What LinkedIn's algorithm does with AI-generated content

LinkedIn does not delete AI-generated posts. It quietly caps how far they travel. 360Brew, the 150-billion-parameter ranking model deployed on March 12, 2026, never asks how a post was written. It reads whether people stayed, came back, or scrolled past. And it remembers.

Does LinkedIn Use AI to Rank Posts, or Just to Build Them?

Yes, LinkedIn uses AI to rank posts, and since March 12, 2026 that ranking runs through one large model instead of a stack of narrow ones. What the model does not do is check whether a human typed the words. LinkedIn VP and Executive Editor Laura Lorenzetti confirmed the platform's position on AI-generated posts: they are not deleted. Their distribution is suppressed, so the post lands with the poster's first-degree connections and stops there. Your analytics still show a post. Your feed still shows a post. It just never leaves the room.

This is the distinction most coverage flattens. The ranking system does not stamp a post as AI-generated and dock its score. It scores behavioral performance: how long readers dwell, how fast engagement arrives, how likely a member is to skip. Generic AI content tends to fail those measurements, which is why the correlation between AI writing and low reach is real. The causation runs through reader behavior, not through a detector reading your syntax. The penalty is behavioral, not syntactic, and that single sentence changes where you intervene.

LinkedIn's stated policy is unusually direct on this point: 'AI-assisted content is welcome as long as it carries original ideas or sparks real conversation.' Read that as a product decision rather than PR. The platform sells AI tooling to its own members. It cannot punish the use of AI without punishing its customers. What it can do is raise the behavioral bar and let the bar sort the output.

Which produces an awkward situation nobody at LinkedIn seems eager to name. LinkedIn embedded AI writing into its own post composer, then tightened behavioral scoring in the March 2026 update, raising the bar for the exact content its own product was helping members generate. We build automation tooling for this platform, and the accounts we watched lose the most reach in the weeks after the update were not running exotic gray-market scrapers. They were writing competent, clean, entirely unremarkable posts with assistance, on a fixed schedule, in one format.

The practical consequence is that the same AI tool produces two different outcomes depending on what you feed it. A draft built from a generic prompt about a generic topic gets capped. A draft built from a specific client outcome, a number you measured, or an argument you would defend in a room full of people who disagree gets distributed. The failure mode we see most often has a shape worth naming: the interchangeable draft, a post that any account in your industry could have published without changing a word. The model has no category for that. It does not need one. Readers skip it, and the skip is the signal.

Behavioral Performance, Not AI Detection, Determines LinkedIn Reach

The clearest primary source on how reach is decided is LinkedIn's own engineering blog on feed dwell time. The feed ranking model estimates P(skip), the probability that a member's dwell time on a given update falls below a natural skip threshold the team calls T_skip. P(skip) sits alongside P(action), the probability of an engagement, and the expected network effects from shares. LinkedIn reported that A/B testing the dwell time model produced fewer skipped updates and more time spent engaging. Skipping is not the absence of a signal. It is a signal, and it is scored.

So when a post gets restricted to the poster's immediate neighborhood, nothing was flagged. What happened is arithmetic. The post went to a small initial cohort, the cohort's dwell readings came back below threshold, and the model concluded that pushing it further would generate more skips than engagement. Every feed slot it occupies is a slot that could hold something with better predicted behavior. The suppression is not a punishment applied to your post. It is your post losing a bid.

LinkedIn documents the graduated response in five levels: demotion in feed ranking, restriction to the poster's immediate neighborhood, limited surfacing across the site, making the content undiscoverable site-wide, and disabling the poster account in extreme cases. Most AI-assisted content that underperforms never gets past level one or two. The account-level outcomes at the far end of that spectrum are reserved for behavior that looks like abuse, not for writing that reads flat.

There is no label to dispute and no appeal queue, because there is nothing to appeal. A predicted behavioral score is not an accusation. We get asked some version of 'how do I get my account whitelisted' often enough that it deserves a direct answer: there is no list. The only exit from a low-distribution pattern is publishing content that produces the readings the model rewards, repeatedly, until the trajectory moves.

One pattern from our own data is worth flagging because it fools people into thinking they got away with it. LinkedIn measures dwell on the feed and dwell after the click as separate things. Generic AI posts frequently clear the on-feed scroll threshold, because the hook is competent and the formatting is clean, then fail the after-the-click measure when readers open them and close them within a second or two. The reactive classifier picks up that click-bounce pattern during its scan of viral spread signals, which runs every few hours rather than continuously. This is why some AI posts look normal for the first two to three hours, then quietly plateau with no second wave of distribution. The first-wave numbers were real. The second wave was cancelled after review.

