Every founder I know has tried it at some point. You sit down to write a LinkedIn post, you cannot find the right opening, you paste your bullets into ChatGPT or Claude and ask for a draft. The draft is fine. It is grammatical, it makes the points, it ends with a question. You publish it.
Eight hours later: 31 impressions, 2 reactions (your cofounder and your aunt), 0 comments.
This happens to almost every B2B founder in 2026, and it is not because the AI is bad. The frontier models in May 2026 can write a competent LinkedIn post in 800 milliseconds. The problem is that LinkedIn's algorithm, Google's helpful-content system, and most human readers all use roughly the same heuristics to detect AI writing, and an unedited model output trips all twelve of the patterns below.
This is the working list I use when I review every draft before publish at Conbound. Most of them are five-second fixes once you see them.
The Quick Answer
The patterns that get flagged are not "the post was written by an AI." They are "the post has the surface features of generic AI writing." Strip those features and you can ship AI-assisted content that ranks and gets engagement. The twelve below are the highest-signal ones we found across roughly 4,000 B2B posts.
1. Em dashes everywhere
This is the loudest signal in 2026. ChatGPT 4 and 4.5 use em dashes (—) at 6 to 9 times the rate of human writers. Once you see the pattern, you cannot unsee it. The fix is one keyboard shortcut: find every em dash, replace with either a comma, a colon, a period, or a parenthetical aside set off with commas. Your post will read more like a human in under 30 seconds.
The deeper reason: the em dash is a way for the model to maintain rhythm without committing to a sentence structure. A human writer would have made a decision (this is a subordinate clause, this is a separate sentence). The model dodges the decision with punctuation. Readers feel the dodge even if they cannot name it.
2. The "X, not Y" pivot
"It is not a content problem, it is a positioning problem." "They do not need more posts, they need a system." This pattern is so universal in AI-generated B2B content that I now reject any post that contains more than one instance of it.
The fix is to say the positive directly. "You have a positioning problem" is stronger than "you do not have a content problem, you have a positioning problem." The double-frame is rhetorically lazy, and B2B readers in 2026 have read it ten thousand times.
3. Three short sentences stacked for drama
"The data is live. The scoring is yours. The decision is now."
Models default to this rhythm because it produces high "punchiness" by token, but it has no information density. Combine two of the three sentences. Pick the one that carries the actual point and let the other two die.
4. The setup-colon-list
"Three things changed: speed, cost, and trust."
Every AI-written B2B post has at least one of these. The fix is to integrate the list into flowing prose: "Speed changed first, then cost, then trust." The colon-list is a visual tell because it always appears one paragraph after a fake-personal opening. Human writers use lists too, but they earn them with longer setup.
5. Wikipedia-tone passive constructions
"It can be observed that engagement is correlated with posting cadence." "It has been shown that founder-led posts outperform company posts."
Models love these because they sound neutral and authoritative. Real writers say "engagement goes up when you post more" and "founder posts outperform company posts." The active version is shorter, sharper, and harder to second-guess.
6. Zero proper nouns
This is the single highest-signal pattern across all twelve. Open any flagged post and count the named entities: companies, people, dates, dollar amounts. If you find fewer than three across 200 words, the post reads as AI even if you wrote every word yourself. Real writers anchor.
"Most B2B companies are not posting enough" reads as AI. "Notion posted 8 times in May, Datadog posted 23 times" reads as human. The rule of thumb: one named entity per 75 to 100 words.
7. The "navigate the complexities of" family
There is a list of about 40 phrases that have become AI tells in 2026 because the models reach for them at training-time-distorted rates. The ones I strip without thinking:
- navigate the complexities of
- in today's fast-paced
- delve into
- tapestry of
- myriad of
- robust and scalable
- cutting-edge
- state-of-the-art
- it is worth noting that
- it has become increasingly clear
- the importance of cannot be overstated
Each of them appeared in less than 0.1% of human-written B2B writing pre-2022. Each now appears in 8-15% of AI-generated B2B writing. Readers know. Sometimes consciously, more often as a vague feeling that "this sounds like ChatGPT."
8. The "moreover" cluster
"Moreover," "Furthermore," "Additionally," "In addition," "Thus," "Therefore" at the start of a sentence is a 90% AI signal. Human writers in 2026 connect ideas with "And," "So," "But," "Then," or by just running the new idea into the previous one without a transition. The five-syllable Latinate transitions are a model trained on academic corpus accidentally writing like a 1962 textbook.
The fix is to delete the word and capitalize the next one. The sentence almost always reads better.
9. The "In conclusion" close
If your post ends with "In conclusion," "To sum up," "In summary," or a paragraph that recapitulates the post you just wrote, kill it. Strong posts end at the last specific claim. Recap paragraphs assume the reader is going to forget what they just read, which is condescending and also bad for the algorithm. LinkedIn's recommendation system reads the last 50 characters of a post as a strong signal. Make those characters land.
