Who Is Counting the AI Slop?

On July 30, 2026, LinkedIn added an option to the menu on every post: Seems like AI slop. Not “appears AI generated”, not “low quality”. Slop, a piece of internet slang, now part of the interface of a platform owned by Microsoft. The same day, LinkedIn removed its own enhance your post button. That had used a language model to rewrite whatever members typed. It was replaced with a proofreader. Chief product officer Hari Srinivasan described the replacement as something that corrects your words without changing your voice. Both moves landed in a week when the numbers about AI content were everywhere. Which raises a question worth asking before accepting any of them: who’s holding the thermometer, and do they also sell air conditioning?

I’ve watched this platform fill up with machine-written posts, and I’m glad somebody finally put the word slop in the menu. I’m less glad about who’s doing the counting. I ghostwrite for a living, so I’ve got a stake here: every inflated number makes people more suspicious of anything polished, and every shaky detector puts a careful human writer at risk of being called a fraud.

What does LinkedIn’s AI slop button really do?

Less than the name suggests. Clicking it hides the post from your own feed and returns a message thanking you for the feedback. Beyond that, a LinkedIn spokesperson said flagged posts get reduced algorithmic reach, comparable to marking something as not interested.

The flags also train LinkedIn’s classifiers. There’s no removal, no public label, and no stated penalty on the account. One outlet noted plainly that what happens to a flagged post beyond reduced reach remains unresolved. That suggests the policy is still being worked out.

So it’s a personal mute button with a downrank attached and a training signal underneath. That’s something, and it’s also a long way from the enforcement mechanism the wording implies, and worth knowing before anyone panics about being reported.

The numbers do not agree with each other

Pangram, an AI detection company, scanned roughly a million posts through an opt-in Chrome extension and found more than 40 per cent of long-form LinkedIn posts fully AI generated, with X around 25 per cent.

Originality.ai, a competing detection firm, reported 53.7 per cent of long-form posts as likely AI in January, based on 3,368 posts from 99 influential profiles. In July, the same company scanned 5,000 posts and reported 81 per cent. Pangram’s browser extension users, meanwhile, see roughly two thirds of what they scroll flagged.

Four figures for one phenomenon, from four in ten to eight in ten. Every one of them has been reported as a fact. That spread should make any editor stop before running the number. Four in ten and eight in ten describe two different platforms, and a reporter who prints either one as settled fact has done the detection industry’s marketing for free.

They’re not directly comparable, and the firms say so. Different models, different sampling, and a real distinction between “fully AI generated” and “likely AI”. But the gap is the story. When two thermometers disagree by forty degrees, the useful question stops being what the temperature is.

Who benefits from a large number?

The studies come from companies whose business is detection.

Pangram raised nine million dollars in July, led by Menlo Ventures, in the same week its LinkedIn figures were circulating.

GPTZero has raised 13.5 million. Copyleaks around seven million. Originality.ai runs profitably selling to publishers and SEO teams. Substack launched a detector through a partnership with Pangram, and its chief executive said outright that he didn’t want Substack turning into LinkedIn.

The cause is more ordinary than a conspiracy and harder to correct for. The organizations producing the measurements are the ones that profit when the measurement is alarming, and the press cycle those measurements generate is what funds the next round.

The detection firms are also more careful in their own materials than the coverage of them is. Pangram’s chief executive calls his figures a lower bound, on the reasonable grounds that people who install an AI detection extension aren’t a random sample of internet users. That caveat rarely survives into the headline. That caveat matters, and dropping it is on the press. A lower bound from a self-selected sample is a sales figure with a decimal point, and it deserves the same skepticism anybody would give a vendor’s white paper.

Do AI detectors work?

Reasonably well, and the answer has a large asterisk on it.

Pangram claims a false positive rate of 0.01 per cent, and a critic writing against the detection industry still concedes its performance has held up across independent studies. This isn’t snake oil, and the first generation of detectors it improved on was.

The asterisk is where the researchers established the rate. Controlled benchmarks aren’t the informal, multilingual, stylistically mixed environment of a real social feed, and nobody has independently tested that model on non-native English at scale. Which matters, because the people the model is most likely to flag wrongly are exactly the people writing in a second language, or writing in a formal tone, or writing with the kind of consistency that good editing produces.

It makes me angry that the people most likely to get tagged as machines are the ones who worked hardest on their English, and they’re the least able to push back when a stranger’s click cuts their reach. A tool that punishes careful writing is broken, whatever its benchmark says.

The fine print is more revealing still. Originality.ai’s terms of service instruct users not to apply its output to any purpose carrying a legal or material impact on a person, employment decisions included. The company selling detection disclaims the use most buyers have in mind. It also states openly that it’s built for publishers and marketers, not students, and won’t defend its accuracy for academic use. The same probability score, sold as a business tool to one buyer and an accusation to another.

Why did LinkedIn add an AI slop button?

It would be tidy to conclude that a nine million dollar startup pushed Microsoft into action. It’s not what happened.

LinkedIn has better data than any external detector. It built the writing button, so it knows at the point of generation how much content came out of it. It doesn’t need a browser extension sampling a million posts to describe its own platform.