Inside LinkedIn's 360Brew AI Ranking System

On March 12, 2026, LinkedIn replaced its fragmented ranking models with a unified LLM-powered retrieval and ranking system called 360Brew: a 150-billion-parameter decoder-only foundation model built on LLaMA 3 and fine-tuned on LinkedIn's proprietary networking and interaction data. Before this, different surfaces ran different specialized models with different features. Now one model handles retrieval and ranking together. That consolidation is the reason the March update felt less like a tweak and more like the ground moving.

The important architectural change is what the model can read. Earlier ranking systems inferred quality from engagement reactions after the fact. 360Brew evaluates semantic content quality directly, which means it can score a post before the crowd has voted on it. LinkedIn's description of how it identifies generic or AI-generated text is specific: absence of concrete details, absence of personal experience signals, and template-like structure. Cadence is a quality feature now. A post whose paragraphs are all the same length, whose sentences all land in the same rhythm, and whose ending is a question aimed at nobody in particular is legible as a template to a model built on next-token prediction. Predicting text is precisely what it was trained to do.

360Brew also includes a Generative Recommender, a sequential model that processes more than 1,000 of a member's historical interactions using a transformer with causal attention. This treats feed activity as an ordered sequence rather than a bag of independent events. The consequence for publishers is not obvious until you sit with it: your posting history is a trajectory the model has a representation of, and each new post is scored in the context of that trajectory, not in isolation.

Underneath, the infrastructure is fast enough that the scoring is close to live. LinkedIn's nearline pipeline refreshes content and member embeddings within minutes of post creation or an engagement change, and candidate retrieval for the feed completes in under 50ms end to end. Because the embeddings are LLM-derived, semantically related topics link without sharing keywords. Topic authority operates at the level of meaning rather than string matching, so writing about the same subject in different words still accrues to the same authority.

Here is the pattern we see in our own automation data, and it is the part most guides miss. Accounts that publish AI content with no variation in post length, sentence structure, or cadence show a consistent signature in the first 60 to 90 minutes. The initial test cohort, roughly two to five percent of the network, receives the post. Dwell comes back low relative to that account's historical baseline. The GR model's sequential representation of the account shifts. Then the next post starts with a lower prior probability of distribution expansion, before a single engagement signal has been collected on it. The suppression is cumulative, not episodic. You are not being graded post by post. You are being graded on the curve you are drawing.

Why Some AI Posts Clear LinkedIn's First Classifier but Fail the Second

LinkedIn's spam and quality enforcement runs in two layers with different technology and different timing. The proactive layer uses deep neural networks served through TensorFlow and Pro-ML, running synchronously at post creation inside a 200ms latency window, before the post enters feed distribution at all. The reactive layer uses Boosted Trees models that scan for viral spread patterns every few hours after publication. Two models, two clocks, two chances to be wrong about you.

The proactive classifier is the one that surprises people, because it evaluates member-side features alongside content signals: account tenure, network diversity, content polarity, and engagement velocity history. A thin account posting at high frequency draws a higher P(spam) score independent of what the post says. Suppression can therefore originate at creation time, in the 200 milliseconds before any human has read a word. If you have ever published something you were proud of and watched it die instantly with no explanation, this layer is a likely culprit, and the cause was your account profile rather than your paragraph.

The reactive scanner exists to catch what the first layer misses. A post that clears creation-time filtering can still be downgraded hours later if its engagement velocity signature matches known spam cascades: a burst that arrives too uniformly, from too narrow a slice of the network, at too consistent an interval. This is the machinery behind the two-hour plateau described earlier. The post was not reversed. Its second-wave expansion was declined.

For the AI-content question specifically, LinkedIn's approach is what the company frames as AI solving AI. Human editors label thousands of posts, and classification models train on those annotations to identify patterns that distinguish original thinking from generic output. LinkedIn claimed 94% accuracy in early tests. It has not published a false-positive rate, and that omission is the number that matters most to anyone writing in a second language, in a regulated industry, or in a house style that reads formal by necessity. Accuracy without a false-positive breakdown tells you almost nothing about your personal risk.

The enforcement stack does work in aggregate. LinkedIn reports a 48% reduction in spam and low-quality content impressions from the combination of creation-time classifiers, viral-spread monitoring, and member reports. A number that large is not achieved with a light touch, and systems tuned to cut half of anything tend to cut things adjacent to the target.

