LinkedIn and Substack Now Score Your Posts for AI. Here Is What Actually Gets Flagged

In the same week at the end of July, two of the platforms people use most for professional writing both shipped a detection feature aimed at their own users. LinkedIn removed the AI tool it had spent two years pushing people toward, added a button for reporting posts as “seems like AI slop,” and started quietly reducing the reach of posts its classifiers flag. Substack partnered with the AI detector Pangram so any reader can scan a newsletter and see an estimated human-versus-AI split.

Neither company is banning AI assistance. Both say, in their own announcements, that the target is narrower: content where nobody involved actually wrote or knew the material, passed off as if a person did.

That distinction matters more than the headlines suggest, because it changes what “using AI on a LinkedIn post” or “using AI on a newsletter” now means in practice.

LinkedIn Retired Its Own AI Writer

LinkedIn’s “Enhance your post” feature, which used AI to rewrite what you typed, is gone. In its place is a plainer proofreading feature that corrects your text without changing your voice, a narrower job than the tool it replaced.

Alongside that, a “Seems like AI slop” button now sits in the ellipsis menu on every post, and is coming to comments too. Chief product officer Hari Srinivasan explained the mechanics in the announcement on LinkedIn: flagged posts feed classifiers that reduce the post’s reach in recommendations outside your network, similar to how marking something “not interested” works today. LinkedIn is also privately nudging accounts through their own analytics dashboard when a post reads as heavily AI-generated, before anyone else has flagged anything.

None of this removes a post. It throttles distribution and, over time, trains a model to recognize the pattern LinkedIn wants less of.

Substack Bet On Readers Doing The Checking

Substack’s approach hands the tool to readers instead of applying it centrally. CEO Chris Best laid it out in Against Claudefishing: any post, note, reply, or comment over 100 words, published after July 21, can be scanned to see how much of it was likely written by hand versus with AI assistance. The result is visible only to whoever requests it, so it is not a public score attached to your byline.

Writers get the same tool in reverse. You can run Pangram on a draft before publishing, add an optional “How I make this” statement to explain your process, and report a scan you believe is wrong. Substack frames this as a transparency feature, not a filter, but the effect for a writer is the same practical question LinkedIn’s button raises: what is this actually measuring?

What The Detectors Actually Measure

Both tools are built on the same statistical approach every AI detector uses: how predictable the phrasing is, how uniform the sentence structure is, how much a passage resembles the smooth middle of the distribution AI models tend to produce. We covered the reliability limits of that approach in where AI writing detection actually stands, and the limits apply here too. Substantial editing after generation moves a piece across the line. Very plain, very consistent human writing can occasionally score the other way.

What that means in practice: these tools are not detecting “AI touched this text.” They are detecting text that reads as unedited and unowned, which is a much narrower thing, and one you have direct control over.

Before:

In today’s fast-paced professional landscape, effective communication has never been more important. Organizations that prioritize clear, authentic messaging consistently see better engagement and stronger relationships with their audience.

After:

We changed our standup format three weeks ago and cut it from 30 minutes to 12. The trade-off: two people now say they feel less looped in on what other teams are doing, which we have not solved yet.

The first version could describe any company on any day, which is exactly the pattern both platforms built their classifiers to catch. The second has a number, a date, and an admitted problem, none of which a generic prompt would produce unless you supplied them.

Write From Your Own Draft, Not A Blank Prompt

The practical shift is where the sentence starts. A prompt like “write a LinkedIn post about our product launch” hands the tool the substance as well as the phrasing, and that is the combination both platforms are now built to notice. Three of your own bullet points, run through a tone pass, is a different thing structurally, even before you get to how it reads.

A Wrivio Context for polishing a post you already drafted could say:

Rewrite this LinkedIn post for clarity and a professional but not corporate tone. Keep every number, name, and specific claim exactly as written. Do not add an opening line about trends, growth, or fast-paced industries. Do not add a closing question. Shorten only where a sentence repeats a point already made.

Press Ctrl+Shift+Space, paste the draft, and check that the diff only touched phrasing, not the argument. If the rewrite is doing more work than that, the post did not have enough of your own material to begin with, and that is worth knowing before you publish, not after a reach drop or a reader’s scan.

What This Signals Going Forward

Two of the platforms people publish professional writing on both moved in the same direction within the same week, which is a stronger signal than either move alone. The C2PA provenance requirements under the EU AI Act are a related but separate track, aimed at disclosure rather than reach or reader trust. Either way, the response that survives is the same: keep the substance yours, and use a tool to fix how it reads.

Common Questions

Will LinkedIn’s slop button get my post taken down?

No. It feeds a classifier that reduces the post’s distribution in recommendations outside your network, comparable to how a “not interested” mark works. LinkedIn has not described it as a removal mechanism.

Can other people see that my Substack post was scanned by Pangram?

No. A scan result is visible only to the reader or writer who requested it, not published alongside the post.

Does editing a draft with a rewriting tool trigger these detectors?

Not reliably by itself. Both tools are built on statistical patterns that pick up unedited, generic phrasing more than they pick up “AI was involved at some stage,” a distinction covered in more depth in our detector reliability piece.

Should I add a disclosure statement to posts I edited with AI?

Substack’s “How I make this” field is optional and framed around explaining your process, not confessing to a violation. Whether disclosure is advisable elsewhere depends on your platform and audience; there is more on the underlying reasoning in the case for using AI carefully rather than not at all.

Is this the same requirement as the EU AI Act’s content marking rules?

No. Article 50’s provenance marking targets synthetic content that could mislead someone about its origin, a narrower legal obligation. LinkedIn and Substack’s tools are platform product decisions aimed at feed quality and reader trust, not compliance.

Download Wrivio for Windows to polish a draft you already wrote, locally if you want the text to never leave your machine before you publish it.