LinkedIn, AI slop, and what changed
Four fifths of long-form posts now read as AI. LinkedIn shipped a button to report them, and a model that reads.
Two things happened to LinkedIn in 2026 and they are related. The feed filled up with generated text, and the ranking system was replaced by something that reads text rather than counting clicks. The second is the reason the first stopped working.
How much of LinkedIn is AI-written?
The scale first, because it is worse than it feels. According to Originality.AI, which sampled 5,000 public LinkedIn posts of at least a hundred words from July 2026 across nine topics, 81.2% classified as likely AI. In late 2024 the same method put roughly half of long-form posts in that bucket. Whatever the error bars on any single classification, the direction is not ambiguous: the long post has become the format most likely to be machine-written, and readers have learned that.
What has LinkedIn done about AI slop?
LinkedIn has responded in two ways, and only one of them gets written about. The visible one is a report option in the three-dot menu for posts and ads that seem generated, which the company has confirmed is rolling out. Hari Srinivasan, LinkedIn’s chief product officer, has said AI slop is a top priority. A human-reported slop signal is a meaningful thing to exist, because it converts a matter of taste into a number attached to your account.
What is 360Brew, and why does it matter?
The invisible change matters more. LinkedIn’s research describes 360Brew, a 150-billion-parameter decoder-only model built on a Mixtral 8x22 base and trained on LinkedIn’s own data, which handles more than thirty ranking and recommendation tasks without task-specific fine-tuning. It reads a member’s history as text in context rather than consuming hand-engineered features, and the paper reports that performance improves as more of that history is given to it and that it does better than the previous production system for members with fewer interactions. The paper was withdrawn from arXiv over licensing at submission rather than for any finding, which is worth stating if you are going to cite it.
The consequence is the part to internalise. A ranking system built on engineered features asks what happened to posts like this one. A model that reads asks what this post says and whether it is worth showing to this person. Under the first, a well-formed post about a common topic inherits the average performance of well-formed posts about common topics. Under the second, it has to be worth reading on its own terms, and the average is not a floor you inherit. That is why generated posts stopped working without anybody needing to build a detector: when four fifths of the feed is the same shape, being that shape is not a signal of quality, it is a signal of the average.
Is it still fine to write with a model?
This is not an argument against writing with a model, and the distinction is one LinkedIn has been careful about too. A post drafted with help that carries something only you know is not slop. A post that could have been written about any company by anybody is slop whether a model wrote it or not, and plenty of it predates the models. The test that survives both is whether the post contains a fact that could only come from you: a number from your own business, a thing that happened on a date, a decision you made and regret.
What changes about how you should post?
Two practical consequences follow. The first is that length has stopped being a proxy for effort, because generating length is free, so a short post with one real number now outperforms a long one with none. The second is that the report button changes the risk profile of publishing: for the first time on this platform, a reader who finds your post hollow can attach that judgement to it rather than just scrolling past.
What has not changed is what the feed is for. LinkedIn ranks for whether a specific member wants to read a specific thing, and it now has a model good enough to make that judgement from the text. The way to be read is unchanged and only harder to fake: say something that is true of you, and say it where the person it is true for can see it.