the post you were about to outsource. Run the four steps. Ship the human version. See who actually engages.