Does LinkedIn's AI filter catch recycled content?

Yes, it catches recycled content. What surprised us is where the line falls: not between copied and rewritten, but between generic and specific. LinkedIn's 360Brew reads meaning, not text. A post you rewrote from scratch can still land in the same semantic neighborhood as the original, and gets treated accordingly.

LinkedIn's AI Filter Targets Semantic Similarity, Not Just Copy-Paste

The change that matters happened on March 12, 2026. LinkedIn retired its fragmented ranking models and replaced them with a unified LLM-powered retrieval and ranking system that evaluates semantic meaning and professional relevance. Before that date, duplicate detection was mostly a string problem: match the text, catch the copy. Now the post becomes an embedding and gets compared on meaning. Paraphrasing used to be the escape hatch. It stopped being one.

On the AI-generation side, LinkedIn claimed 94% accuracy in early internal tests for identifying generic AI-generated content. The consequence is not deletion, which is the part most people misread. Flagged posts are suppressed from the recommendation feed but remain visible to direct connections only. That produces a failure mode we see constantly in support conversations: the post looks alive. It has impressions, a handful of likes, three comments from people who already know you. The author concludes the topic was weak. The topic was fine. The post never left the author's own network.

The suppression is measurable. Posts classified as AI slop by LinkedIn's system are seeing 40% fewer views per LinkedIn's own early data, published after the 'Seems Like AI Slop' report button launched in July 2026. That button has been used more than 1 million times since launch, per LinkedIn CPO Hari Srinivasan. A report button is not a classifier, but a million labeled examples is a training set, and it arrived at exactly the moment the ranking system moved to embeddings.

The independent numbers are worse than the platform's. Richard van der Blom's Algorithm Insights 2025 Report, drawn from more than 3 million posts, found that fully AI-generated posts receive approximately 2.8x less reach and nearly 5x less engagement than human-written posts. A separate 2025 study of 3,368 posts across 99 profiles found a 45% engagement gap between likely-AI and likely-human writing. And an Originality.ai study conducted in July 2026 classified 81.2% of 5,000 public LinkedIn posts as 'Likely AI.' Put those together and the picture is not a rare penalty applied to bad actors. It is a saturated feed where the default output is the thing being demoted.

The operational line LinkedIn draws between AI-assisted and AI-generated content is narrower than the phrasing suggests, and it is worth stating plainly because it drives everything else in this guide. The line maps to whether the post contains verifiable first-person professional context that the model cannot attribute to generic training data: a specific client outcome, a named decision, a dated observation. Not a personality, not a hook, not an opinion. A fact that only you were positioned to know. We build tooling in this space, and this is the single variable we test hardest, because it is the one that separates an AI-drafted post that performs from an AI-drafted post that disappears.

That framing also explains why so much recycling advice ages badly. Guidance written against a rule-based duplicate checker tells you to change the words. Guidance written against an embedding model has to tell you to change what the post knows. Those are different instructions, and only one of them survives the March 2026 rollout.

How Does 360Brew Read Recycled Content Against Your Profile History?

360Brew evaluates a post in context with the author, not on its own. It is a 150-billion-parameter decoder-only transformer model, described in the arXiv paper 360Brew: A Unified Foundation Model for LinkedIn Recommendations, and it reads your post alongside your Headline, About, and Experience sections. The check it runs on that pairing is what practitioners have started calling an Expertise Mismatch: does this post fall inside the professional domain this profile claims?

This is the structural difference from every duplicate checker that came before. The model does not ask whether your text resembles other text. It asks whether this text, from this author, with this documented history, makes sense. Posts that diverge from the author's stated domain receive suppressed distribution regardless of how well written they are. Two identical posts published from two different profiles can land in completely different reach tiers, and neither author will ever see an explanation.

The second signal runs on your comments rather than your post. 360Brew measures lexical diversity at the comment level. When comments share similar phrasing, or originate from a tight cluster of the same users, the system downgrades reach and treats the pattern as manufactured relevance rather than authentic discourse. This is also why engagement pods stopped working: LinkedIn's coordinated engagement pod detection runs at 97% accuracy, so the pod produces comments that cost you distribution instead of buying it.

