How human readers spot AI-generated LinkedIn posts

In our session data, most readers register an AI post before they finish the first line, no detection tool required. A "Here's the thing" opener, one sentence per line, a closing "What do you think? Drop a comment below." Those habits show up in aggregate data, and the feed algorithm treats every skim-past as a reason to show you to fewer people.

The Signs of an AI-Written LinkedIn Post Most Readers Catch Immediately

You can usually name the tell before you finish reading the first line. The opening is the loudest signal: a contrarian hook, a humble confession, or a flat shock statement designed to stop the scroll. In Adrian Vega's analysis of 500 AI-generated LinkedIn posts, 82% used just three identical opening structures. Regular LinkedIn readers register these the way an editor registers passive voice, not consciously at first, but fast enough to stop reading.

Phrase frequency is the second giveaway, and it is the one that survives any amount of topical variety. "Here's the thing" appeared 170 times across those 500 analyzed AI posts, roughly 1 in 3 posts, against a baseline of about once per 50 articles in normal professional writing. "Let that sink in" showed up at 28 times the normal rate, and "Read that again" at 22 times the normal rate. A single use of any of them reads as mechanical to someone who spends real time on the platform.

The closing line is just as predictable. "What do you think? Drop a comment below," or some variation that poses a question anyone could answer without knowing a single specific thing about the topic. We build tools in this space, and the closer is the part we watch most closely, because a post that invites generic affirmation rather than asking the reader to draw on their own experience is a structural problem, not a style preference. It tells you the post had nothing at stake.

None of this is random. These patterns reflect how language models are trained to produce professional-sounding content. The model learned that these phrases correlate with positive feedback in its training data, so it reproduces them at a steady clip regardless of who the author is or what the post is about. The result is content that pattern-matches to engagement without containing any of the friction that genuine engagement actually needs.

That is why the surface scan works so well. You are not detecting intelligence or its absence. You are detecting a small, repeating vocabulary of moves that one kind of writer makes and another kind almost never does.

ChatGPT LinkedIn Post Patterns: Formatting That Gives It Away at a Glance

The fastest tell is visual, and it lands before a single word is read. Ninety-one percent of the 500 AI-generated posts in Adrian Vega's analysis used one-sentence paragraphs with a blank line between each sentence. That produces a distinctive column of stacked fragments, and experienced readers associate the shape itself with AI posts. The eye recognizes the silhouette before the brain parses the meaning.

This is not a hunch dressed up as an observation. Peer-reviewed stylometric research indexed on PubMed Central (PMC11231544) identified 33 lexical, syntactic, and structural features that separate ChatGPT-authored text from human text with up to 100% accuracy at the document level and 92.3% accuracy at the paragraph level, using XGBoost ensemble learning. The point worth sitting with: detection works on structure alone, with no surface keyword matching. Strip out every banned phrase and the skeleton still gives the post away.

Human writing breaks the pattern in ways we can see in account analytics. When we run A/B tests across post types, the posts that hold readers carry the irregularities a model smooths away: variation in paragraph length, mid-thought asides, an incomplete sentence, a specific named detail only the author would have reached for. Those are hard for a model to reproduce consistently, because they come from someone thinking through an idea while writing it, not from a system generating tidy output. Uniformity is the tell. Mess, within reason, is the proof of a person.

In our own data, the downstream effect is visible in account analytics. AI posts with polished, uniform formatting tend to show high impression counts but low post-click rates. Human posts with conversational, variable formatting show lower impressions but post-click rates 2 to 3 times higher. That post-click behavior feeds back into LinkedIn's ranking system as an active engagement signal, which means the messy human post is quietly being rewarded for the same quality that makes it look less professional at a glance.

If you want one formatting check to run before publishing, it is this: does the post look like a list of captions, or does it look like someone wrote paragraphs? The first shape is the one the algorithm and the reader have both learned to discount.

What the Engagement Data Shows About AI LinkedIn Posts

The engagement penalty is real and it is measurable. AI-generated LinkedIn posts received 45% less engagement than human-written posts in Originality.ai's analysis of 2,726 posts published from December 2022 through October 2024. The gap is not flat across the platform, which is the part most guides miss. Human posts outperformed AI posts by 80% in Innovation and Strategy and by 73% in Marketing and Branding, the two categories where audiences are most tuned to personal voice and genuine perspective.

