An 8,795-post study, LinkedIn's 360Brew algorithm, and the creators recovering reach in 2026.
Over half of long-form LinkedIn posts are now AI-generated. They get 45% less engagement than human-written content. The reason isn't taste — it's a 150-billion-parameter algorithm called 360Brew specifically designed to detect generic content. Two cohorts are winning: people not using AI at all, and people who train AI on their own voice first.
A designer with 20,000 LinkedIn followers told me last year that he'd been averaging over a million impressions every couple of weeks. Then he started feeding his posts through ChatGPT to save time. Within a few months his reach collapsed. He stopped using AI, went back to writing himself, and the million-impression cadence came back.
This pattern isn't isolated. It's measured.
In late 2024, Originality.ai analysed 8,795 LinkedIn long-form posts published between January 2018 and October 2024. They classified each post as likely-AI or likely-human using their detection model, then looked at engagement.
The combined effect: LinkedIn is flooded with longer, more polished, AI-generated content that performs measurably worse. Algorithm analysts at multiple firms have reported organic reach dropping roughly 50% year-on-year through 2025-2026. The platform's solution isn't to suppress AI as a category — it's to suppress what most AI tends to produce.
In January 2025, a paper titled "360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation" appeared on arXiv. The lead author was Hamed Firooz, head of LinkedIn's Foundation AI Technologies team. The paper described a 150-billion-parameter language model — derived from Meta's LLaMA 3 architecture, fine-tuned on LinkedIn's proprietary data — that ranks every piece of content on the platform.
360Brew has been deployed gradually through 2025 and is now the dominant ranking signal. Unlike older recommendation systems that scored content on engagement velocity and follower graph alone, 360Brew reads the actual text of every post and judges it on semantic quality. It's specifically built to identify templated structures, generic vocabulary, the predictable rhythm of LLM output, and the absence of personal anecdotes or specific examples.
In March 2026, what's been called the "Authenticity Update" added natural language processing classifiers on top of 360Brew. Reporting from independent algorithm analysts suggests AI-detected posts now see roughly 30% less reach and 55% less engagement under the new system.
LinkedIn's stated position is that they don't downrank AI-generated content as a category. That's technically true. What they downrank is generic content — content that pattern-matches to the average of its genre.
The catch: most unedited AI output is genetically generic. Large language models are trained on the average of all writing on the internet, then optimised to produce safe, useful, on-topic responses. The natural output is the median of the genre. That median is exactly what 360Brew is built to suppress.
So the practical effect is identical to suppressing AI content. The mechanism is just more honest: it's not about whether you used a robot, it's about whether what you produced sounds like everyone else.
The specific signals 360Brew uses, based on the paper and subsequent reporting:
A cybersecurity professional with 14,000+ followers documented her experience in detail. For six weeks she fed her weekly newsletter into ChatGPT and posted the output 3-4 times a week. Initial engagement looked fine — likes ticked up 15%. Then comments dropped 62% across the six-week stretch. A post about zero-trust architecture, on a topic where she has genuine expertise, received 117 impressions against her 14,000 followers.
She paused AI use for 30 days and wrote 8 posts by hand, grounded in real client work. Average impressions jumped to 2,100 within 72 hours. Comment depth recovered. About a third of the comments asked follow-up questions, which is the engagement signal 360Brew weights heaviest.
The platform-level pattern matches:
Buffer ran a seven-day experiment in 2024 where the CEO posted AI-generated content exclusively. Average impressions were 11% higher than a typical week. Average engagement was 75% higher.
Worth flagging. Three caveats: the experiment ran on a high-baseline account with established credibility (the algorithm rewards account history heavily). The sample was seven days. And it ran before the March 2026 Authenticity Update, which is where the steepest detection improvements came online. The Buffer result is real, but it's not the modal experience for someone with 1,000-50,000 followers running AI content in 2026.
The cohort growing reach right now is the one running a specific pattern: use AI to draft, then rewrite the draft in your own voice. The key is the rewrite — and more importantly, training the AI on your voice before the first draft, so the rewrite is small.
The mechanism is straightforward. AI defaults to "AI voice" because that's the average of its training data. Feed it 500-800 words of context about how you specifically write — sentence length range, words you'd never use, signature phrases, point-of-view, the rhythm of how you make a point — and the default shifts. Output starts pattern-matching to you instead of to the genre average.
This is what every creator outperforming the algorithm in 2026 is doing, whether they call it that or not. Some build it themselves over months of trial and error. Some hire someone to build it. Some pay a ghostwriter and skip the AI entirely. The principle is the same: distinctive voice beats polished content.
We've built voice systems for businesses across wellness, B2B outbound, financial planning, and photography. The mechanism is consistent — voice context fed before any AI task — but the calibration is different per operator. A nutritionist's voice prompt is built differently from a B2B sales founder's, which is built differently from a portrait photographer's. The output reads as them, not as the genre.
If your LinkedIn reach has dropped over the last 12-18 months and you've been using AI to draft posts, this is probably why. The fix isn't to abandon AI. The fix is to stop using it as a generic writer and start using it as your specific writer.
Three options, in order of cost and effort:
The middle option is what most people should do. The first one is what most people will end up doing by accident, because the second one feels like more work than it is.
Either way, the era of generic AI content on LinkedIn is over. The algorithm isn't punishing AI — it's punishing the average. The way out is to stop being average.