A Million People Clicked LinkedIn's Slop Button. By Monday It Meant Your Posts Get Less Reach.

LinkedIn turned on its "Seems like AI slop" button on July 30. Chief product officer Hari Srinivasan posted his first progress report on August 20, twenty-one days later.

Cybernews quoted him putting the launch two weeks earlier, and Fortune described the same window. The Register, working from the identical post, said three weeks.

On its own the discrepancy is trivial. What makes it worth a paragraph is where it seems to begin. If Srinivasan wrote two weeks, then the reporters who quoted him accurately printed an interval his own launch date contradicts, while the outlet that did the arithmetic quietly overruled the executive it was citing. Newsrooms make that trade constantly and usually nobody notices. It happened here inside a story about how hard it has become to tell human content from machine-made content.

What was actually disclosed

Everything published this week traces to one post, written on the platform it describes by the executive responsible for it. Srinivasan said more than a million people had clicked the button, and that members were seeing roughly 40% fewer views of what LinkedIn classifies as slop compared with a few weeks earlier. He allowed that the job was unfinished, writing that the company knows it has "more to do."

Nothing else came with it. There was no denominator, no stated baseline period, no breakdown by content type or geography. When Fortune had sought comment on the original July announcement, a LinkedIn spokesperson pointed the outlet back at Srinivasan's post.

Within six days the figure had run through Slashdot, Cybernews, BigGo, The Register, PCWorld and Fortune, and three details had changed along the way.

A million of what, exactly

Srinivasan's phrasing was that over a million people had clicked. LinkedIn's communications team later told Moneywise the number counts unique members who used the feedback flow rather than total clicks.

Reporters should treat that clarification as a point in the company's favor, since a million distinct people is a much harder number to manufacture than a million clicks, which a few hundred motivated users could run up in an afternoon. Companies tend to volunteer that kind of detail when the underlying figure holds.

PCWorld reported that the button had been used over a million times.

The reversal that matters to marketers

Srinivasan's 40% was a claim about what audiences see. Members are being served less of the material LinkedIn's classifiers tag as slop, and that is the whole of what he said.

PCWorld hedged it, reporting that views had fallen 40% in some cases, a qualifier that appears nowhere in the original.

The Register flipped it, reporting that people who copy-paste AI-written posts are seeing around 40 percent fewer views. That version describes author-side reach, which is a claim about what happens to your post rather than what turns up in your feed, and the two only connect if you assume a mechanism LinkedIn has never published.

Marketers should care which one is true. A statement about feed quality costs a content team nothing. A distribution penalty on AI-assisted posts is the kind of finding that sends a content lead into Monday standing with a proposal to stop drafting in ChatGPT, and the version with consequences attached is the one that traveled furthest.

Two things shipped in the same release

Even at face value, the 40% cannot be assigned to anything. LinkedIn launched the button and updated its detection classifiers together. Srinivasan has said that "no single piece of feedback determines how content is distributed," and that the company weighs many signals at once, but LinkedIn has not explained how the reports feed into reach beyond confirming that they do.

So a million people pressed a button, models were retrained on what they pressed, distribution moved by a reported 40%, and no outsider can say which of those did the work. LinkedIn has little reason to sort it out in public. A crowdsourced cleanup makes for better copy than a classifier update, and the button is the half of the operation members can actually see.

A second change got far less coverage and carries more weight for anyone posting on behalf of a brand. Flagged posts now lose reach the way content marked "not interested" does, and the author gets a private notice in their analytics dashboard suggesting the post read as inauthentic. Srinivasan presented this as feedback rather than penalty. For a company page or a ghostwritten executive program it functions as a mark only the wearer can see, applied by a classifier nobody outside the company can audit, against a definition LinkedIn concedes keeps moving.

What the platform is up against

The scale here explains why LinkedIn was willing to publish an unauditable number in the first place.

Srinivasan says the company blocks hundreds of thousands of automated comment attempts daily and has stopped billions of other automation attempts in recent months. At the same time LinkedIn is retiring "enhance your post," the AI writing tool it built for its own members, and replacing it with a proofreader meant to leave a writer's voice alone. Microsoft's professional network spent two years handing users a slop machine, and this month it handed them a button to report the output.

Everyone else is somewhere similar. Spotify says it pulled 75 million bulk uploads and duplicate tracks over twelve months against an estimated 100 million on the service. Substack has partnered with Pangram to flag AI-written newsletters. Digg shut down its Reddit competitor in March, blaming bot traffic it could not get under control. Cloudflare now reports that automated requests outnumber human ones across the web, a threshold the company had expected to take longer to reach. Fortune reached for the dead internet theory to describe all of this, which is glib but not far off.

The measurements from outside the building

Two firms have tried to quantify how much of LinkedIn is machine-written, and their headline numbers look impossible to reconcile.

Pangram scanned close to 57,000 items between April and June and found 41% of LinkedIn's long-form public posts and 30% of its public comments entirely AI-generated. Its cross-platform comparison is the more damaging finding. X came in at 29% and Reddit at 13%, and while LinkedIn accounted for only about a third of everything Pangram scanned, it produced 62% of all the AI content the firm flagged. Originality.ai reviewed 5,000 public July posts and reported that 81.2% showed more than moderate AI use.

Those two results are routinely presented as though one of them must be wrong. Neither is. Pangram counts only posts running past 250 words that come back fully machine-written, while Originality.ai looks at anything over 100 words and calls it likely AI once its model clears 50% confidence, with a 15% allowance built into the scoring. They set different thresholds because they are answering different questions. The 81% is the figure headed for a thousand agency decks, and it will not mean there what it means in the study.

LinkedIn disputed both sets of findings without publishing anything comparable of its own.

Nobody has defined the word

Srinivasan has been unusually candid that the category is unstable. "Slop is hard to define and the definition changes," he wrote when announcing the button, which is the stated reason it exists, since user reports are what tune the models. He has also been careful to separate using AI from producing slop, on the grounds that plenty of people run a model over a thought they already had. "AI slop is a top priority for all of us," he wrote.

Follow the chain and the problem gets clearer. LinkedIn is suppressing distribution on a category it admits it cannot pin down, calibrated by a million people who each brought their own private definition, benchmarked against outside detectors that are answering a narrower question about who or what did the typing. None of those instruments asks whether a post is accurate. A vendor can publish a capability claim its product cannot deliver, in prose a human wrote by hand, and nothing in this apparatus will flag it.

The failure mode running the other direction is already visible. TikTok labeled videos by Nikolai Savic, a creator with five million followers, as AI-generated. He had edited them himself, and he told the Wall Street Journal it damaged his credibility with his audience. Detection systems get it wrong in both directions, and for brands the costly error is the one that mistakes real work for output.

The version that stuck

What has settled into general circulation this week is that LinkedIn now penalizes AI-assisted posts by roughly 40%. Anyone could have come away from the coverage believing that, and the company never said it.

A single unaudited first-party metric picked up three different timeframes and two units of measure, then reversed direction, in less than a week, while passing through publications that check claims for a living. The claim it carried was about synthetic content.

Nobody fabricated anything at any point. The number simply had to be interesting and hard to check, which describes most of what LinkedIn is now trying to demote.