BlogDeep dive 1

LinkedIn Never Said It Detects AI 94% of the Time

What LinkedIn's 94%, 40% and million-flag figures count, and the one number it never published: how often good posts get caught.

5 October 2026 6 min read 15 sources cited
Contents
  1. The 94% comes with no false-positive rate
  2. The 40% is a platform total, graded by LinkedIn's own yardstick
  3. A million flags, and no single one decides your reach
  4. AI detectors and LinkedIn measure different things
  5. Judge your draft by its claim
  6. Sources

In May the press reported that LinkedIn catches AI slop with 94% accuracy. The Next Web called it LinkedIn's "94 per cent accuracy claim."30 LinkedIn never claimed accuracy. Laura Lorenzetti, its VP and Executive Editor, wrote: "In our initial testing, we're correctly identifying generic content 94% of the time."9 That counts generic posts in early tests, and it leaves out the number you need: how often a good post gets caught by mistake.

This is one of four deep dives behind my essay on LinkedIn's AI slop crackdown.

In short

  • LinkedIn's 94% counts generic posts caught in early tests. LinkedIn never called it AI detection and never published how often good posts get caught.
  • The 40% is a platform-wide drop in views of what LinkedIn classifies as slop, compared with a few weeks earlier. LinkedIn published no per-post penalty and did not tie the drop to the flag button.
  • Over a million people have clicked "Seems like AI slop." LinkedIn says no single piece of feedback decides how a post is distributed.
  • AI detectors and LinkedIn measure different things. A platitude typed by hand can pass a detector and fail LinkedIn, and a sharp ChatGPT draft can do the reverse.

The 94% comes with no false-positive rate

A filter that marked every post as generic would catch every generic post too. A catch rate alone says nothing about the good posts swept up with the bad, and LinkedIn has not said how many there are. The Next Web noted it "has not shared data on false positives."30 The Decoder was blunter: "How often legitimate posts get wrongly flagged as 'AI slop' is anyone's guess."43

A wrong call costs reach with strangers. Such content "is less likely to be widely distributed beyond a person's immediate network."9 For a founder whose buyers have never heard of them, strangers are the whole point of posting.

The 40% is a platform total, graded by LinkedIn's own yardstick

In August Hari Srinivasan, LinkedIn's Chief Product Officer, said he hesitated to share numbers because every feed is different, then gave one: "overall members are now experiencing 40% less views on what we classify as AI slop from just a few weeks ago."3

That is total views across the platform, and it tells you nothing about any single post. The words "what we classify as" carry weight too. The same system that defines slop is reporting its own success, and in those weeks LinkedIn was "ramping up a series of new and improved classifiers."10 Nobody outside LinkedIn can split the drop into less slop shown and a definition that moved.

Engadget ran the news as "LinkedIn says its AI slop button is working."37 LinkedIn announced the button and the new classifiers in the same late-July post,10 and the August post gives the 40% with no method and no cause.3 If someone tells you the flag cut slop by 40%, LinkedIn never said so.

A million flags, and no single one decides your reach

The same August post carried the third number: "Over a million people have now clicked 'seems like AI slop,'" in about two weeks.3 It does not say how many posts were flagged or what one flag weighs. On the weight, Srinivasan was explicit: "Importantly, no single piece of feedback determines how content is distributed," and "we've built safeguards to help prevent individual feedback from being used to unfairly target other members."3

When enough members flag a post, its author sees a tip in Post analytics. The Help Center says the tip "does not mean that your post was taken down, or that a policy decision was made about your content."1 Flags still count for something. "These reports help inform our classifiers," LinkedIn says.16 Engadget, citing an unnamed spokesperson, reported in July that flagged posts would see "reduced algorithmic reach similar to" a "not interested" click.67 LinkedIn has published no figure.

If you get the tip:

  1. Leave the post up. The tip means no takedown and no policy decision.
  2. Look for a claim someone could disagree with. If there is none, that is the fix for your next post.
  3. If there is one and readers flagged it anyway, look at how it sounds. Srinivasan has said that many posts and comments "written in the same AI style" make everything feel less human.53

AI detectors and LinkedIn measure different things

LinkedIn publishes no share of its feed that it calls slop. The "X% of LinkedIn is AI" figures come from companies that sell AI detectors.

  • Pangram, more than 40% of long posts "flagged as fully AI-generated."40 Its sample came from people who installed its browser extension and chose to share data, people who wanted an AI detector on their feed.41 Its 0.01% false-positive rate is the vendor's own claim.40
  • Originality.ai, 81.2% of 5,000 public posts found by topic search scored likely AI in July 2026.38 Its earlier study, of posts by 99 influential profiles, found 53.7%.39 The samples differ, so the jump means little.