There is one clean exception. In Leadership and Inspiration, AI posts outperformed human posts by 75%. That sector structurally expects inspirational statements, so the model's reflex to produce them stops being a tell and starts being a fit. If your niche runs on motivational framing, AI posts blend in. Everywhere that readers come for a specific point of view, they stand out and pay for it.

Polish, counterintuitively, makes things worse. Posts scoring low on an AI-polish scale, 1 to 3 out of 10, averaged 2.1% engagement. Posts scoring high on polish, 8 to 10, averaged 0.4%. That is a 5-to-1 gap, and it is driven by how readers respond to recognizable AI structure, not by topic or timing. The more a post looks like flawless machine output, the harder readers behaviorally penalize it. Rough beats smooth.

The account-level cost compounds over a year. Richard van der Blom's Algorithm Insights 2025 Report found that views fell roughly 50%, engagement dropped 25%, and follower growth declined 59% year-over-year, with generic AI content named as a substantial driver. Read those three figures together and the lesson is blunt: AI posts do not just underperform on the individual post. They erode the distribution your whole account depends on.

Does LinkedIn's Algorithm Detect AI-Generated Posts?

LinkedIn does not officially penalize AI-generated content as a category, and this is the distinction most discussions get wrong. According to LinkedIn Engineering lead Hristo Danchev's March 2026 blog post, the platform deprioritizes content that fails to hold attention. The mechanism is engagement behavior, not creation method. AI posts get suppressed because they produce low dwell time and weak engagement signals, not because anything stamped an AI label on them.

Detection systems do exist. LinkedIn has built internal classifiers that correctly identified generic AI-generated content 94% of the time in early tests, and flagged posts are suppressed from recommendations rather than removed from the platform. In May 2026, LinkedIn rolled out updated detection measures aimed at posts that are generic, repetitive, or lacking a clear personal perspective. So the category does get watched, but the consequence is reduced reach, not a takedown.

Scoring starts before any human sees the post. LinkedIn's online classifiers label content at creation time, within roughly 200 milliseconds, sorting it into spam, low-quality, or clear. A post carrying AI-pattern formatting and AI-pattern phrasing enters that scoring step at a disadvantage from the first instant. The race is already half-lost before the post reaches a single feed.

The practical takeaway changes how you should think about the problem. LinkedIn is not primarily hunting for AI authorship. It is hunting for posts that readers scroll past without stopping, and AI posts produce exactly that behavior more reliably than human ones. That is the route by which they accumulate suppression. Fix the reader response and the algorithmic response follows, which is a more useful goal than trying to disguise a post from a classifier.

It also means the two layers reinforce each other. A post that trips the 94%-accurate classifier loses recommendation reach, and the audience it does reach skims past, which feeds the behavioral signal that suppresses it further. Detection and disengagement are not separate risks. They are the same outcome arriving twice.

Behavioral Signals: What Happens to an AI Post in the First Two Hours

The clearest post-publish separator we track is reply latency. AI posts tend to draw a burst of emoji reactions in the first 30 minutes, served by tier-1 connection notifications, and then almost no substantive comments. The reason is structural: the post does not pose a real question or take a position that requires genuine knowledge to answer. Reactions are cheap. Comments need something to push against, and AI posts rarely give the reader anything.

Genuine content behaves differently from the start. Accounts posting a specific claim or a personal stake generate first comments within 10 to 15 minutes and hold comment velocity for 2 to 3 hours. LinkedIn's Community-Focused Feed Optimization added contribution probability as a ranking objective alongside click probability, and it introduced freshness adjustments because prompt creator responses affect ranking. Reply latency is a documented algorithmic signal, not a practitioner superstition.

Connection-tier distribution exposes an AI post faster than any phrase-level check. On a well-written human post, engagement fans outward: tier-1 connections engage first in the 0 to 30 minute window, tier-2 follows as reshares propagate over 1 to 3 hours, and tier-3 impressions generate late engagement across 3 to 24 hours. AI posts from the same account show a flat, front-loaded curve, heavy tier-1 reaction and almost no tier-2 or tier-3 propagation, because the reshare-probability model scores the post low after seeing passive-only engagement from the first sample audience.

The weighting is what makes this expensive. LinkedIn's multi-objective ranking combines passive consumption signals like dwell time and clicks with active consumption signals like comments and reshares, and the active ones carry more weight. A member may be 100 times more likely to read a post than to reshare it, yet resharing receives higher weight because it generates more network engagement. AI posts that earn reads but not replies are structurally disadvantaged even when their raw impression counts look healthy.

