LinkedIn AI Slop Is 40% of Long Posts. The Engagement Data Splits by Topic

Pangram scanned 1,002,627 social posts and found that more than 40% of LinkedIn posts over 250 words were written entirely by a machine. That figure now circulates as proof that AI posting has stopped working. The engagement data says something more awkward. Likely-AI posts lose badly in analysis and marketing, and beat human posts in leadership and inspiration. Slop is not failing everywhere. It is failing in the half of the feed where your buyer reads.

What LinkedIn actually means by "AI slop"

The phrase arrived as a joke and is now a product category with a written definition. LinkedIn published its own on 20 May 2026, in a post by VP of Product Laura Lorenzetti.

Slop, in LinkedIn's telling, is content that reads as polished and carries no real experience, perspective or insight behind it. Three things are in scope: generic AI posts with no point of view, bot and template comments, and tools that publish at scale.

Notice what is absent from that list. The drafting tool is not the target. The missing person behind it is.

The test is originality, not tooling

This is the distinction most published guidance gets wrong, and it changes what a team should actually do. LinkedIn said plainly that AI-assisted content is still welcome when it contains original ideas or starts a meaningful conversation.

So "did a model touch this" is not the question being asked of your post. The question is closer to "could anybody else have published this".

A sentence that 400 competitors could have written is the failure state, whoever typed it. I would stop asking teams to declare which tool they drafted with, and start asking them what in the post cannot be copied. Those two audits produce very different content calendars.

How much of the feed is actually machine-written

Two studies landed within weeks of each other with wildly different headline numbers. The gap is entirely method, and reading them as a range is a mistake.

Pangram measured what people were served. It scanned 1,002,627 posts of 50 words or more, collected through an opt-in browser extension between 24 April and the end of June 2026, across LinkedIn, X, Reddit, Substack and Medium. Across all five platforms and all lengths, 13.8% was fully AI-generated. On LinkedIn longform it passed 40%.

Originality.AI measured what search returns. It pulled 5,000 public posts of 100 words or more from 90 topic-and-date searches in July 2026, and classified 81.2% as likely AI.

One number describes a feed. The other describes a search result set, which is a different population with a heavy tilt toward accounts optimising for discovery. Both are useful. Neither is "the percentage of LinkedIn that is AI".

Two studies, two very different populations

Every row above comes from a party with a commercial interest in the answer. Two sell AI detection. The third is grading its own homework and has not released a false positive rate.

The engagement penalty is real, and it is not evenly spread

The most repeated statistic in this debate is that AI posts get 45% less engagement. It traces back to a single Originality.AI study of 3,368 long-form posts from 99 influential profiles across 11 industries, running January to November 2025.

Six blog posts restating that number is still one source. So read the study's own category table, where the average stops being useful almost immediately.

Where AI posts lose, and lose heavily

The gap is widest wherever the reader arrived wanting information. In innovation and strategy, likely-AI posts averaged 143 interactions against 708 for likely-human ones. In marketing and branding it was 207 against 771. In healthcare, 116 against 208.

These are categories where a reader is trying to learn something specific and checkable. Generic text has nothing to give them, so they keep scrolling.

Where AI posts win, which nobody quotes

In leadership and inspiration, likely-AI posts averaged 2,635 interactions against 1,504 for human ones. That is 75% higher, the single largest category gap in the study, and it runs the opposite way to the headline.

Technology and finance also came out marginally ahead for AI, by roughly 7% each. The mechanism is not mysterious. Inspirational posts are consumed as a mood rather than as information, and fluent, familiar, emotionally legible prose is the thing language models produce best.

My read is that LinkedIn does not have a slop problem so much as a genre problem. Half its feed was always low-information content that performed on tone, and a machine does tone cheaply and well.

Interactions are likes plus comments averaged per post. Gap is computed directly from the two columns, so small rounding differences from the study's own labels are expected. Category baselines differ by two orders of magnitude, which is the first reason a single blended average across them is close to meaningless.

Engagement did not fall. Reach did.

Almost every write-up of this story assumes that saturation depressed engagement across LinkedIn. The largest public benchmark does not support that, and neither does LinkedIn's owner.

Socialinsider analysed 1.3 million posts across 16,645 LinkedIn business Pages. Average engagement rate by impressions came in at 5.20%, up 8% year on year. Text posts rose from 4.00% to 4.50%, the biggest gain of any format at plus 12%.

In the same dataset, views fell 36% year on year across every page size. Follower growth for pages between 100,000 and 1 million fell from 21.6% to 6.4%.

Engagement rate is a ratio: interactions divided by impressions. When the feed stops showing a post to the people who would have scrolled past it, that ratio improves while the commercial outcome gets worse.