The automation-specific failure mode is the one we can speak to directly. When tools rotate between AI drafts at high frequency without behavioral variation, same publishing window, similar word counts, no reply engagement in the first hour, the proactive network responds to those account-level features before any reader sees the post. It is not evaluating whether the writing is good. It is evaluating whether the account behaves like a person. Consistency, which every scheduling guide recommends as a virtue, is legible as machinery when it extends to word count and minute of day.

LinkedIn Built AI Writing Tools Into Its Composer, Then Crowdsourced Their Detection

In July 2026, LinkedIn shipped a member-facing 'Seems like AI slop' report button. In the same period it pulled back its own 'Enhance your post' AI writing feature and replaced it with a more conservative proofreading tool. Both moves happened quietly, and coverage treated them as two separate product notes. They are one decision.

The contradiction deserves a plain statement. The platform that built AI writing into its composer is now asking members to flag AI writing so it can train better suppression. Some of the accounts being reported are accounts that used LinkedIn's own suggestions to write the post. Pulling 'Enhance your post' back to proofreading is an admission that the feature was producing exactly the output the ranking model had been retrained to demote. It is rare for a platform to disagree with itself this legibly in public.

The report button is not decorative. It feeds the reactive model's member-report signal layer, the same layer used for spam cascade detection. That has a consequence practitioners should think about before they dismiss it: a small number of readers who recognize a templated format can trigger reactive review on an otherwise clean account. Not a ban, not a strike, just a downgrade in the second wave. You will never be told it happened.

The mitigation is structural variety, not vocabulary variety, and this is where most advice gets it backwards. Swapping out the word 'delve' does nothing, because neither the classifier nor the human reporter is matching on your dictionary. Both are matching on shape. The hook line, the line break, the bulleted list of three, the one-word paragraph for emphasis, the question at the end: that arrangement is the fingerprint. We have watched posts with genuinely original content get reported and capped because they wore the uniform. Vary the shape across posts. Some should be a single block of prose. Some should open in the middle of an argument with no hook at all.

LinkedIn's stated goal remains what it says it is: suppress content that lacks original insight, expertise, or perspective, without punishing the use of AI tools. That policy is coherent on paper. In production, the line between the two is drawn by a behavioral model and a report button, neither of which reads policy. If your writing produces the readings of low-value content, you are treated as low-value content, and the fact that you had something real to say does not enter the calculation.

The Engagement Signals LinkedIn Uses to Rank Posts in 2026

Engagement is weighted non-linearly, and the gap between signal types is wide enough to change how you think about a post's goal. AuthoredUp's analysis of more than 621,833 posts suggests a save carries roughly 5x the weight of a like, and a substantive comment of 15 or more words carries approximately 15x. These are not officially disclosed figures and should be read as directional, but they line up with the logic of the P(action) model: the signals that cost a reader more attention predict future attention better. Chasing likes optimizes for the cheapest signal on the board.

The macro picture explains why so many accounts felt the change at once. After the March 2026 update, post views declined approximately 50% year over year, engagement dropped 25%, and follower growth declined 59%, per Dataslayer's April 2026 analysis and Whitehat/Socialinsider 2026 benchmarks. In the same period, content creation on the platform rose 14%. More supply, stricter behavioral scoring, same finite feed. The competition is not for attention in general. It is for the small number of slots that clear the quality bar.

P(skip) is not scored in a vacuum either. It is read against the account's historical baseline, which means an account with two years of consistent publishing and a new account with sparse history receive different treatment on identical content. This cuts both ways. A long history of strong dwell buys tolerance for an experiment that lands badly. A short history means every post is load-bearing, and the account that posts fifteen times in its first month is teaching the model a curve it will honor.

The single most leveraged moment in a post's life is earlier than most people schedule for. LinkedIn's embedding refresh pipeline updates a post's representation within minutes of creation and again after each meaningful engagement event. A first comment or save arriving in the first 5 to 10 minutes, particularly from a non-follower or a second-degree connection, updates the embedding while the ranking model is still inside the candidate-expansion decision window. Later engagement of the same volume arrives after the decision. In our workflows, the reliable difference between a post that expands and one that does not is often nothing more than whether a genuine first response landed before the window closed.

None of this resets between posts. The sequential GR model processes the full trajectory, so a sustained run of structurally similar posts with weak engagement compounds into deeper suppression rather than a series of independent disappointments. The corollary is the encouraging half: a run of posts that hold attention compounds the same way. The model is not scoring your last post. It is scoring the sequence your last post extends.

