LinkedIn's "AI Slop" Button: Nobody Bans You, You Just Stop Being Shown

Your posts still publish. Your connections still like them. Nothing in your notifications has changed.

But the comments from people you have never met stopped. The profile views from outside your company stopped. And there is no message anywhere explaining why, because nothing was taken away from you — something just stopped being given.

What LinkedIn actually shipped

On 30 July 2026, LinkedIn rolled out a report option letting users flag posts that appear to be low-quality AI-generated content — colloquially, the "seems like AI slop" button. Chief Product Officer Hari Srinivasan described the problem as a top company priority, noting that automated defences already block hundreds of thousands of fake comment attempts a day.

Two mechanisms sit behind it, and they work differently:

1. User reports. A reported post sees reduced reach, in much the same way a post marked "not interested" does. It is a distribution signal, not a moderation action.
2. Algorithmic detection. Independently of reports, LinkedIn identifies generic and repetitive content and decreases its visibility outside the author's own personal network.

That second phrase is the one to sit with. Your first-degree connections still see your posts. Everything beyond them — the second and third-degree reach that is the entire reason people post on LinkedIn — is what gets trimmed.

This is the modern shape of a penalty: not a ban, not a strike, not even a notification. Just a quiet reduction in the number of strangers who will ever see you. You cannot appeal something that never announced itself.

How the detection was built — and what that tells you

LinkedIn describes an "AI solving AI" approach: human editors annotated thousands of posts as either generic or original, those examples trained machine-learning models to recognise the pattern at scale, and the resulting system runs continuously over the feed. LinkedIn has said that in testing it correctly identified generic content in 94% of cases.

Treat that number carefully — it comes via press reporting rather than a published methodology, as does the widely-quoted claim that over 40% of long-form posts on the platform are fully AI-generated. Both are worth knowing and neither should be cited as an official statistic.

But the training method is the genuinely useful detail, because it tells you what the classifier learned. It was not trained to detect "was a language model involved". It was trained on a human judgment of generic versus original. Those are different targets, and the difference is your entire strategy:

• A post written by a model, containing a specific number from your own quarter, a named client situation, and a conclusion that contradicts the usual advice → original.
• A post typed by hand at 11pm containing "Here are 5 lessons I learned about leadership" → generic.

The tool is not the variable. The presence of something only you could have written is the variable. This is the same conclusion Snapchat's Spotlight policy arrived at from the opposite direction, which we covered in Snapchat's AI video recommendation rules.

The two supporting signals nobody is talking about

Alongside the button, two smaller moves say more about direction than the button does:

LinkedIn replaced its "enhance your post" AI writing tool with a proofreading feature. The platform that spent two years offering to write your post for you now offers to check your spelling instead. When a platform removes its own generation feature, that is not messaging — that is a product bet on where the value has moved.

LinkedIn says it will privately notify users whose posts appear inauthentic due to heavy AI use, framed as helping them refine their writing rather than penalising them outright. If you get one of these, it is not a warning shot to be argued with. It is the only visibility you will ever get into a system that otherwise operates silently — read it as data.

Diagnosing whether this is happening to you

Because there is no notification, you have to infer it. The signature is specific:

• Impressions holding roughly steady, but the audience composition narrowing. LinkedIn shows you viewer companies and titles — if those have collapsed toward your own employer and immediate industry, that is the shape of network-limited distribution.
• Engagement increasingly from people you know. Same reach, different people.
• The drop coincides with a format change, not a topic change. If you started using a template or a generator around the time the numbers moved, that is your correlation.

What it is not: a sudden drop to near-zero. That pattern points elsewhere — an account issue, a link penalty, or ordinary volatility. Before concluding anything from one bad week, run the arithmetic properly; this diagnostic covers how to avoid drawing conclusions from noise.

What to change, concretely

Not "use less AI". That is the wrong axis, and following it will cost you output without fixing anything. Change the inputs:

1. Put one unshareable fact in every post. A number from your own work, a thing a customer actually said, a decision you regret. Anything that cannot be produced by someone who does not have your week.
2. Kill the structural tells. The five-lessons list, the one-line-per-paragraph cadence, the rhetorical question opener, the "Thoughts?" close. These were trained on as generic markers — not because they are bad writing, but because they are everyone's writing.
3. Take a position that costs you something. Generic content is generic partly because it is safe. A post that some readers will disagree with is structurally impossible to classify as templated.
4. If you generate at volume, vary the input, not the output. Rewriting one draft into ten versions produces ten generic posts. Ten genuinely different starting observations produce ten originals. This distinction is the whole game for anyone running multiple accounts or channels — the same logic we apply in reviewing AI content at scale.

For what it is worth, this is also the honest boundary on tooling generally — including ours. Automation can carry the parts of the job that are mechanical: the posting, the timing, the per-account bookkeeping. It cannot supply the thing the classifier is looking for. That has to come from you, and if it does not exist, no volume of output substitutes for it.

FAQ

Does being reported once hurt me?

A single report is described as functioning like a "not interested" signal — a small negative input, not a strike. There is no evidence of a permanent mark on the account. Sustained reporting across many posts is a different matter, but one annoyed reader is not a crisis.

Should I stop using AI on LinkedIn entirely?

The stated target is generic and repetitive content, not AI assistance as such — and LinkedIn continues to ship AI features itself. Abandoning the tools while keeping the templated structure would change nothing. Keep the tools, change what you feed them.

Does this apply to comments as well as posts?

LinkedIn has specifically cited automated defences blocking large volumes of fake comment attempts daily, so comment-side enforcement clearly exists and appears more aggressive than the post-side signal. If part of your strategy is high-volume commenting, that is the riskier half. Our reading of the platform's stated automation rules is in LinkedIn's automation policy in plain terms.

The uncomfortable part of this change is that there is no fight to have. The comfortable part is that the fix is not more effort — it is one specific sentence per post that only you could have written.