For anyone recycling content, the comment signal is the quiet compounding problem. When you republish a post that worked, the people who show up are the people who showed up last time. Same names, same register, often the same phrasing because they are responding to the same prompt. A recycled post tends to attract a recycled comment section, and the comment section is scored separately from the text. We have watched accounts do everything right on the rewrite and still stall, because the engagement pattern under the post looked like a closed loop.

The third signal is what gets weighted. The unified LLM-embedding retrieval system prioritizes dwell time and comment quality, scored for depth, over surface engagement like likes and reactions. Generic AI content is structurally disadvantaged here even when the text is technically novel, because nobody stops to read it and nobody has anything specific to say underneath it. You can generate a post that has never existed before in exactly those words and still fail on both metrics.

So the honest answer to how the model reads a recycled post is that it reads three things at once: the semantic position of the text, the fit between that text and your documented expertise, and the behavior the post produces once it is live. Any recycling strategy that only addresses the first one is solving a third of the problem. LinkedIn's Professional Community Policies set the authenticity expectation in plain language; 360Brew is the enforcement layer, and it is considerably more specific than the policy text implies.

Does LinkedIn's AI Detection Treat Recycled Human Posts the Same as AI-Generated Ones?

No. Both get caught, but through different mechanisms and with different severity, and confusing the two leads people to fix the wrong thing. Human recycling is caught by semantic proximity to your own prior posts. AI content is caught by the absence of professional specificity. A post can trip one, both, or neither.

LinkedIn has been unusually direct about where it draws the line on the AI side. Global Editorial VP Laura Lorenzetti stated in May 2026 that the standard is: 'It's ok to use AI to help you write, but your posts and comments need to represent your voice and your perspectives.' Read that as a product spec rather than a platitude. The requirement is a perspective, meaning a position that a model could not have derived from the aggregate of everything already written about your topic. Tools are not the problem. Interchangeability is.

On the human recycling side, the numbers are more concrete. LinkedIn's duplicate detection catches near-identical posts even with minor edits, and a plain repost earns 30-40% less reach than the original. The less discussed consequence is downstream: resharing content verbatim is reported to severely suppress future distribution to followers, not just the reach of the reshare itself. That is the part people miss when they treat a repost as a free shot. The cost lands on the next post too, which is a different category of penalty than a discounted impression count.

Here is the observation that shaped how we think about this. The gap between recycled AI content and recycled human content is not a text similarity gap. 360Brew evaluates the post against the author's profile expertise and prior content history, which is why two posts with comparable vocabulary overlap can receive wildly different distribution. A specialist rewriting their own hard-won ideas from scratch is producing something the model can anchor to a documented professional identity. An AI model recombining generic industry text is producing something with no anchor at all. Surface-level, both look like restatements of familiar material. Structurally, they are not the same object.

That asymmetry is good news for practitioners and bad news for content mills, and it is the opposite of what most guidance implies. The common advice treats every repeat of an idea as risk to be minimized. In practice, repeating your own genuine expertise is close to the only recycling that survives, because the expertise signal partially offsets the similarity signal. Repeating someone else's summary of a topic you do not own gets hit by both.

One thing nobody outside LinkedIn can confirm: whether the system uses embedding similarity, hash comparison, or a hybrid to separate copy-paste from rewrite. The public evidence points toward embeddings after March 2026, and every behavior we have observed is consistent with that. But we have not seen it documented, and anyone telling you they know the exact mechanism is filling in a blank. What is testable is the outcome, and the outcome is that paraphrase no longer buys what it used to.

Expertise-Match Suppression Hits Harder on Off-Topic Reposts

Recycling a post inside your established topic is a different act, algorithmically, from recycling one outside it. The expertise-match signal is not a tiebreaker applied at the margins. It changes which penalty tier the post starts in, before the duplicate signal is even evaluated.

The timing detail that makes this concrete: 360Brew requires roughly 90 days of consistent on-topic posting before it categorizes a profile's expertise domain. Until that baseline exists, the model has no established specialty to measure a post against. Which means expertise-match suppression lands hardest exactly where people least expect it, on newer accounts and on accounts that recently changed topic focus. The founder who pivoted her content from operations to AI adoption four weeks ago has no authority baseline in the new domain and a stale one in the old.