The timeline runs the other way, too. Disclosure labels for AI content date to 2024. The downranking initiative launched in May 2026, announced by editorial vice president Laura Lorenzetti, who wrote that overusing AI at scale dilutes what real conversation produces. The July studies arrived after the platform was already acting. The driver is more prosaic. A feed people distrust is a feed people scroll less, and less scrolling is less inventory to sell. That’s a revenue problem visible in LinkedIn’s own dashboards, and it doesn’t require anyone else’s research to notice.

Microsoft also has a position to protect. It’s selling AI assistance into every product it owns, so “AI is bad” isn’t a message it can afford. “Slop is bad, and AI used well is fine” is exactly the line Srinivasan drew when he wrote that AI and slop aren’t the same thing.

Why not stop AI slop at the compose box?

The announcement never addresses the obvious point. LinkedIn doesn’t need members to identify AI writing, because it’s better placed to identify it than anybody outside the company. It built the writing button. Every post that came out of enhance your post was generated inside LinkedIn’s own product. That means the platform knew at the moment of generation. It doesn’t need a browser extension sampling a million posts to describe its own feed.

And if a startup running a Chrome extension can flag content at scale, a company owned by Microsoft can do it in the compose box. If the goal were fewer generated posts, the intervention writes itself: a message before publication saying this reads as machine written, try again in your own words.

They’re not doing that. The flag arrives after publication, from another member, and removes nothing.

That tells me the company cares more about the complaint than the problem. A platform that can see exactly which posts came out of its own generator, and still waits for members to point at them, has picked the cheap fix on purpose.

The arithmetic behind that choice

Two reasons, and neither is about authenticity.

The first is inventory. Blocking a post at the compose box means one fewer post in the feed. Posts are what the advertising sits between, and a platform doesn’t reduce its own supply to solve a quality complaint. A hide button removes nothing. The post still exists, still reaches the poster’s network, and still carries ads for everyone who didn’t flag it.

The second is liability. Intervening before publication makes LinkedIn the accuser. Every false positive becomes an angry member with a legitimate grievance and no meaningful appeal, and the platform owns every one of those conversations. Handing members a button moves the accusation off the company and onto the membership. Same signal, no exposure.

The private analytics nudge confirms both readings. LinkedIn is willing to tell a poster their writing came across as inauthentic, privately, after the fact, with nothing attached to it. That’s the exact limit of what the company will do: enough to show concern and never enough to cost anything. Which makes the feature easier to describe than the branding suggests. It’s a hide button with a training signal attached. Members do the labeling for free, the classifiers improve, and the feed stays exactly as full as it was.

What part of AI slop has nobody solved?

A button that lets one member reduce another’s reach, anonymously, with no adjudication and nothing to appeal, is open to obvious misuse. A ghostwriter quoted in coverage of the launch put it plainly, warning the feature looks ripe for abuse against legitimate posts. The competitive version is the sharpest. In any field where practitioners compete for the same clients on the same platform, a free anonymous way to suppress a rival’s visibility won’t sit unused.

Consultants, coaches and freelancers all compete on that platform for the same attention, and I’d bet the first real use of this button in a crowded niche is a quiet shot at a competitor. LinkedIn handed out an anonymous weapon with no appeal attached. That’s careless, and the people who’ll pay for it are small operators who live or die on reach.

And human intuition about machine writing is worse than people believe.

What people flag is short sentences, clean structure, plain vocabulary and a consistent tone. Those are the surface features of generated text. They’re also what careful editing produces, and what anyone writing in a second language tends to produce. The same problem shows up wherever machines are asked to judge writing: the signal they measure is style, and style isn’t authorship.

LinkedIn is at least aware of the distinction. It’s testing a private note in a poster’s analytics when members have flagged their writing, framed as feedback about coming across as inauthentic. That’s a more sensible design than a public label, and it’s the first time writers have had any signal at all about how their work reads to strangers.

What to make of all this

Three things seem solid.

The problem is real. Anyone who has scrolled LinkedIn recently doesn’t need a study. The default output of a language model asked for a post is the same shape every time, and that shape now defines the platform.

The size of it is unknown. Four in ten and eight in ten are different worlds, and the people producing both figures sell the remedy. Treat any specific percentage as an unverified claim.

The remedy is aimed at the visible symptom. A button addresses what members complain about in public. It doesn’t address ad density, feed ranking, or the fact that the platform shipped the generator and is now shipping the alarm.

For anyone writing professionally, the practical consequence is smaller than the noise around it. Material that only one person could have written can’t be generated, can’t be mistaken for generation, and doesn’t depend on a detector to prove it. The difference between a machine and a person was always about who had been somewhere and could say what it was like.

The thermometer question stays useful. When the next number arrives, check who’s holding it before you repeat it. And look hard at a platform that sold its members a machine to write their posts and then handed them a button to accuse each other of using one. I think that’s cynical, and writers who spend their energy dodging detectors are letting the company off the hook for a mess it built.

Frequently Asked Questions