What AI-Content Guides Get Wrong About the 360Brew Update

Most coverage treats 360Brew as a black box that lowered everyone's reach in March. That framing is not wrong so much as useless, because it implies nothing you can do. The mechanism is specific and it points somewhere. The Generative Recommender inside 360Brew processes more than 1,000 historical interactions as an ordered trajectory with causal attention, which means publishing structurally similar posts repeatedly deepens suppression rather than resetting the count each time. A guide that says 'reach is down, post more' is recommending the thing that makes it worse.

The second gap is the big one. No competing page we have read explains that the suppression is behavioral rather than syntactic. The platform does not label a post as AI-generated and demote it. It reads dwell time, P(skip), and engagement velocity, which are downstream outcomes, then demotes on those. This is not a semantic quibble. If you believe the mechanism is detection, you invest in disguise: rewording, humanizer tools, prompt tricks to sound less like a model. If you know the mechanism is behavior, you invest in the thing that changes reader behavior, which is having something specific to say. One of those strategies survives the next model update.

The 94% accuracy figure gets quoted everywhere, and the behavioral fingerprints feeding that classifier get quoted nowhere. From what we observe in automation data, the signals a model trained on human-annotated posts would weight most heavily are the ones humans use when they eyeball a post and think 'a machine wrote this': uniform post length across an account, identical cadence, no variation in sentence rhythm, and the absence of micro-engagement such as return visits or saves from people who do not follow you. Real posts collect odd, asymmetric attention. Templates collect flat attention.

The two-wave architecture is also flattened in most coverage into a single filter. It matters for strategy that a post can pass the proactive classifier at 200ms and still be caught hours later by the reactive model if its engagement velocity matches a spam cascade signature. Single-wave thinking produces incomplete mitigation: people optimize the draft, watch the first hour, see healthy numbers, and conclude they are clear. The second review has not run yet.

One number reframes the whole competitive picture. Originality.ai sampled 5,000 public LinkedIn posts of at least 100 words in July 2026, across 90 topic-and-date searches, and classified 81.2% as 'Likely AI' at a 15% AI allowance threshold. Take that seriously and the pessimistic reading inverts. Your specific, experience-backed post is not competing against the full publishing volume of the platform. It is competing against the fraction that clears the behavioral bar, and that fraction is much smaller than the raw post count suggests. The bar went up, and almost nobody is clearing it.

How to Publish AI-Assisted LinkedIn Posts Without Triggering Suppression

Start from LinkedIn's own standard, because it is a usable test rather than a slogan: AI-assisted content is welcome when it carries original ideas or sparks real conversation. Translate that into an editorial check. Does the post contain a specific detail, a named experience, or a perspective that could not be produced by a generic prompt on this topic? If a competitor could publish your post verbatim under their own logo without anyone noticing, the model has nothing to score and readers have no reason to stay.

Vary structure before you vary vocabulary. The proactive classifier and the human with the report button are both matching on template shape, so changing the hook format, the body arrangement, and the closing type across your posts reduces the recognition signal in a way that word substitution does not. Write one post as a single unbroken paragraph. Start the next in the middle of an argument. End one with nothing, no question, no call to action, because a post that does not ask for engagement is unusual enough to be read.

Treat the first 60 to 90 minutes as a model-updating event rather than a distribution gate. The embedding refresh runs within minutes of creation and again after each meaningful engagement change, and the GR model updates its representation of your account's trajectory on that early behavioral data. Publishing when your audience is awake matters less for the raw impressions than for what the model learns about you in the window when it is still deciding. A genuine first response inside the first 5 to 10 minutes, ideally from someone outside your follower graph, is the highest-value intervention available.

Do not run high-frequency publishing without behavioral variation. Same window, similar word counts, no reply engagement in the first hour: those are account-level features the proactive network scores before any reader sees the post. This is the piece of advice that costs us something to give, because we sell automation, and the honest version is that the scheduling discipline most tools optimize for is the exact regularity the classifier reads as machinery. Irregular cadence with real interaction beats perfect cadence with none. If your tool cannot vary the timing, vary it yourself.

The editing pass that catches the most damage takes about five minutes. Read your draft and mark every sentence a language model could have written from a generic prompt with no knowledge of your specific situation. In a typical AI-assisted draft, that is most of the middle. Those sentences are the suppression risk, not because a detector will catch them but because nobody will finish reading them. Replace each one with a specific client outcome, a number you measured, or something you noticed in your own work that you have not seen written down elsewhere. If you cannot fill the space that way, you did not have a post yet. That is a better outcome than publishing one and teaching the model something about your account you will spend a month undoing.

Frequently asked questions

Does LinkedIn penalize AI-generated content, or does it penalize low-quality content that AI tends to produce?