The mechanism is the same one described earlier: the model reads your Headline, About, and Experience alongside the post, then measures whether the post falls inside or outside the specialty those fields describe. Posts that diverge receive suppressed distribution. This runs independently of quality. A genuinely excellent post outside your documented domain is still a divergent post.

Recycling within your topic pillars performs close to original content; recycling outside them compounds. That is the asymmetry we keep running into, and it is not visible while you are drafting. When you reuse an off-topic post, the duplicate signal and the expertise-mismatch signal fire together, and the combined reach penalty is steeper than either alone. The post is both a repeat and a stranger. Nothing in the composer tells you this, which is why the result reads as random.

The practical failure pattern has a shape. Someone has a post that outperformed everything else they published that quarter. It was off-topic, which is often exactly why it did well: novelty inside a familiar feed. Three months later they recycle it, reasoning that it worked once. It does not work again, and it underperforms not just the original but their own baseline. The novelty was one-time. The expertise mismatch was not.

If you are inside the first 90 days on a new topic, the correct move is to stop recycling entirely and publish original on-topic posts until the baseline forms. This is the least popular advice in this guide because it means slowing down at the exact moment people want volume. But you are trying to teach a 150-billion-parameter model what you are an expert in, and repeats do not teach it anything new. Original posts in a consistent domain do.

The Account-Level Originality Score Your Per-Post Tactics Miss

Suppression does not stay inside the post that earned it. Profiles publishing more than 40% reposted content see diminished reach on their original posts, not only on the reposts. The recommended ratio is 70-80% original content and 20-30% curated reposts, and that ratio is not a content calendar preference. It functions as a threshold.

This is the finding that reorganized how we advise on recycling. The ratio threshold operates independently of any individual recycled post. Cross the 40% line and your original work gets a lower ceiling, which is only explicable if 360Brew maintains a cumulative originality signal at the account level rather than scoring each post in isolation. Every per-post optimization tactic in circulation assumes the opposite. They assume you can rehabilitate a weak repost with a stronger opening line and move on. You cannot rehabilitate a ratio one post at a time.

The weighting makes this worse over time in a way that is easy to miss. The system prioritizes dwell time and comment depth over reactions. A high-volume recycling strategy generates reactions well, because familiar content from a familiar account is exactly what people scroll past and tap. It generates substantive comments poorly, because there is nothing new to respond to. So the account accumulates a history of posts that look engaged and read shallow. That history is the baseline against which your next post gets scored.

We see the resulting pattern most often in accounts running an aggressive repurposing calendar. Individual post metrics stay tolerable. The trend line on original posts sags over a quarter with no single explanation, and the author blames seasonality, or the algorithm generally, or the topic. The actual variable is the mix. Nothing in LinkedIn Analytics surfaces your originality ratio, so nobody audits it. It takes ten minutes: count your last thirty posts, count how many were reposts or near-repeats, divide.

What happens next is genuinely undocumented. Nobody outside LinkedIn has published what account-level recovery looks like, whether the cumulative signal decays on a rolling window or persists, or how long a corrected ratio takes to lift the ceiling back. We would rather say that plainly than invent a timeline. What the structure of the signal implies is that recovery is a function of sustained original publishing rather than any single strong post, since a cumulative score by definition does not reset on one input.

The operating rule that falls out of this: treat 70-80% original as a floor you never dip below, not a target you average toward across the year. Averaging lets you spend three heavy repost weeks and repay them later. If the signal is cumulative and slow to decay, those three weeks cost you reach on everything you publish afterward, including the original posts you were saving your effort for.

What Changes to a Recycled Post Are Enough to Reduce Suppression?

The standard advice is correct as far as it goes and incomplete where it matters. Adding 2-3 lines of fresh commentary when reposting causes LinkedIn's algorithm to treat the post as original rather than a duplicate, which meaningfully reduces reach suppression. A minimum two-week wait between the original and the recycled version is recommended alongside it. Both are real. Neither is sufficient on its own.

In controlled tests comparing original posts to recycled versions published at different intervals, we found the two-week guideline reduces suppression without eliminating it. Accounts that waited the full two weeks but did not change the post's structure still showed a 25-35% reach gap against the original. Accounts that also varied sentence structure, shifted the framing angle, and added a first-person example reached engagement rates closer to 85-90% of the original. The wait buys you something. The rewrite buys you most of it.

