Open LinkedIn today and you can feel it. The posts have started to rhyme. The same openers, the same neat contrasts, the same tidy little question at the end. AI did not just arrive on the feed; it gave the whole feed a house style.
That house style has a small reach cost, and it is older than most people think. We studied the impressions of every eligible post by thousands of authors, comparing each writer only against their own other posts, so audience size never enters the comparison. Four templated turns of phrase each pull a post a little below its author's own baseline, and they already did in 2025.
Each runs about 4% to 7% below the same author's normal reach in our English data; the steepest, the "stop X, start Y / the key is" advice frame, runs about 6.7% below. (The original analysis is Do AI posts get less engagement on LinkedIn?.)
So this page tells two stories: a map of the house style, and what actually costs reach in 2026. The second one moved in late July 2026, and the move was about substance more than phrasing: posts with no concrete detail and no lived experience lost the most ground. The uncomfortable part of the history is that almost none of these turns were invented by a machine.
They were invented by the best humans on the platform, refined for years because they worked. The machine only copied them, at scale, until they read as a stamp instead of a skill, and readers started discounting the stamp.
The recognizable shapes
If you have scrolled LinkedIn in 2026, you already know these by feel. Here they are by name. (The full breakdown of all eleven is in How to spot an AI-written LinkedIn post; this is the short tour.)
The em dash. The long dash dropped mid sentence. It sat under 2% of posts for years, then jumped to 9.5% in 2024 and 15.6% in 2025 as AI tools spread. It went from rare to everywhere in two years.
"Here's how / Here's what." The promise opener. Chris Donnelly (1.2M followers) opened a post with "Here's the breakdown:" and it pulled more than 23,000 likes. It works, which is exactly why it is everywhere.
"It's not X, it's Y." The contrast formula. "It's not about the cost, it's about the value." It is the single most common move among top creators.
The closing question. "What about you?" The reflex ask for comments, now so automatic it reads as a script.
The reveal bridge. "The result?" "Plot twist:" The mini cliffhanger before the payoff.
One creator, Allie K. Miller (1.6M followers), summed up the collective recognition in a single post:
"Insanely obvious signs you used AI: 'It's not X, it's Y', equal-length bullets, ending a post with a weird question, a parade of em dashes, certain words (harness, supercharge)." When a list like that gets thousands of likes, the patterns are not a secret anymore. The whole platform can see the seams.
The four turns that cost reach
Recognition is one thing. Reach is another. We measured the impressions of each post against the same author's other posts, so a creator's audience size cancels out of the comparison.
On that within-author basis, four specific phrasings each drag a post a little below its writer's own normal. The cost is older than 2026: rerun on a larger and more recent base, the same four turns already cost about as much between June and December 2025, and January 1, 2026 marks no break in our data.
Turn of phrase | What it sounds like | Reach cost, Dec. 2024 to May 2026 (English, within-author) | Same turn, June to Dec. 2025 |
Generic advice frame | "Stop X. Start Y." / "the key is" | −6.7% | −6.6% |
"Here's what nobody tells you" | −4.3% | −4.1% | |
Reveal bridge | "The result?" / "Plot twist:" | −4.8% | −3.0% |
"It's not X, it's Y" | −4.9% | −5.1% |
Read the table honestly. These are measured across our English posts, each author compared only to their own work, and all four clear statistical significance in both periods. The cost is real, it is measured within each author, and it is a second-order lever: audience still drives reach far more than phrasing does.
Cleaning these turns recovers a few percent on the most templated posts, not an order of magnitude. It will not double anyone's reach, and we are not going to pretend it does. Nor did their cost jump in late July 2026: at equal substance, the turns flagged by our detector cost about 4% before July 27 and show no extra cost we can detect since. What moved after that date is the substance of the post, which we come back to below.
These are the turns in the wild, reworded from real 2026 posts so no one is named:
Generic advice frame. "Stop describing the tool. Start owning the result." "Stop chasing likes, start solving problems." Replace the template with the concrete, topic-specific action.
The "Here's" opener. "Here's what nobody tells you when you run a team this size." Open straight on the substance instead of announcing it.
The reveal bridge. "Teams stitch a workflow out of five different apps. The result? The frontline gets lost." Chain the consequence directly: "...so the frontline gets lost."
The contrast formula. "That's not a branding question. It's a system question." State it plainly ("This is a system question") and skip the negate-then-reframe pivot.
Two other moves that often get lumped in with "AI writing" do the opposite. A candid "Honestly..." runs above its author's average in every period we measured (about +4% in 2025), and a closing question added about +3% in 2025. The P.S. or CTA sign-off is the one that changed: it carried a small bonus in 2025 and no longer does in our latest data, so keep it only when it says something.
Stripping the candid aside and the question out of a post to "sound human" would be the wrong move. Drop the four shared templates, and put the saved effort into having something specific to say.
