LinkedIn's originality scoring and what low-signal posts cost

LinkedIn deployed a 150-billion-parameter ranking model called 360Brew that reads your posting history as an ordered sequence of over 1,000 interactions. It is not scanning for a list of AI-associated words. It is scoring the semantic quality of what you publish against the behavioral signals your audience returns. A post that clears a vocabulary check but fails on sentence-length variance, low specificity, or vague first-person framing still lands in the suppressed tier. When that pattern holds across three consecutive posts, the account-level penalty compounds before the fourth post is ever submitted.

How LinkedIn's AI Content Detection Caps Your Reach

LinkedIn VP and Executive Editor Laura Lorenzetti confirmed the system in a statement reported by Entrepreneur: flagged posts are not removed. Their distribution is capped to roughly the poster's first-degree network. Second-degree connections and hashtag followers, where most reach growth occurs, never see the content. The post exists. It just stops spreading.

LinkedIn officially targets three content categories for this reach suppression: generic AI-generated posts lacking original insight, automated AI commenting tools, and attention-bait videos engineered for engagement without substantive value. The framing matters because none of these categories is defined as "AI-generated" in isolation. A post can be AI-assisted and clear the classifier. A post can be human-written and get flagged. The criterion is signal quality, not authorship.

The detection system uses machine learning models trained on human-annotated examples. LinkedIn describes this internally as "AI solving AI." The same engineering discipline that has driven LinkedIn's feed quality work since 2017 now applies specifically to identifying low-originality content.

Content creation on LinkedIn grew 14% year-over-year, a figure LinkedIn VP Lorenzetti attributes directly to AI adoption. That surge is the stated justification for building the suppression system. More content makes the feed more competitive, not more useful. LinkedIn's response was to tighten the distribution gate, not expand capacity. The practical effect for any individual creator: doing what everyone else is doing is now the fastest path to reaching almost no one.

360Brew and the Originality Score: What the 150-Billion-Parameter Model Measures

360Brew replaced thousands of prior specialized ranking systems with a single LLM-based model. A LinkedIn-authored research paper published in January 2025 describes its architecture: it treats each member's 1,000-plus historical interactions as an ordered sequence, scoring content semantically rather than by isolated post-level signals. Prior LinkedIn ranking systems evaluated posts in relative isolation. 360Brew evaluates a post in the context of what the viewing member has read, reacted to, and scrolled past.

This has a direct consequence for AI-generated content. When the model surfaces a post to a given member, it is predicting the probability of meaningful engagement: will this person dwell on it, comment substantively, save it, or share it? If the post's text patterns match what that member has historically skipped, the predicted engagement probability falls, and the post does not surface.

LinkedIn's official feed ranking page provides the signal weights: dwell time is weighted 2.8x heavier than likes. Comments of 15 or more words boost reach 2.5x more than short replies. Saves and shares outweigh likes as quality signals. The most counterintuitive weight is the passive one. Choosing to keep reading, without clicking anything, carries more algorithmic weight than the active gesture of clicking Like.

Profile-content alignment carries the heaviest algorithmic weight in the ranking formula overall. A post whose topic, vocabulary, and framing match the creator's headline and About section is treated as more relevant to the audience following that creator. This is where generic AI content most often fails: it does not sound like the person whose profile it is posted under, and the model scores that mismatch.

Does LinkedIn Penalize AI-Generated Content, or Does It Cap Reach Instead?

Classification starts before most creators finish reading what they have submitted. LinkedIn's SVM-based classifier labels a post as spam, low-quality, or clear within 200ms of submission. That initial gate runs synchronously. A deep neural network then runs asynchronously for deeper analysis on content that passes or falls in the borderline range.

LinkedIn's own A/B tests documented a 48% reduction in spam and low-quality content impressions via virality predictor classifiers alone, as reported in LinkedIn's 2017 engineering post on feed quality. The detection architecture has evolved since then, but the underlying logic, using predicted engagement to gate distribution before it starts, has not.