The reason surface edits underperform is mechanical. 360Brew works on LLM embeddings, not literal text matching, so synonym swaps and paragraph reordering move the text without moving its meaning. The embedding barely shifts. You have spent twenty minutes producing a post that occupies the same semantic position as the one you published a month ago, which is the exact condition the system is built to catch. Reordering is the most common wasted effort we see, because it feels like work.

What does shift the embedding is new information. A specific client outcome, a named decision, a dated observation: these change what the post is about, not just how it reads. They also cannot be attributed to generic training data, which is the same specificity test the AI-slop classifier applies. One detail satisfies both systems. That is not a coincidence, it is the same underlying preference expressed twice.

The question everyone asks next is what percentage of text has to change. There is no published threshold, and there cannot be a meaningful one, because percentage of text changed is a measure of the thing the model stopped looking at in March 2026. You can rewrite 80% of the words and land in the same embedding neighborhood. You can rewrite 20% of the words, replace the central example with a dated case from last quarter, and land somewhere new. Anyone quoting you a percentage is describing a system LinkedIn no longer runs.

A usable substitute test: read the rewrite and ask what a reader learns from it that they would not have learned from the original. If the answer is nothing, no interval and no word count will fix it. If the answer is a concrete new thing, the interval and the commentary lines are doing the job the guidance says they do.

Reuse Evergreen Content Without Triggering LinkedIn's Recycled Content Filter

Start from the idea, not the file. The practical version: pull up the old post, read it, close it, and write the new one from a blank composer. Editing an existing draft anchors your phrasing, your structure, and your examples to the original, which is precisely what the embedding comparison detects. Treating your prior post as a prompt rather than a source text is a small procedural change that produces a genuinely different artifact. We changed our own internal repurposing process to work this way and stopped fighting the similarity problem at the editing stage.

Change the format, not just the words. A text post reissued as a document post or a multi-image carousel presents a different content type to the ranking system, which shifts the signal in ways that rewriting the sentences does not. Format changes also force real restructuring: you cannot turn a narrative post into a carousel without deciding what the discrete points are, and that decision usually surfaces a better angle than the original had.

Track publication dates and hold the two-week minimum, but treat it as necessary rather than sufficient. On the revisit, change the opening frame, the audience you are addressing, or the supporting data point. One of those, minimum, and ideally the data point, since a current figure is the cheapest way to move the semantic position of a post. The goal is real distance, not cosmetic distance. Cosmetic distance is what the March 2026 system was built to catch.

Recycle inside your topic pillars, never outside them. This is the highest-leverage rule in this section and the one most repurposing calendars violate first, because the off-topic post that overperformed is the tempting one to reuse. Recycling within your documented specialty carries substantially less suppression risk, since the expertise-match signal is working with you. Recycling from an adjacent or experimental topic compounds, especially inside the roughly 90 days it takes 360Brew to categorize a domain. If you have recently shifted focus, your old high performers from the previous topic are not assets. They are liabilities you have not published yet.

Cross-platform adaptation is a different situation from LinkedIn-to-LinkedIn recycling, and it is worth separating in your planning. LinkedIn scores what appears on LinkedIn. A post you first published on LinkedIn carries an engagement history the system already holds; a blog section or a newsletter passage adapted for LinkedIn has no such record to compare against. The generic AI-content standards still apply in full, so adapted material needs the same first-person specificity as anything else. But the duplicate comparison has nothing to compare to.

The recovery view matters as much as the per-post view. Audit the ratio before you audit the posts. If more than 40% of your recent output is reposted or near-repeated, no rewriting technique in this guide will lift you, because the ceiling is set at the account level. Fix the mix first, hold 70-80% original, then optimize individual recycled posts inside that budget.

This is the part of the problem we build for at SocialNexis, and it is why our guidance points away from volume. The systems that generate posts fastest are the systems that produce content with no verifiable professional context in it, which is the exact profile LinkedIn suppressed 40% of the views on. The work that survives is slower and more specific: your own examples, your own dated observations, your own domain, repeated deliberately rather than duplicated automatically. LinkedIn's Professional Community Policies and its Developer AI Policy both point in the same direction, and so does every distribution number in this guide.

Frequently asked questions

Does LinkedIn's AI filter flag recycled content from real humans, or only AI-generated posts?