Our session data lines up with what LinkedIn's feed engineering documents. Replying to comments within the first two hours generates roughly 30% more engagement across a post's lifecycle, and LinkedIn's own feed quality engineering documents that its virality prediction watches the temporal velocity of likes, shares, and comments in real time. The window is short and the signal is live. A post that has nothing to reply to inside that window simply forfeits it.

The first two hours are where an AI post quietly loses. It collects its emoji, fails to start a conversation, never propagates past the first ring of connections, and the algorithm reads all of that as a verdict.

Repeated AI Posts Create Suppression at the Account Level, Not Just the Post Level

The damage does not stay on the post that earned it. LinkedIn's 360Brew generative recommender, a 150-billion-parameter model based on LLaMA 3 and deployed in March 2026, processes more than a thousand of a member's historical interactions as a chronological sequence. It treats long dwells, likes, comments, and shares as an interleaved behavioral narrative. Non-engagement, scrolling past without interacting, is an active training signal via hard negatives, which means every skim-past on an AI post compounds its suppression rather than simply registering as nothing.

Because the model reads engagement history as a sequence, a run of AI posts builds what we call behavioral debt. A string of low-dwell, comment-poor posts progressively lowers the account's predicted engagement probability. The consequence is the part operators find surprising: subsequent posts, even genuinely good ones, receive a smaller initial distribution sample, because the model expects them to perform poorly based on recent history. You are no longer being judged post by post. You are being judged on your trailing pattern.

Recovery is possible, but it is slower than people hope. In our experience it takes a sustained run of high-engagement posts to reweight the model's read of the sequence, typically 3 to 4 weeks at a normal posting cadence. The suppression is not permanent, and it is not a manual penalty you can appeal. It is just a prediction that has to be earned back. A single strong post does not erase weeks of weak ones.

Underneath all of this sits a documented threshold. LinkedIn's feed engineering describes a dwell-time cutoff, Tskip, below which an update shows near-zero probability of generating any engagement. LinkedIn measures both passive scroll-past time, with at least half the update visible, and post-click reading time. An AI post that readers skim past without pausing falls under that threshold, and falling under it is itself the training signal that teaches the system to show the account's future content to fewer people.

That is the real cost of AI slop, and it is the one rarely priced in. It is not the single post that flops. It is that each flop teaches the ranking model a little more about how little to trust the next thing you publish.

How to Spot AI-Generated LinkedIn Posts in Your Own Writing Before Publishing

The opening line carries the most signal, so start there. If it leans on a contrarian hook, a humble confession, or a shock statement, rewrite it around a specific date, a named person, or a concrete situation from your own experience. A sentence that only you could have written is immune to AI-pattern recognition, because no model would generate that exact combination of specifics. The paragraph shape matters almost as much. If more than two-thirds of your paragraphs are a single sentence, restructure: a paragraph that develops one idea across 2 to 4 sentences is a marker of editorial voice that models consistently underweight, and it breaks the stacked-fragment silhouette that 91% of analyzed AI posts share. Breaking that silhouette also gives a reader a reason to slow down, and that dwell-time behavior is exactly what the algorithm rewards.

The closing line determines whether anyone comments. If the post ends with "What do you think? Drop a comment below," replace it with a question that forces the commenter to draw on specific knowledge or experience. In our data, posts with friction-inducing CTAs generate comment-to-impression ratios 4 to 6 times higher than generic closers, and those comments carry more weight in LinkedIn's contribution probability scoring. The CTA is not decoration. It is the single lever with the largest measured effect on whether the post starts a conversation.

A phrase check catches what the structural fixes miss. Search the draft for "Let that sink in" (28 times the normal rate in AI posts), "Read that again" (22 times the normal rate), and "Here's the thing." Their baseline in genuine professional writing is low enough that a single appearance reads as a mechanical tell to anyone who spends regular time on LinkedIn. Stylometric research confirms structural patterns alone enable detection at 92.3% accuracy at the paragraph level, so cleaning the vocabulary is necessary but not sufficient. The goal is output that passes a human-reader check, not just an automated tool.

Once those checks are done, the test becomes simple. Open the post the way a stranger would, give it two seconds, and ask whether anything in it could only have come from you. If the answer is no, the algorithm will reach the same conclusion, and faster.

Frequently asked questions

How can you tell if a LinkedIn post was written by AI without using a detection tool?

Look for three patterns that appear in the majority of AI-generated LinkedIn posts: opening phrases like 'Here's the thing' or confessional openers starting with 'I used to believe,' one-sentence paragraphs with blank lines between every line, and a generic closing like 'What do you think? Drop a comment.' In a study of 500 AI posts, 82% used just three identical opening structures and 91% used one-sentence-per-line formatting.