What "low-quality" means in practice is not defined by a word list. The synchronous classifier flags structural patterns: uniform sentence length, low specificity, dense transition phrases. A post with no AI-associated buzzwords but those structural properties is labeled low-quality by the same gate that catches an unedited AI draft.

The post does not disappear after flagging. Its distribution ceiling is simply lowered before it ever reaches second-degree connections. From the outside, the creator sees normal-looking post metrics at low volume. Without a benchmark for expected reach, most creators do not realize anything is wrong.

In November 2025, LinkedIn updated its platform Terms to allow training its generative AI models on member public posts and profile data by default, with opt-out required. This means the AI-detection training corpus is continuously refreshed from the same feed it is policing. As AI writing patterns shift, the detection models update in kind.

The Six Signals LinkedIn's Originality Classifier Actually Scores

Reverse-engineered from observed post performance, the six signals the originality classifier scores are: vocabulary repetition rate, sentence-length variance, transition phrase density, specificity of examples (numbers, names, dates), tone consistency, and presence of personal markers anchored to lived professional experience. These are not an official LinkedIn disclosure. They are derived from watching what moves posts out of the suppressed tier and what keeps them there.

Sentence-length variance is the signal most tools get wrong. A post with zero instances of "delve" or "leverage" but composed entirely of sentences between 12 and 16 words triggers the same low-signal classification as a post dense with AI-pattern vocabulary. The classifier is scoring cadence uniformity, not just word choice. A banned-word audit does not touch this signal.

Personal markers must be specific to score well. "I learned this the hard way" carries no verifiable professional context and scores low on personal-marker density. "I learned this when we missed our Series A close date by 11 days" anchors the first-person claim to a dateable, role-specific event. We observe this distinction consistently across accounts. Vague "I" statements do not substitute for grounded ones.

Transition phrase density operates independently of vocabulary signals. A high rate of "however," "therefore," and "additionally" is a structural AI pattern the classifier penalizes on its own terms. This is why lightly edited AI drafts still get flagged: the writer removes the vocabulary tells but leaves the connective tissue intact, and the density signal fires anyway.

Account-Level Penalty Accumulation: When Three Bad Posts Cost You a Fourth

An Originality.AI study analyzed 3,368 long posts from 99 top LinkedIn profiles covering January through November 2025 and found 53.7% were classified as likely AI-generated. Engagement gaps between human-written and AI-generated posts varied significantly by sector: in healthcare, human-written posts outperformed AI posts by 44%; in government, the gap was 40%. The sector variance matters because it suggests audience sophistication and content category interact with how the signal weights play out.

The account-level mechanic is the one most content audits miss. We have observed that accounts recycling the same five to seven abstract nouns across three or more consecutive posts accumulate a depressed relevance score at the account level, not just the post level. Common offenders: "strategy," "impact," "journey," "ecosystem." The fourth post in that streak takes a distribution penalty even if its own text would pass the classifier on a fresh account. Single-post audits miss this entirely because they do not examine posting history.

High-cadence AI-assisted posting without sufficient editing depth compounds the problem through a separate mechanism. Followers begin scrolling past at increasing speed after the second or third low-dwell post from the same creator. That behavioral pattern updates the classifier's predicted P(Skip) score for that creator-follower pair. The result is that the fifth post in a high-cadence streak fights a worse prior than the first, even when its content has nominally improved.

The instinctive fix is to reduce posting frequency. That alone is not sufficient. Reducing frequency without improving specificity does not recover reach. The P(Skip) score recovers when followers encounter posts worth reading, not simply when they encounter fewer posts. Both levers must move together.

Content creation growing 14% year-over-year means the volume of content your audience is filtering has grown at the same rate. The distribution gate tightens as supply increases. Accounts coasting on volume face a compounding disadvantage as the signal-to-noise threshold rises.