Where they actually came from
This is the part that gets lost in the panic. These are not AI inventions. They are the signature moves of the most successful writers on LinkedIn.
We profiled 100 of the biggest creators on the platform (a median of roughly 79,000 followers). The "AI" moves are their moves:
Move now read as "AI" | Top creators who use it |
"It's not X, it's Y" | 100% (28% in nearly every post) |
"Here's how / Here's what" | 98% |
"The key is / Stop doing X" | 100% |
A question at the end | 98% |
Gary Vaynerchuk (5.9M followers) writes "It's not always how much money you make, it's how much you spend." Justin Welsh (853k) opens with "Here's." These posts win, thousands of likes at a time. The patterns work because they were refined over years by people who write for a living.
So why do they now read as a robot? Because of how language models learn. As the writer Ann Handley (511k followers) put it in a widely shared post: "AI models love the em dash because humans do.
They are trained on millions of human-written sentences." The model studied the best creators, absorbed their highest-performing moves, and now serves them back to everyone, all at once, in every post. What gives it away was never the move itself; it is the saturation, the same handful of turns repeated across the whole feed.
The backlash, and the over-correction
The community noticed, and the mood turned fast. Some of the most-liked posts of the year are now about the tells themselves. There are parody lists of "phrases that scream AI." There is a running joke about the em dash being put "on trial."
But the backlash misfired on aim, not on instinct. People started stripping em dashes, clean bullet points, and any sharp sentence out of their writing, and several well-known writers pushed back on the punctuation panic specifically.
Netflix co-founder Marc Randolph (380k followers) noted he has written with em dashes for nearly 50 years. The em dash is 400 years old, and a single dash is not a verdict.
They were right about the dash and wrong to extend the amnesty to everything. The reach data draws a sharp line between the two. A lone em dash is cosmetic and carries no measurable reach penalty. The four templated turns above are structural, and they carry a small one, as they already did in 2025.
So the readers tearing dashes out of their drafts were sanding the wrong surface. The templated turns cost a few percent, and a post with nothing concrete in it costs far more, a gap that has widened since late July 2026. The crowd went after the most visible tell and left the bigger problem in place.
The same author, two versions
The cleanest evidence sits inside individual creators, where audience is held constant by definition. We lined up a handful of authors' posts that carry one of the four turns against their own posts without them. The templated posts sat below.
A recruiter leans on the reveal bridge ("...companies bring it in-house too early. The result?"). A creator who writes about personal branding uses the contrast pivot ("That's not a branding question. That's a system question."). In both accounts, the posts that did better were the ones that opened on something specific and lived-in: a real hiring situation, what clients actually do before they get in touch.
That comparison is correlational, since topic and format vary inside one account too. The clean posts also differ in more than phrasing: they usually carry a detail or a story that the templated ones lack. Our larger 2026 data points the same way, with substance moving reach more than the turn of phrase does.
What LinkedIn is doing about it
The platform decided to weigh in. In May 2026 it announced it would demote "generic or repetitive" content that "lacks any real unique perspective," and reported flagging generic content with about 94% accuracy in early tests. It banned no specific phrase; the stated target is emptiness, the repeated and the templated. Our English data shows no break in the cost of the turns around that announcement.
On July 30, 2026, LinkedIn went further. It announced classifiers that identify AI slop and low-quality content, to show less of it in suggested posts and in posts from outside your network, and added a "Seems like AI slop" report button. In our data, the week of July 27 is when posts with nothing concrete in them started losing more ground. In our study of AI slop on LinkedIn (40,937 posts by 660 creators), posts with no concrete detail and no lived experience sat 13% below their author's average post from April 1 to July 26, 2026, and 23% below since. Among the 342 creators who published them on both sides of the date, they fell a further 15%. The extra reach a concrete detail brings jumped 21% at July 27, the largest move of the 24 dates we tested.
Phrasing did not follow the same curve. Against a comparable post by the same creator, with the same substance, a post rated 30 points more AI-like cost about 9% before July 27, and the turns flagged by our detector, "It's not X, it's Y" included, about 4%. Neither shows an extra cost we can detect since. The timing matches LinkedIn's announcement, but we do not see how LinkedIn ranks posts, so we do not claim a cause. (We unpack what that means for reach in Does LinkedIn penalize AI content?.)
Where That Leaves AI Writing on LinkedIn
So this is the state of play in 2026. A feed that converged on a shared set of moves, partly because models trained on the best creators, partly because everyone imitates the same ones, and a community that can now spot all of it.
The patterns are already thinning out: in our English data, the share of posts carrying one of the four turns fell from 26% in April 2026 to 14% in September. They no longer signal "good writer." They signal "wrote like everyone else."
The interesting writing now is the writing that does not rhyme with the rest of the feed. Not because it dropped the em dash. Because there is an actual person behind it. Since late July 2026 that shows in reach too: against a comparable post by the same creator, a lived experience brings 25% more impressions, up from 7% before, according to our AI slop study.
