Two weeks ago LinkedIn added an option to the three-dot menu on every post and comment. Seems like AI slop. One click, any member, no explanation required.
Hari Srinivasan, LinkedIn’s Chief Product Officer, published the early numbers. More than a million clicks inside two weeks, and content the platform classifies as AI slop now receives about 40% fewer views than it did before the button existed. He framed it as progress, with good intent, multiple signals feeding the decision, and safeguards in place.
Richard van der Blom, who has spent years measuring how the LinkedIn algorithm behaves, was one of the few people in the field who didn’t cheer. His newsletter this week takes the update apart, and his objection is the right one: the platform handed a consequential judgment to a crowd with no training, no accountability, and no shared definition of the thing they’re judging.
He’s correct and I want to push the argument one step further, because there’s a defect underneath the ones already named, and it’s the reason this button will behave worse over time instead of better.
What makes the AI slop button different from other report buttons?
Think about what happens when you flag something on any platform. Spam, harassment, a fake profile, a scam DM. You click, the thing may or may not come down, and you get nothing out of it. Your position is unchanged. The only motive available to you is that the content shouldn’t be there.
Now consider this button in a feed where attention is fixed and supply isn’t.
Every post competes with every other post for the same finite pool of eyeballs. If a rival’s post gets 40% fewer views, that reach doesn’t evaporate. It goes somewhere, and some of it goes to the people still visible. Which means the person clicking the flag has a direct material interest in the outcome, and the size of that interest scales with how well the flagged post was performing.
That’s a new thing. LinkedIn has built the first moderation control where the reporter is paid, in the only currency the platform trades in, for a successful report.
You don’t need a conspiracy for that to matter. You need ordinary human behavior under an incentive. Nothing in any system is more reliable than that. A tool that rewards its own use gets used more, and it gets used for the reward whatever the stated purpose was.
Van der Blom noticed the symptom without naming the cause. In the comments on the Chief Product Officer’s own post, a member joked that he personally accounted for close to half the million clicks. The joke isn’t the story. The pride is. Somebody felt good enough about mass-flagging other people’s work to announce it in public, underneath the executive who built the feature.
Is LinkedIn’s AI slop button moderation or a dogpile?
I’ve written before about how a pile-on works, and the shape is always the same. First a label makes somebody acceptable to hurt, because once a person is an AI grifter or a slop merchant they stop being quite a person and the ordinary rules switch off. Then the group senses permission. Then people rush in, because there’s no cost to landing a blow and real reward for landing one.
Read that back with the new button in mind. The label is supplied by the platform. The permission is official. The cost is zero, since flagging is anonymous and requires no justification. And now, for the first time, the reward is measurable in reach.
The same machinery drives review bombing, which authors have been dealing with for years. A coordinated group decides a book shouldn’t do well, and a mechanism designed for honest signal becomes a weapon because the mechanism cannot tell motive from judgment. Nothing about LinkedIn’s version prevents that. A competitor with a network can bury a post, and neither the author nor the platform will ever know it happened.
Can both LinkedIn statements be true?
LinkedIn maintains that no single click determines distribution. LinkedIn also confirms that content classified as AI slop is getting 40% fewer views.
Both can be technically true and the combination still tells you nothing useful. Sure, one click may not sink a post. Ten might. Or the click may feed a classifier that then makes the call, in which case the click is doing the work at one remove and the disclaimer is a technicality.
What’s missing is the thing that would settle it. There’s no notification. You’re not told your post was flagged, by how many people, or whether the reduction applied. There’s no appeal, because there’s no accusation to answer. The penalty is real and invisible, which is the worst possible design: a punishment you cannot detect, contest, or learn from.
What is the button measuring?
Nobody has defined AI slop. Not the platform, not the people clicking. There’s no agreed line between AI-written, AI-assisted, and AI slop, and yet millions of members are now drawing that line in about two seconds, on instinct, with no training whatsoever.
So what fills the gap? Texture. How the writing feels. Even sentence lengths, tidy transitions, a certain evenness of tone.