Dwell Time, the Golden Hour, and Stage 2 Distribution

LinkedIn's 2024 engineering blog introduced the Auto Normalized Long Dwell Model: a binary classifier that predicts whether a viewer's dwell time on a post will exceed a context-dependent percentile threshold, normalized by content type and creator type. A short text post and a long article are not held to the same absolute dwell-time standard. The model adjusts the threshold based on what is typical for that format and for that creator's audience.

LinkedIn A/B tests on the dwell model confirmed two outcomes: skipped updates decreased and time-on-feed increased across the sample population. Both are direct optimization targets for LinkedIn's feed team. A post that drives dwell time contributes to platform health metrics; one that drives skips costs them.

In the first 60 to 90 minutes after a post goes live, LinkedIn samples a subset of first-degree connections and measures their response. If engagement rate and dwell time in that window clear the threshold, the algorithm expands distribution to second-degree connections and hashtag followers. A post that underperforms in this window rarely recovers, even if later viewers engage more deeply. The initial sample window is the primary distribution gate, not a soft signal.

The account-level dwell-time dynamic feeds into this gate in a way that compounds over time. We have observed that followers who have scrolled past multiple low-dwell posts from the same creator respond at lower dwell rates even when the creator publishes improved content. The golden-hour sample for a creator with depleted dwell-time credit starts from a weaker prior. Recovery takes multiple posts with sustained specificity improvement. A single well-written post after a streak of low-signal content does not reset the score.

What the Banned-Word Approach Gets Wrong About LinkedIn's Originality Detection

Most AI content tools and most advice columns offer the same fix: here is a list of words that sound like AI. Remove them. Post cleared. This solves the wrong problem.

LinkedIn's originality classifier scores structural patterns: cadence uniformity, specificity density, and personal marker quality. A post that removes every flagged word but retains uniform sentence length, dense transition phrases, and no specificity anchors will still land in the suppressed tier. The vocabulary is surface. The structure is what the classifier reads.

Trust Insights practitioner analysis found that posts with concrete specifics, including company names, exact metrics, and specific timeframes, receive 3 to 4 times the reach of generic content on the same topics under the 360Brew ranking model. Niche-specific insights consistently outperform broad motivational content. The gap is not between human and AI authorship. It is between specific and vague content, regardless of who or what wrote it.

This reframe changes what editing means. Removing "synergy" and replacing it with "alignment" does not change the specificity score. Adding the name of the organization where you observed the outcome, the quarter it happened, and the metric that moved does change the specificity score. These are different operations. Most content review workflows only include the vocabulary pass.

Fix the Structure, Not Just the Words: A Specificity Checklist by Post Length

The editing depth required to clear the distribution floor is not uniform across post length. Short posts under 600 characters need one high-specificity anchor to shift out of the low-quality tier. One anchor is sufficient because the classifier normalizes the specificity signal against content length. A single named metric, date, or company name in a brief post moves the score meaningfully.

Posts over 1,200 characters require at least three specificity anchors distributed across the post body. A strong opener followed by two generic supporting paragraphs does not pass the length-normalized threshold. This is the systematic error in flat editing checklists: they apply the same standard regardless of post length, so long posts are routinely under-edited and short posts get over-labored.

Specificity anchors that score well: a named organization, an exact percentage or dollar figure, a specific date or timeframe, a named product or event, a role-specific outcome stated with precision. "We grew revenue" is not an anchor. "We added $2.3M in ARR in Q3 2024" is. The anchor does not need to be dramatic. It needs to be grounded in something a reader could, in principle, verify.

After editing for structure, a quick audit catches the common gaps. For a long post: does each major section contain at least one concrete referent? Does any first-person framing reference a specific professional context rather than generic reflection? Is sentence length clearly varied between short and long sentences in the same paragraph? For a short post: is there one dateable or named claim anywhere in the body? When these checks pass, the classifier's structural signals improve regardless of which tool produced the original draft.

Frequently asked questions