LinkedIn penalizes low-quality content, not AI-generated content specifically. The algorithm reads behavioral signals: dwell time, skip probability, and engagement velocity. Generic AI content tends to fail those behavioral tests because it lacks specific detail and personal experience, which reduces the time readers spend with it. The penalty applies to the behavioral outcome, not the origin of the writing.

How does LinkedIn's algorithm actually detect AI-written posts in 2026?

LinkedIn uses machine learning systems trained on human-annotated posts to identify patterns distinguishing original thinking from generic content. The proactive classifier runs at post creation time with 200ms latency. A second reactive model scans engagement velocity patterns every few hours. LinkedIn claims 94% accuracy in early tests but has not published a false-positive rate. A crowdsourced 'Seems like AI slop' report button introduced in July 2026 also feeds into the reactive model.

What is LinkedIn's 360Brew model and how does it change what gets suppressed?

360Brew is a 150-billion-parameter decoder-only foundation model built on LLaMA 3, deployed on March 12, 2026. It unified LinkedIn's fragmented ranking models into a single system that evaluates semantic content quality and processes more than 1,000 of a member's historical interactions as an ordered trajectory. The sequential nature means repeated publication of structurally similar posts compounds suppression over time rather than being treated as independent events.

What behavioral signals cause LinkedIn to stop distributing a post beyond first-degree connections?

The primary signals are P(skip), the estimated probability that dwell time falls below LinkedIn's T_skip threshold; low P(action), meaning weak probability of any engagement; and a mismatch between the post's early engagement velocity and the account's historical baseline. LinkedIn's five-tier treatment spectrum applies graduated responses from feed demotion to restriction to the poster's immediate neighborhood and beyond.

What engagement signals does LinkedIn use to rank posts in 2026, and which matter most?

LinkedIn's ranking model weights saves approximately 5x a like and substantive comments of 15 or more words approximately 15x, per AuthoredUp analysis of 621,833+ posts. These are not officially disclosed figures. Official signals include P(action), P(skip), and expected network effects from shares. The first engagement from a non-follower or second-degree connection is especially weighted because it updates the post's embedding during the distribution-expansion decision window.

Does AI-assisted content get less reach on LinkedIn, or only fully AI-generated posts?

LinkedIn's official policy distinguishes the two: AI-assisted content carrying original ideas or sparking real conversation is acceptable. Only content that lacks original insight, expertise, or perspective is targeted for suppression. If an AI-assisted post generates genuine dwell time and engagement, the algorithm treats it identically to a human-written one. The problem is that AI-generated templates tend to produce the behavioral failures the algorithm penalizes.

Why did LinkedIn post reach drop 50% after the March 2026 algorithm update?

After the March 2026 360Brew rollout, post views declined approximately 50% year-over-year, engagement dropped 25%, and follower growth declined 59%, per Dataslayer April 2026 analysis and Socialinsider 2026 benchmarks. Content creation rose 14% in the same period. The new model's stricter behavioral scoring, combined with rising content volume, compressed the quality-signal competition that determines which posts get expanded distribution.

What is the difference between LinkedIn's proactive spam classifier and its reactive viral-spread scanner?

The proactive classifier runs synchronously at post creation with 200ms latency, evaluating content signals and account-level features like tenure and network diversity before the post enters feed distribution. The reactive Boosted Trees scanner runs every few hours and monitors for viral spread patterns matching known spam cascades. A post can pass the proactive classifier and still be demoted by the reactive scanner if its engagement velocity triggers a spam-cascade signal hours after publication.

Does using a scheduling tool or automation reduce LinkedIn organic reach?

Automation reduces reach when it produces account-level behavioral patterns the proactive classifier flags: same publishing window, similar word counts, no reply engagement in the first hour. The classifier evaluates these account-level features alongside content signals and can assign a higher P(spam) score before any reader sees the post. The risk is not the scheduling tool itself but the predictable, low-variation pattern it enables across multiple posts.

What content patterns trigger LinkedIn's AI-slop suppression, and how is the demotion applied?

Patterns that trigger suppression include uniform post length and cadence, the hook-bullet-CTA structure AI writing tools default to, absence of micro-engagement such as saves from non-followers, and low dwell time relative to the account's historical baseline. Demotion follows a five-tier spectrum: feed ranking demotion, restriction to the poster's neighborhood, limited site surfacing, site-wide undiscoverability, and account disabling. The penalty is cumulative through the GR model's sequential trajectory representation, not episodic.

Sources and further reading

• LinkedIn Engineering: how 360Brew and the Generative Recommender work
• LinkedIn Engineering: dwell time and the P(skip) ranking model
• LinkedIn Engineering: five-tier content suppression and spam detection metrics