Each of these asks whether software wrote the text. LinkedIn asks whether the text says anything, and for the private tip it sends authors, it chose people over a detector.

“we want members to get feedback from real humans on what sounds authentic”

Hari Srinivasan, Chief Product Officer, LinkedIn

The rest of his sentence: "not just have an AI detector review it and get it wrong."10 Put two posts through both instruments and the gap is plain.

Illustration

Typed by hand, says nothing

“Show up every day. Add value at every touchpoint.” A detector should call it human. It is exactly the generic content LinkedIn says it shows less of.

Illustration

Drafted in ChatGPT, says something

A number from your own pipeline and a claim a peer could argue with. A detector can call it AI. It should pass LinkedIn's test, because it adds something.

Save this

Before you repost a LinkedIn AI statistic

  1. What does it count? Generic posts, AI-written posts, views and flags are four different things.
  2. Is there a false-positive rate? A catch rate without one is half a number.
  3. Is it one post or the whole platform? A drop in total views says nothing about yours.
  4. Who drew the line? If the system that defines slop also reports its own success, the line can move.
  5. Does the source sell an AI detector?
  6. Does it ask "did software write this?" or "does this say anything?" LinkedIn says it cares about the second.

handful.cc/blog/linkedin-slop-numbers

Judge your draft by its claim

Srinivasan's shortest version: "AI and slop are not the same thing."10 So how much of a ChatGPT draft you changed is the wrong measure. You could rewrite every word and still have a generic post.

My test is one sentence. Can you state the post's claim in a line a colleague could disagree with? "Consistency builds pipeline" fails, because nobody argues with it.

LinkedIn even wrote the brief: "We want the real take, the dive into what the latest headline means for me, how a new automation tool saves time (or not), a sly sense of humor."20 The percentages will keep moving. That brief is the part worth keeping.

Next in the series: The LinkedIn Algorithm, According to LinkedIn.

Sources

  1. 01LinkedInBest practices for content created with the help of AI, LinkedIn Help Center, ~2026-08 (live by 2026-09-13). linkedin.com
  2. 03LinkedInFollow-up on "Seems like AI slop" ("One more follow up on AI Slop"; duplicate of [19]), Hari Srinivasan, CPO, LinkedIn post, 2026-08-19. linkedin.com
  3. 09LinkedInKeeping conversations real on LinkedIn, Laura Lorenzetti, LinkedIn Pulse, 2026-05-20. linkedin.com
  4. 10LinkedInAI slop is a top priority for all of us, Hari Srinivasan, CPO, LinkedIn post, 2026-07-30. linkedin.com
  5. 16LinkedInHow LinkedIn is Continuing to Tackle AI Slop, LinkedIn Pressroom, 2026-09-15. news.linkedin.com
  6. 20LinkedInKeeping conversations real (announcement post and member comments), Laura Lorenzetti, LinkedIn post, 2026-05-20 (member comments third-party). linkedin.com
  7. 30PressLinkedIn cracks down on AI slop with 94% detection accuracy, The Next Web, 2026-05-20. thenextweb.com
  8. 37PressLinkedIn says its AI slop button is working, Engadget, 2026-08-21. engadget.com
  9. 38Third-partyLinkedIn AI study (updated series), Originality.ai, 2026-07-30. originality.ai
  10. 39PressOver half of LinkedIn posts are now likely AI, PPC Land, 2026-01-22. ppc.land
  11. 40Third-partyAI Content Is Everywhere on Social Media, Especially LinkedIn, Pangram Labs, 2026-07-09. pangram.com
  12. 41PressAI slop writing has taken over the internet, particularly LinkedIn and X, The Register, 2026-07-09. theregister.com
  13. 43PressLinkedIn's war on AI slop is ... an admission that the platform lost control of its feed, The Decoder, 2026-05-20. the-decoder.com
  14. 53LinkedInOn AI slop... First, I see it too, Hari Srinivasan, CPO, LinkedIn post, 2026-05-21. linkedin.com
  15. 67PressLinkedIn is testing a 'seems like AI slop' reporting tool, Engadget, 2026-07-30. engadget.com

Comments that add something, on the posts that matter.

Handful finds up to five LinkedIn posts a day from people who need what you sell, and drafts a comment in your voice. You read it, edit it, and post it yourself.

Start free →