What words and phrases does AI overuse on LinkedIn that readers recognize?

'Here's the thing' appeared 170 times across 500 analyzed AI LinkedIn posts, roughly 1 in 3, compared to once per 50 articles in normal professional writing. 'Let that sink in' appeared at 28 times the normal rate and 'Read that again' at 22 times. Other common signals include 'I'll be honest with you,' 'nobody is talking about this,' and openers that frame the post as a revelation most professionals are missing.

Does LinkedIn's algorithm penalize AI-generated posts, and if so, how?

Not directly as a category. LinkedIn's algorithm suppresses content that produces low dwell time and weak engagement signals, which AI posts do more consistently than human posts. LinkedIn has also built internal detection with 94% accuracy for identifying generic AI content, suppressing those posts from recommendations rather than removing them. In May 2026, LinkedIn updated its detection to target posts that are generic, repetitive, or lack a personal perspective.

What behavioral signals does LinkedIn track that AI-generated posts consistently fail to produce?

LinkedIn tracks dwell time (including a minimum threshold below which a post generates near-zero engagement probability), reply latency from the author, and comment velocity in the first two hours. AI posts typically generate emoji reactions but not substantive comments, produce a flat engagement curve without second- and third-tier propagation, and fail to trigger the reshare signals that LinkedIn's multi-objective ranking weights most heavily.

Why do AI LinkedIn posts get fewer comments even when they get plenty of likes?

AI posts typically close with zero-friction CTAs that invite affirmation rather than requiring the commenter to draw on specific knowledge or experience. A post that anyone can respond to with 'Great point!' generates shallow engagement. Posts with a specific, slightly contestable claim or a question requiring genuine experience to answer generate substantive comments 4-6 times more often per impression, and LinkedIn's contribution probability scoring weights those comments more heavily.

How does LinkedIn's dwell time scoring work, and why do AI posts underperform on it?

LinkedIn measures two types of dwell time: passive scroll-past time with at least half the post visible, and post-click reading time. Below a documented threshold called Tskip, updates show near-zero probability of generating any engagement. AI posts with predictable formatting and recognizable phrase patterns cause experienced readers to scroll past quickly, producing dwell times that fall below this threshold and training the algorithm to suppress future posts from that account.

What is the 'golden hour' on LinkedIn and why do AI posts miss it?

The first two hours after posting are the window in which LinkedIn's virality prediction monitors the temporal velocity of likes, shares, and comments in real time. Replying to comments within this window generates approximately 30% more engagement across a post's total lifecycle. AI posts miss this window because they generate few substantive comments to reply to, and their authors often schedule posts without planning for real-time engagement during the critical early period.

How does posting AI content repeatedly hurt your LinkedIn reach over time, not just on one post?

LinkedIn's 360Brew ranking model processes a member's historical engagement as a chronological sequence. A series of low-dwell, comment-poor AI posts lowers the account's predicted engagement probability, so subsequent posts receive smaller initial distribution samples. Recovery typically takes 3-4 weeks of high-engagement posts to reweight the model's assessment. The suppression compounds: each poorly-performing post reinforces the pattern in the sequence the algorithm uses to predict future performance.

What structural patterns make AI LinkedIn posts recognizable to experienced readers?

The most consistent patterns are single-sentence paragraphs separated by blank lines (found in 91% of 500 analyzed AI posts), numbered lists without genuine analysis between points, a predictable three-part structure of hook, content, and CTA, and endings that pose generic questions. Stylometric research found these structural patterns enable AI detection at 92.3% accuracy at the paragraph level using 33 linguistic features, without any surface-level keyword matching.

How do you use AI to write LinkedIn posts that still generate real engagement and comments?

Use AI to draft structure, then rewrite the opener with a specific date, event, or named person from your own experience. Break the one-sentence-per-line pattern by combining related ideas into 2-4 sentence paragraphs. Replace any generic CTA with a question that requires specific knowledge to answer. Remove any phrase from the high-frequency anomaly list: 'Let that sink in,' 'Read that again,' 'Here's the thing.' The goal is output that passes a human-reader check, not just an automated detection tool check.

Sources and further reading

• LinkedIn Engineering's documentation on feed dwell time and the Tskip threshold
• LinkedIn Engineering's March 2026 post on the 360Brew generative recommender
• peer-reviewed stylometric research on structural AI text detection accuracy