10. Triple-adjective lists
"Comprehensive, robust, and scalable." "Strategic, thoughtful, and intentional." "Bold, creative, and human."
Three is the model's default cardinality for an adjective list because it sounds balanced. Real writers use one adjective. If you cannot decide between three, you have not figured out what you actually mean.
11. The fake personal opening that turns into a generic essay
"Last week, I was on a call with a founder who is trying to figure out how to scale her LinkedIn presence." Two paragraphs later: "There are four key reasons this matters in 2026."
The opening promises a story. The body delivers a listicle. Readers feel the bait and switch and bounce. Either commit to the story (single anecdote, specific names, specific tension) or open with the abstract claim directly. The fake-personal opening is the most common LinkedIn AI tell in mid-2026.
12. No specific opinion that someone could disagree with
This is the deepest one. Every LinkedIn post that gets engagement contains at least one claim that another reasonable person could argue with. AI-written posts default to consensus statements ("good content matters," "voice is important," "consistency wins") because the model is trained to avoid claims that might be wrong.
If a reader cannot find a single sentence in your post they could disagree with, the post is not worth engaging with. The fix is harder than the other 11 because it requires the writer to actually have an opinion. No tool can give you one. But every tool can help you put one on the page once you know what it is.
The rewrite pattern that works
For every draft I get back from an AI tool, I run three passes in this order:
Pass 1: find-and-replace. Em dashes to commas, "moreover" to deletion, "X, not Y" to a positive claim, smart quotes to straight. This pass is 90% mechanical and takes under a minute.
Pass 2: named-entity injection. Read each paragraph and ask: where could a specific name, number, date, or dollar amount go? Add at least one per 100 words. This pass takes 5 to 8 minutes and is the highest-leverage edit.
Pass 3: opinion check. Read the post and ask: what does this say that someone could disagree with? If the answer is nothing, rewrite one sentence so the answer is something. This pass takes 5 minutes and is the difference between a flat post and one that gets 30 comments.
We built Conbound to do these passes automatically, including the named-entity injection from Crustdata's index of 200M+ B2B companies. But you do not need a tool to do them. You need a checklist and 12 minutes per post.
Why this matters beyond LinkedIn
Google's helpful-content system, rolled into core ranking in March 2024, uses many of the same surface features to detect AI content at scale. Pages full of em dashes, "moreover" transitions, zero named entities, and triple-adjective lists are getting downranked across the board, not just on LinkedIn. The same checklist works for your blog. It works for your landing pages. It works for your sales emails.
The wave of "AI content is dead" predictions in 2024 was wrong. AI-assisted content is fine, even excellent, when it has been edited to remove the model fingerprints. The wave is just shifting. The companies that win in 2026 will be the ones whose AI-assisted content reads as fully human, because they invested ten minutes per post in the patterns above.
The honest position is this: the model is not the bottleneck anymore. The editor is. If you do not have an editing pass that catches the twelve patterns above, your AI tool is producing content that loses to no content at all, because zero impressions is better than 30 impressions on a post that signals "AI sludge" to a reader who now associates it with your brand.
FAQ
Will Google penalize AI-generated content in 2026?
Not for being AI-generated. Google's official policy is that content quality is what gets penalized, not the production method. In practice, AI-generated content that fails the helpful-content criteria (E-E-A-T, expertise, original perspective) does get down-ranked, and unedited AI output usually fails on at least two of those.
How do AI detection tools like Originality.ai actually work?
They look for surface features: punctuation patterns, sentence-length distribution, vocabulary entropy, transition word frequency. They are 60-80% accurate on raw model output and 30-50% accurate on edited output. The patterns in this post are the same ones the detectors use.
What is the fastest way to humanize an AI-generated LinkedIn post?
Run passes 1 and 2 from above (find-and-replace plus named-entity injection). It takes under 10 minutes and removes 70% of the AI signal. The opinion check is the deeper fix but the first two get you most of the way for daily LinkedIn writing.
Can I just train an AI on my own writing to fix this?
You can train on your voice (Conbound does this; so does Supergrow). It helps with rhythm and vocabulary. It does not fix the absence of specific opinion, because the model can only output the kinds of opinions it has seen you write before. The opinion check pass is still on you.
Which AI tools produce the most human-sounding LinkedIn content?
We ran the honest matrix on the four most-used B2B tools here. The short version: AuthoredUp is not an AI generator, so it is fine by default. Conbound (mine) runs the twelve-pattern strip pass at draft time. Taplio and Supergrow both fail at least seven of the twelve in their default output, so they require human editing before publish.