This is the finding I would put in front of a founder before any slop statistic. Your engagement rate can climb every quarter while your actual reach halves, and a dashboard reporting the percentage will call that an improvement. The same measurement trap shows up in the piece on measuring discovery you cannot see.

The counter-evidence is worth stating just as plainly. Microsoft reported LinkedIn revenue up 12% for fiscal 2026, with double-digit member growth for a fifth consecutive year. A platform in engagement collapse does not file that.

What LinkedIn changed, and what it left alone

The May rule, which suppresses rather than removes

On 20 May 2026 LinkedIn said it would detect and downrank generic AI content. Flagged posts are not deleted. They stay visible to direct connections and followers, and the recommendation engine stops amplifying them beyond that circle.

The company claimed 94% accuracy in early tests. It has not published a false positive rate, and that omission matters more than the accuracy headline. Detection tools are documented as harsher on non-native English writers, so the cost of a false positive falls unevenly on real people with real expertise.

The July button, and what one click actually does

On 30 July 2026 LinkedIn added a "Seems like AI slop" option to the post menu. It shipped alongside new classifiers and a private dashboard signal for creators, and with the removal of its own "enhance your post" writing feature. The replacement proofreads without rewriting the author's voice.

Chief product officer Hari Srinivasan said on 20 August that more than a million people used the option in its first two weeks. He added that members were seeing 40% fewer views of content the platform classifies as slop.

One click does not demote anybody. It hides the post from the person who clicked and feeds the classifier a training signal. Distribution only moves when many members flag the same thing.

The detail I find most telling is the retirement of "enhance your post". LinkedIn spent two years shipping the feature that manufactured the problem, then withdrew it and asked members to report the output. That sequence is worth remembering the next time a platform offers to write for you.

What still breaks through

Strip out the advice that amounts to "be authentic" and two things survive contact with the data above.

One number that nobody else has

The categories where human posts win are exactly the ones where the reader wanted information. That points at a cheap and unglamorous answer: publish a figure your own business produced.

Sample size, time window, what you excluded. A number stated that way is checkable, which is why it is also quotable, and why a competitor cannot repost it as their own without attribution. The wider case for that is in the piece on original research as a content moat, and the verification discipline behind it in the note on case studies with verified outcomes.

Most B2B teams have three or four such numbers sitting in a CRM and have never published one. That is the whole opportunity, and it costs an afternoon rather than a budget.

A claim a reasonable person could argue with

LinkedIn's ranking model reads more than likes. In a 2026 paper describing Feed SR, its transformer-based feed ranker, LinkedIn engineers describe scoring on clicks, likes, comments and dwell time, using up to 1,000 recent impressions per member. The model lifted time spent by 2.10% in an online test.

Dwell time and comments both reward a post someone needs to think about. A post that states a position invites a reply. A post that agrees with everybody invites a scroll.

The same architecture rewards topical consistency, because a member's history is read as a sequence rather than as isolated events. The practical consequence is that posting on one subject for a year beats posting on eight, and that argument extends beyond LinkedIn into how community mentions drive AI visibility.

Where this argument is weakest

Three problems, in descending order of size.

First, every prevalence figure here comes from a company that sells AI detection. Pangram puts its false positive rate at roughly one in 10,000. Originality.AI claims 99% accuracy in recent testing. Neither has been independently audited, and detectors are known to misfire on non-native English writing. If both are 5% wrong in the same direction, several conclusions above move.

Second, the engagement study is correlational and its authors say so. It cannot separate "readers dislike machine text" from "accounts that automate also write about subjects with smaller audiences". Both explanations fit the same table.

Third, the category split rests on one study, one 11-month window, and 99 profiles. Government posts averaged 8,524 interactions while healthcare averaged 116. Populations that different should probably never be averaged, which is my objection to the 45% headline and also a caution against over-reading my own version of it.

The strongest case against the whole framing is simply LinkedIn's results. Revenue up 12%, members growing double digits for a fifth year, and a slop button pressed a million times in a fortnight without any reported effect on that trajectory. A platform can be full of low-quality content and commercially healthy at the same time. Reading the first fact as a prediction of the second has been wrong for a decade.

Three checks to run before you post

None of these require a tool, a detector or a policy document. They take about 90 seconds on a finished draft.

Then change what you report internally. Track impressions and comments separately, never engagement rate alone, because the rate rises when reach falls. That is the same denominator problem examined in the analysis of AI content saturation and search visibility, and the volume economics behind it in the piece on what the content flood actually costs.

Frequently asked questions

Where to start this week

Pick the single dullest number in your business that a customer would find interesting. Time to first response, win rate by lead source, average implementation length, anything you already record. Publish it once, with the sample and the period stated, and say what it excludes.

While you do that, pull your last 12 months of LinkedIn impressions and put them next to your engagement rate. If the rate rose while impressions fell, you have been reading a ratio and calling it growth. Fix the report before you fix the posting.

Related reading