That’s the exact error the publishing industry is making right now with detection tools, which I wrote about when book deals started collapsing over suspicion and a score. Texture is what professional editing produces. A well-structured post from somebody who took the trouble to organize their thinking reads more like a machine than a rushed one from somebody who didn’t.
Van der Blom points out who pays for this most. Non-native English speakers get hit hardest. Somebody with twenty years of real expertise who uses a tool to make that expertise legible in a second language gets flagged for the polish, not for the absence of substance. The button punishes the accommodation, not the emptiness.
It also kills the learning curve. A member experimenting with these tools for the first time produces something rough, gets flagged and buried, and never finds out why or how to improve. The correction arrives as a silent penalty instead of as feedback. That’s not correction at all.
Does LinkedIn sell the tools it punishes people for using?
Yes, and this is where the whole thing tips from poorly designed into indefensible.
LinkedIn will rewrite your About section with AI. LinkedIn will turn your comment into a post with AI. Parts of that sit inside paid subscriptions. The platform built the features, sold them, and then handed members a button to punish content that sounds like the features.
A company cannot ship the tool, charge for the tool, and then treat the tool’s output as a violation. Pick one.
What does the spam comparison tell you?
Van der Blom includes the detail that makes the priorities visible. Flagging genuine spam, a fake profile, a scam DM, takes five or six steps and days of waiting, and in his experience comes back with no violation found roughly nine times in ten. He reported a clearly fake account offering Sales Navigator seats at twenty dollars a month. LinkedIn found nothing wrong.
So the platform can ship a one-click control that cuts distribution by 40% inside a fortnight, while an obvious scam survives a multi-step report and a human review.
That’s not a resourcing problem. That’s a statement about what the company considers urgent, and the answer is the thing that makes the feed look bad, ahead of the thing that takes money from members.
What would a good version look like?
The problem the button is aimed at is real. There’s a flood of empty content on every platform, and I have put numbers to it. Being against this implementation isn’t being against the goal.
Van der Blom’s proposed alternatives are better than what shipped, and they share one property: each names something checkable instead of something felt. This contains no real experience or evidence. This is misleading, exaggerated, or questionable on the facts. This is repetitive engagement bait or recycled content.
A member reading any of those has to point at something in the post. That’s a higher bar than a vibe, and a higher bar is exactly what removes the value of a spite click.
I’d add three more, all aimed at the incentive problem instead of the definition problem.
Make the flag cost something. A required sentence, even one nobody reads, changes who bothers. Free actions attract volume and volume is what a bad-faith clicker has to offer.
Weight flags by the flagger’s history. Somebody whose flags are usually upheld should count more than somebody who flags everything. This is solved technology and every mature moderation system uses it.
Tell the author. A penalty applied in silence teaches nobody anything and cannot be appealed. If the goal is a better feed instead of a quieter one, the person who wrote the post needs to know.
What should you do about the AI slop button this week?
Van der Blom’s advice is to stop editing posts for tone and start stress-testing them for substance, and his three questions are good enough that I’ll pass them on instead of invent worse ones. Could you defend this to somebody’s face who wanted to tear it apart, backing every claim with a number, a name, or a specific moment? Is there a sentence in here that only you could have written, one requiring your exact experience and your exact failure? Would this survive somebody flagging it out of spite, holding up under real scrutiny?
The third is the one that does the work, and notice what it implies. If the accusation would feel almost fair, the problem was never the flag.
My own version of the test is shorter. Every piece carries something nobody could fake: a number, a name, a date, a client outcome, a specific afternoon. That rule is why a small consistent cadence beats a large one, since you cannot produce that many genuine specifics in a day and pretending otherwise is how people end up producing the thing the button was built for.
The button is badly built and it’ll get abused, because it pays the people abusing it. Both things can be true at once, and neither changes what protects you. A post rooted in a life nobody else lived doesn’t read as slop to any reasonable person looking. That’s the only defense available, and the one worth having anyway.