LinkedIn shipped a “Seems like AI slop” button on 30 July, tucked into the three-dot menu on every post. Chief product officer Hari Srinivasan posted an update this week: over a million people have clicked it, and posts flagged this way are now getting roughly 40% less reach than they were getting a few weeks earlier, according to LinkedIn’s own numbers as reported by Engadget and others. The button sits next to a quieter change — LinkedIn also killed “enhance your post,” the AI feature that used to rewrite your draft for you, and replaced it with a narrower “proofreader” that’s supposed to leave your actual words alone rather than generate new ones.
The trigger for all this is a study nobody at LinkedIn commissioned. Pangram Labs, an AI detection company, ran a browser extension that scanned roughly a million posts people actually scrolled past across LinkedIn, X, Reddit, Substack and Medium since April, and found that 41% of LinkedIn’s longform posts were fully AI-generated — reported first by 404 Media and picked up widely since. LinkedIn accounted for 62% of all AI-flagged content Pangram found, despite being only about a third of what the extension scanned. Pangram’s own researcher called that a floor, not a hard number, since people who install a detector extension aren’t a random sample of LinkedIn users. Fair caveat. It’s still a lot of posts about “5 lessons I learned from a failed product launch” written by a model that has never launched anything.
Why I’m writing about a corporate networking site
I don’t post much on LinkedIn, but I post some things there, the same as most people in games and VFX do — render breakdowns, a shot from a project once it’s cleared for public work, the odd “here’s what I shipped” update. That’s the part of this story that isn’t really about thought-leadership spam. It’s about what happens when a platform starts training a million users to hit “report” on anything that pattern-matches to synthetic, and the pattern it’s training them on is too polished, too consistent, too clean.
That description also fits a lot of finished 3D work. A well-lit hero render, a texture set with consistent PBR values across every surface, a lookdev pass where the materials all agree with each other under the same HDRI — that is what several hours of knowing what you are doing looks like. But it is, visually, close enough to the smoothed-over, slightly-too-perfect surface language that AI image tools default to that a viewer skimming a feed on their phone can mistake one for the other in under a second, and the “report” button is now one tap away.
I want to be precise about what I can and can’t back up here. LinkedIn hasn’t published category-level data on who’s getting flagged, and I haven’t found reporting that ties this specifically to 3D artists or game-dev portfolios — the closest documented version of this problem is on Steam, where indie developers have had to change capsule art after players mistook clean, competent illustration work for AI generation. That’s a different platform with a different mechanism, but it’s the same underlying failure: a detector, human or algorithmic, that was tuned on “AI slop” ends up flagging “unusually competent” as a false positive. LinkedIn’s button makes that failure mode faster and more visible than it’s ever been anywhere else, because it’s a single click with an immediate reach penalty attached, and there’s no appeals process described anywhere in the coverage I’ve read.
The actual defense is boring, and I already use it
The honest fix isn’t a watermark or a badge — those get faked or ignored. It’s the same thing that’s always separated a real production pipeline from a generated image: process that can be shown, not just a result that has to be taken on faith. A viewport recording with undo history intact. A wireframe pass before the smoothing. Layered PSDs instead of a flattened final. I’ve written before about recording my process at real speed instead of a sped-up timelapse, specifically because the boring, slow parts are the parts that prove something — a diffusion model doesn’t have twenty minutes of you adjusting a bevel and squinting at it. If platforms are going to start treating “looks too finished” as a signal worth punishing, the answer isn’t to make your work look rougher. It’s to keep the receipts that show how it got that finished, and post those alongside it.
None of this makes LinkedIn’s button a bad idea. A feed that’s 41% synthetic thought-leadership is a real problem and a report-and-suppress mechanism is a reasonable first attempt at fixing it. I’d just watch, over the next few months, whether LinkedIn ever publishes anything about false-positive rates — because a system with a mass-reporting button and no visible appeals process is exactly the kind of thing that quietly punishes the people doing the most careful work, right alongside the people doing none.