For most of August I did something slightly unhealthy. I drafted every LinkedIn post I published, plus about thirty I never published, with an AI assistant open in the next tab. Same notes, three tools. ChatGPT, Claude and Gemini each got identical instructions, and I kept a spreadsheet of what came back.
The first week was demoralising. Every draft opened with a one-line hook, a blank line, then a short story with the lesson tucked into the last paragraph. Two of the three tools reached for “It’s not about X. It’s about Y.” before I had finished reading the prompt. The posts were fine. They were also indistinguishable from the forty posts above and below them in my feed, and one of them earned a message from a colleague who has never once commented on anything I’ve written, asking whether I had “started using the robot.”
By week three I had a method that worked. The drafts still started in a chat window, but the parts that made them mine happened before and after the model touched them. That is what this guide covers, along with what LinkedIn changed this summer that makes the gap between assisted writing and generated writing matter more than it did a year ago.
The short answer Give the AI your raw material, a real event, a number, a sentence someone actually said, instead of a topic. Ask for a draft with a list of things it must not do, in your sentence lengths, never “engaging” copy. Then cut the opening line, the closing lesson and every “It’s not X, it’s Y.” Read the first 140 characters on your phone before you post, because that is roughly where mobile cuts to “see more.” |
Why AI drafts sound like LinkedIn
Ask any assistant for “a LinkedIn post about X” and you get the statistical centre of a genre. Large models learned that genre from millions of public posts, and public LinkedIn posts have converged on one shape: single-sentence hook, white space, short story, tidy lesson, question, hashtags. The model isn’t being lazy. It is giving you the average you asked for.
That average is now largely machine-written, which tightens the loop. Three measurements published this year point the same way, using different samples and different detectors.
How much of long-form LinkedIn is flagged as likely AI
| Study | Sample | Flagged as likely AI |
|---|---|---|
| Originality.ai, July 2026 | 5,000 public long-form posts found through search | 81.2% |
| Pangram, via 404 Media, July 2026 | Long-form posts and short-form posts | 41% and 30% |
| Originality.ai, 2025 | 3,368 long-form posts from 99 influential profiles | 53.7% |
The numbers disagree because the samples do, and detectors return probabilities, not proof. Read the spread as “somewhere between a third and most,” which is enough to explain why readers recognise the shape instantly.
That recognition is the real cost. A post that follows the template is not punished for being AI. It is skipped for being familiar.
What LinkedIn changed in 2026
The platform spent the summer moving from encouraging AI drafting to policing the output. Four changes affect how you should use these tools.
The 2026 changes that matter for AI-assisted posts
| Change | What happened | What it means for your drafts |
|---|---|---|
| “Seems like AI slop” option | Added to the three-dot menu on every post on July 30, 2026, first documented by 404 Media. Tapping it hides the post. | Readers now have a one-tap sanction for template writing. It doesn’t check whether AI was used. It checks whether the post reads like it was. |
| “Enhance your post” retired | LinkedIn pulled its in-composer rewriting tool and is replacing it with a proofreader that corrects text without changing the writer’s voice. | The platform itself stopped rewriting posts. Use AI to fix and tighten, not to generate the voice. |
| Inauthenticity nudge | Chief product officer Hari Srinivasan said LinkedIn will test a private note in a poster’s analytics when their writing may have come across as inauthentic, adding that AI and slop are not the same thing. | Assisted writing is explicitly allowed. Generic writing is what gets flagged, and you may be told before your reach shows it. |
| 360Brew ranking | The AI ranking system that replaced the old feed reads the text of a post to decide who sees it, weights saves heavily, gives hashtags little credit and tolerates links in the body. | Specific nouns in your sentences are how the system finds your audience. A generic draft gives it nothing to match. |
Ten tells that give a draft away
I marked up forty drafts with a red pen. These ten patterns appeared in more than half of them regardless of which tool wrote the draft.
| The tell | Why readers clock it | Do this instead |
|---|---|---|
| One-line hook, blank line, then the story | It is the opening of nearly every post above and below yours. | Start with the second sentence of your draft. It is usually the real one. |
| “It’s not about X. It’s about Y.” | Two of three tools produced it unprompted. Readers have seen it thousands of times. | Say Y once, with the number that proves it. |
| Lists of three ending in an abstract noun | “Speed, trust and clarity” means nothing to anyone. | Two items, or one item with a figure attached. |
| The lesson paragraph | “The takeaway?” tells readers you don’t trust them to get it. | End on the last fact and let the reader draw the line. |
| Emoji bullets | Ticks and rockets are the visual shorthand for generated. | Plain text. If you need bullets, you probably need sentences. |
| “Let that sink in” and “Read that again” | A beat without a thought. | Delete. The line either lands or it doesn’t. |
| Label openers: “Hard truth:” “Unpopular opinion:” | Announcing an opinion instead of having one. | Cut the label, keep the opinion. |
| Words nobody says on a call: journey, leverage, unlock, elevate, game-changer | Instant genre marker. | Use the verb you would use out loud: use, fix, ship, cut. |
| Every paragraph the same length and shape | Machine rhythm. People ramble, then snap. | Put a four-sentence paragraph next to a three-word one. |
| Closing engagement question | “What’s your take?” reads as a request for metrics. | End with what you will do next, or a question only your readers could answer. |
The five-step workflow
Total time on my posts: around twelve minutes, down from forty when I wrote from scratch. The model does step three. You do the rest.
