LinkedIn just deleted its own AI writing tool.

That is the part of this announcement worth sitting with. The “enhance your post” feature, the one that rewrote your draft for you, is gone, replaced by a proofreader that fixes the writing without touching the voice. The company that removed the friction from publishing machine-written text has deliberately put some of it back.

It shipped alongside a button that lets any reader flag your post as AI slop.

Reading as AI became a distribution variable

Most marketing teams already argue about this, and they argue about it as a matter of taste.

Someone in review says a draft feels machine-written. Someone else disagrees. A sentence gets swapped, and the post ships. That process held up fine while the answer only touched brand perception, which is slow to measure and easy to defer.

The answer now touches reach. LinkedIn said it is rolling out classifiers that identify whether a post is slop or otherwise low quality, and that this feeds how much a post gets recommended beyond the author’s own network. Your readers’ aesthetic reaction is now an input to a system that hands out impressions.

That changes the priority order for anyone running B2B content. A post that reads as automated no longer costs only credibility. It costs impressions you already budgeted for, and it costs them invisibly, because a suppressed post looks exactly like a slow week.

We argued a version of this when we wrote that your AI marketing needs a verifier. What changed is that the verifier who now matters does not work for you. They are on the other side of the scroll, holding a button.

What Pangram actually measured

Worth reading the study before the reaction, because the number circulating has been trimmed.

Pangram published its analysis of 1,002,627 posts gathered between April 24 and July 9, 2026 across LinkedIn, X, Reddit, Substack and Medium, collected through an opt-in browser extension. It counted only posts longer than 50 words and ran them through Pangram 3.3, at a stated false positive rate of 0.01%.

The platforms separate sharply. LinkedIn long-form came in above 40% fully AI-generated. X articles landed at 23.9% fully generated with another 22.9% AI-assisted, leaving just 53.2% entirely human. Reddit sat at 4.4% combined, the lowest in the set.

The number that says most about concentration is a different one. LinkedIn produced 62% of all AI-flagged content while representing only 33% of everything scanned.

Fair warning on the method. The sample comes from people who chose to install an extension, which is not a random draw from the platform, and every AI detector carries error even when the declared rate is low. Pangram CEO Max Spero described his own figures as a floor rather than a ceiling. Treat them as an order of magnitude. The order of magnitude is enough to act on.

What the button does, and what LinkedIn declined to say

On July 30, 2026, LinkedIn added “Seems like AI slop” to the report menu on every post, the same menu used to report spam. TechCrunch covered the launch that day. The reporter stops seeing the post, and the signal trains the classifiers.

LinkedIn also put a number on the automation it says it is already stopping: hundreds of thousands of automated comment attempts a day, and billions of automation attempts over recent months.

Then there is the gap, which I would rather name than paper over. Forbes pointed out that LinkedIn never spelled out what happens to a reported post beyond saying the feedback helps improve the feed. We know classifiers trim recommendation outside your network. We do not know how many reports it takes, or whether the effect accrues to the account or dies with the post. Anyone quoting you a threshold today made it up.

A moving standard with no published rubric

Hari Srinivasan, LinkedIn’s chief product officer, said two things that are worth reading side by side. That slop is hard to define and the definition keeps changing. And that people rather than automated detectors should be the ones deciding, because detectors get it wrong.

Each statement is reasonable alone. Together they are uncomfortable.

They mean the standard now shaping your reach is unpublished, moves over time, and gets applied by a tired reader thumbing past your post. There is no described appeal and no definition to audit your content against before you hit publish.

A marketing team cannot optimize against that the usual way, by finding the rule and satisfying it. What it can do is stop manufacturing the thing that triggers the reflex, and there the patterns are stable, because readers are not reacting to a model. They react to a shape.

They react to symmetry. Paragraphs of matching length, lists of three with identical cadence, openings that repeat across posts from the same author, closings that restate what was already said. Human writing is lumpier than that.

They also react to the absence of anything proprietary. When a post could have been written without knowing a single specific fact about your company, somebody else already published it this week in different words. That is the same standard we applied in the LLM judge panel for marketing content QA, where review earns its keep by testing whether the piece says something only you could say.

Two changes worth making this week

None of this requires dropping AI from your writing process. Two things are worth moving.

Change the order of production. When the draft comes from the model first and someone hunts for proprietary detail afterward, the result drifts toward the average shape, because the fact arrives as decoration. Start from the proprietary fact instead, a figure from your operation, a call your team made last month, an error you corrected, and the piece organizes itself around something no model had.

Change what review measures. Most content reviews check spelling, message and call to action. None of those catch the problem described here. Add one explicit question to the process. What in this post could not appear in a competitor’s version with the name swapped? If that answer takes a while to arrive, the draft is not ready.

Then check your own symmetry by hand. Read your last five posts back to back, in order, the way a follower sees them. If they open alike and close alike, the problem is not in any one of them. It is in the template underneath, and that is precisely what a reader detects without being able to explain it.

That exercise works better against your actual archive than against theory, and it takes under half an hour. Now is the calm moment to run it, while the button is still new and the classifiers are learning from whatever people report over the next few weeks.

Let’s find what in your content only your company could sign