Why every LinkedIn post is starting to sound the same

Scroll through LinkedIn for five minutes in 2026 and a quiet pattern becomes noticeable. The posts are different in topic: someone's observation on a product launch, someone else's reflection on a conference, a third person's take on industry hiring trends. The posts are not different in shape. The openers feel related. The middles have the same rhythm. The closings land in places that begin to rhyme.

Three years ago this was not the case. LinkedIn feeds were sloppier, messier, more recognizably individual. Writers sounded like themselves because they were doing the writing. The convergence is a recent development, and it has a specific cause.

均The shape of an average.

When a large language model is trained on billions of words of human writing, what comes out the other end is the middle of that distribution. The model has absorbed many writers' opener patterns and generates the one closest to the center. It has absorbed many writers' transitions and generates the most common. It has absorbed many writers' closers and produces the closer that would offend the fewest editors. The output is defensible. The output is also average.

Average is not the wrong word. It is the precisely correct one. The model's default output is the mean of its training data, adjusted for inoffensiveness. For a single writer producing a single post, this is fine. Average sounds safe. Average reads as polished.

The problem shows up in aggregate. When one thousand writers ask the same tool for drafts, the tool produces one thousand drafts that cluster near the same middle. Each individual draft is fine. The thousand together look like one voice in a thousand lightly altered suits.

LinkedIn currently has something like one billion users and an actively posting subset that keeps growing. A meaningful fraction of that actively posting subset now uses ChatGPT or Gemini or Claude as a drafting tool. The feed's overall shape is being slowly pulled toward the mean of three or four large language models' training distributions. The pull is not fast. The pull is steady.

型Three structural tells in generic AI output.

The convergence is not abstract. Three specific patterns show up in almost every AI-drafted LinkedIn post, and readers now recognize them even if they cannot name them.

One: the category-statement opener. "In today's fast-moving professional landscape…" or "As the world of work continues to evolve…" or "In a recent study…" These openers are the safe choice because they commit to nothing specific. They are also the openers most trained into the model by repetition across business-writing training data.

Two: the neatly balanced middle. The draft presents a claim, a counterclaim, a synthesis. Or three causes of the observed phenomenon, one at a time. Or the tension between two perspectives. This shape is reliable, generalizable, and present in most business-writing training data. Professional writers use it sometimes. The tool uses it almost every time.

Three: the aspirational closer. The post lifts at the end into a generalizable observation about the future, the state of the industry, or the importance of something. "As we navigate this evolving landscape, one thing remains clear…" The aspirational closer is the model's default because many published essays it has seen end this way. A specific professional writer might close this way once in ten posts. The tool closes this way in nine of ten.

A reader who scrolls past five drafts in a row, each from a different writer, each shaped by these three patterns, begins to feel what they cannot immediately name. The writers sound alike. Each post reads as competent. None of them reads as someone.

察What the reader pattern-matches, and when it stops working.

Readers on LinkedIn are extremely good at pattern-matching. They read fast, scroll fast, stop at what surprises them. The posts that stop the scroll are the ones that feel specific, rooted, and attached to the writer's actual situation. The posts that fail to stop the scroll are the ones that could have been written by anyone.

For the first year or two of widespread AI adoption in LinkedIn writing, readers were not yet calibrated. A well-structured AI draft read as competent professional writing. Calibration has since arrived. A version of the "AI smell test" now runs automatically in most readers' attention: does this post sound like someone actual, or does it sound like a tool that has read many posts?

This is not about readers being able to list the tells. They mostly cannot. What they experience is a faint loss of interest. The post is fine. The post is forgettable. The reader moves on without liking, commenting, or saving.

The economic consequence on LinkedIn is real. Posts that read as generic get fewer impressions, because the algorithm's own engagement signals reward posts readers stop on. Posts that read as specific get disproportionate visibility. The incentive, for any professional who wants their LinkedIn to do useful career work, is to avoid the convergence.

早Why this accelerates each quarter.

Three forces are pushing feed convergence faster, not slower, through 2026 and into 2027.

First, adoption is still climbing. The professionals who adopted AI writing tools early (the 2023 and 2024 cohorts) are now being joined by the bulk of the mid-majority: professionals who resisted AI for a year or two and are now trying ChatGPT for LinkedIn. Each month, more drafts on LinkedIn are shaped by the same three or four tools. The averaging effect compounds with more users, not fewer.

Second, the major tools update their outputs in correlated ways. When OpenAI ships a new ChatGPT version with different default cadence, millions of posts start using the new cadence the same week. The platforms that were individual voices are now, effectively, weather systems. The shapes of the feed shift together.

Third, generic AI tools do not maintain persistent voice profiles of their users. Every draft is produced fresh, from the middle of the distribution, with no memory of how this specific writer actually writes. Every draft pulls toward the mean. Nothing pulls away from it.

The only force pulling away from the mean is a persistent voice model that knows how this specific writer writes and insists on that specificity every time it generates a draft. This is what Onaji's Voice Profile exists to do. A writer using a voice-model-driven tool writes against the averaging force; the post lands in their voice, not the tool's middle.

As the feed converges over the next four quarters, writers whose posts sound like themselves will stand out more, not less. The averaging of the feed is not a disaster. It is an opening, for writers willing to use tools that preserve their voice rather than collapse into it.