1. Write the ugly version first. Two to five minutes. Voice memo, bullet points, a pasted Slack message, whatever is fastest. It must contain at least one number, one day or date, and one thing someone actually said. If your notes lack those, the model will invent them, and invented specifics read worse than none.
2. Build a voice file once. Paste three to five things you wrote without AI (emails count) and ask for a description of your patterns: average sentence length, how you open, punctuation habits, words you never use. Save the result. Every future prompt starts with it. This one step did more than any prompt trick I tried.
3. Draft with constraints, not adjectives. Never ask for “engaging” or “punchy.” Ask for a list of things the draft must not do (the ten tells above), a character range, and numbers kept exactly as given. The full prompt is further down.
4. Cut, don’t polish. Read the draft aloud once. Delete the first sentence and the last paragraph, then decide whether either deserves to come back. Check the first 140 characters on your phone, since that is about where mobile truncates and desktop gives you roughly 210. Swap every generic noun for the specific one in your notes.
5. Post, then handle the comments yourself. Comments allow 1,250 characters and display in full. Replies are where readers decide whether a person is on the other end. Never automate this part.
Before and after
Same notes, same tool. The first panel is the raw draft from a topic prompt. The second is what shipped after the workflow above. Red underlines mark tells. Yellow marks specifics that came from my notes.
Before: raw draft from a topic prompt Last week, a client told me something that stopped me in my tracks. “We don’t need more features. We need to understand the ones we have.” [see more fold, about 140 characters] It’s not about building more. It’s about building clarity. Here’s what I learned: ✅ Simplicity wins ✅ Listen before you build ✅ Clarity is a feature What’s the best feedback you’ve received from a client? 👇 #ProductManagement #Leadership #Growth | After: notes, voice file, red pen A client told me on Tuesday that they’d used nine of our forty-one features in six months. They pay for all forty-one. [see more fold, about 140 characters] I had been pitching the roadmap. She wanted a walkthrough of the settings page. So the Q4 plan changed. Two new features got cut. We are spending that time on a fourteen-minute onboarding video and a redesign of the settings page nobody visits. If you run a product with more than thirty features, pull usage by feature this week. Mine were worse than I had guessed. |
Above the fold, the raw draft offers a feeling and a quote that could belong to anyone. The edited post offers two numbers and a contradiction, which is the only reason anyone taps “see more.”
Prompts that held up
Three prompts did most of the work. Paste your voice file above each one.
Draft from notes Using the voice description above and only the notes below, write a LinkedIn post between 900 and 1,600 characters. Do not open with a one-line hook. Do not end with a lesson, a takeaway or a question. Do not use “It’s not X, it’s Y,” emoji, hashtags or bullet lists. Keep every number exactly as written. If a detail is missing, leave a gap in square brackets instead of inventing one. Notes: [paste] |
The de-LinkedIn pass Here is a draft. List every sentence that could appear unchanged in any other LinkedIn post. Do not rewrite them, only list them. Then list every claim in the draft that my notes do not support. |
Fold check Show me the first 140 characters of this draft on their own. Then tell me, in one sentence, what a reader learns from those characters alone. |
Which AI tools are worth using
The categories matter more than the brands.
| Tool type | Good for | Watch for |
|---|---|---|
| General assistants (ChatGPT, Claude, Gemini) | Drafting from notes when given a voice file and a do-not list. Free tiers exist; paid plans follow the model you want. | All three revert to the template the moment a constraint is missing. Differences between them were smaller than differences between my prompts. |
| LinkedIn’s built-in proofreader | Fixing typos and grammar inside the composer without touching structure. | It replaced the rewriting tool on purpose. Do not expect it to draft. |
| LinkedIn writing and scheduling tools (AuthoredUp, Taplio and similar) | Previewing the see-more fold, scheduling, per-post analytics. | Their hook generators produce tells one through three by design. Use the preview, skip the generator. |
| AI detectors (Originality.ai, Pangram) | A rough second opinion on a finished draft. | They return probabilities and flag human writing too. Writing to beat a detector produces stranger text than writing to be read. |
What worked and what kept breaking
What the tools did well • Drafting time dropped from about forty minutes to twelve per post. • Turning a rambling voice memo into paragraphs with the facts in the right order. • Tightening sentences over thirty words without losing the point. • Spotting claims my notes did not support, when asked directly. • Catching typos and doubled words in the last pass. | Where they kept failing • Every tool reverted to the template when one constraint was dropped from the prompt. • Numbers softened into “significant” or “a big jump” unless told to keep them verbatim. • Thin notes produced confident invented details, including a client quote that never happened. • Each request to make a draft “more engaging” made it more generic. • Emoji and a closing question crept back after two or three revision rounds. |
Final verdict
After three weeks and a spreadsheet I am slightly embarrassed by, this is where I landed. I still draft with AI. The posts that got real replies, the kind with a follow-up question instead of “Great post,” were the ones where the model had nothing to invent because my notes already held the number, the day and the sentence the client actually said.
The posts that flopped were the ones I wrote from a topic. “Write about product feedback.” Those came back sounding like everyone, because that is what the model was asked for. I could tell the difference by Friday of week one, and by week three so could the people replying with questions of their own.
So the rule I am keeping: AI gets my notes, my voice file and a list of things it is not allowed to do. It does not get to decide what the post is about, and it never gets the first line or the last. That takes me about twelve minutes a post instead of forty, and nobody has asked about the robot since.