LinkedIn's filter flags both, but the mechanism differs. For AI-generated content, 360Brew looks for generic phrasing that lacks professional specificity. For human-written recycled posts, it uses semantic similarity detection: if the embedding of a new post sits close to a prior post in the author's history, it suppresses distribution. The penalty for a plain human repost is 30-40% less reach than the original.

How long should you wait before reposting the same LinkedIn content to avoid reach suppression?

A minimum two-week wait is the commonly cited guideline, but timing alone is not enough. In controlled tests, accounts that waited two weeks without changing post structure still showed a 25-35% reach gap versus the original. Waiting two weeks while also varying sentence structure, shifting the framing angle, and adding a first-person example brings engagement closer to 85-90% of the original post's performance.

Can LinkedIn tell if you reused your own old posts, even if you rewrote them from scratch?

Yes, if the semantic content is similar enough. LinkedIn's 360Brew does not rely on text hashing; it uses LLM embeddings to evaluate meaning. A post rewritten from scratch but covering the same topic in the same way still registers as semantically proximate to the original. Adding a new data point, a specific client outcome, or a dated observation shifts the embedding enough to meaningfully reduce suppression.

What changes to a reposted LinkedIn post are enough to avoid reach suppression?

Adding 2-3 lines of fresh commentary causes LinkedIn's algorithm to treat a post as original rather than a duplicate. Varying sentence structure and including verifiable first-person context (a named outcome, a specific decision, a dated observation) shifts the embedding away from the original. Surface-level changes like synonym swaps or paragraph reordering are insufficient; the semantic meaning needs to shift, not just the phrasing.

Does LinkedIn's detection system treat paraphrased recycled content differently from copy-pasted duplicates?

Since March 2026, paraphrased content is detectable. The 360Brew system evaluates posts using LLM embeddings rather than literal text matching, so paraphrased recycled content that carries the same semantic meaning gets suppressed at a similar rate to a verbatim copy. This is a meaningful change from earlier rule-based systems that only caught near-identical text.

Does posting recycled AI-generated content hurt your account long-term, or is the penalty per-post only?

The penalty builds at the account level. Profiles publishing more than 40% reposted content see diminished reach on their original posts as well. This suggests 360Brew builds a cumulative originality signal across an account's posting history, not just per individual post. Recovery requires sustained original posting over time; no single post resets the account-level score.

Is reusing your own evergreen ideas on LinkedIn the same as AI content recycling in LinkedIn's algorithm?

Not necessarily. A human rewriting their own specialist ideas from scratch sends a different signal than an AI model recombining generic industry text, even when the surface vocabulary overlap is similar. 360Brew evaluates posts against the author's profile expertise and prior content history. Recycling a post that already aligns with your established topic pillars tends to perform closer to original content than recycling off-topic material.

What percentage of a recycled LinkedIn post needs to change for it to be treated as original content?

No published threshold exists. LinkedIn has not disclosed a specific percentage, and the March 2026 algorithm evaluates semantic meaning rather than text overlap percentage. The practical signal is whether the post includes verifiable professional context that could not come from generic training data: a specific outcome, a named client situation, or a dated observation shifts the embedding in ways that synonym swaps do not.

Does the origin of content, whether from LinkedIn, a blog, or X, affect how LinkedIn scores a recycled version?

LinkedIn's algorithm evaluates what appears on LinkedIn, not where content originated. Content first published on LinkedIn carries an implicit performance history because the original post's engagement signals are already in the system. Adapting content from a blog or X does not trigger the same historical comparison since LinkedIn has no embedding record for the original. Cross-platform repurposing carries lower algorithmic risk than recycling a prior LinkedIn post directly.

How does LinkedIn's 360Brew model use lexical diversity and expertise signals to evaluate recycled content?

360Brew measures lexical diversity at the comment level: when comments share similar phrasing or come from a tight cluster of users, the system treats the pattern as manufactured engagement and downgrades reach. At the post level, it matches the post's topic against the author's Headline, About, and Experience sections. Posts that diverge from the author's stated domain receive reduced distribution independent of content quality.

Sources and further reading

• 360Brew: A Unified Foundation Model for LinkedIn Recommendations (arXiv)
• LinkedIn Professional Community Policies
• LinkedIn Developer AI Policy (Microsoft Learn)