AI Content Is Everywhere on LinkedIn: How to Write Posts That Don’t Look Like Slop

Here’s the number that’s making the rounds this week: 41% of longform LinkedIn posts are fully AI-written. That’s not a guess – Pangram’s July 2026 study pulled it from over a million real posts. AI content is now everywhere on LinkedIn, and the platform is fighting back. Which means the way you’ve been writing posts might already be getting throttled.

This tutorial isn’t about whether to use AI (you should). It’s about how to use it right now, in a feed where your reader assumes every post is fake until proven otherwise.

What just happened, in 30 seconds

Pangram, an AI-detection company, released a Chrome extension in April 2026 that passively scans what real users actually see while they scroll. Users who opted in helped build a dataset of 1,002,627 posts across LinkedIn, Medium, Substack, X, and Reddit, each scanned with Pangram 3.3 – a model the company claims has a 0.01% false positive rate – on any item longer than 50 words.

The result: LinkedIn had the highest AI share of any platform. It made up a third of scanned items but accounted for 62% of all AI content flagged. For comparison, Reddit and Substack sit around 10% for longform – roughly a quarter of LinkedIn’s rate.

LinkedIn saw this coming. In May 2026, LinkedIn’s product team announced changes targeting engagement bait, recycled thought leadership, and posts with obvious signs of AI construction like the “it’s not X, it’s Y” phrasing. Translation: the platform now downranks the same output its own “Enhance post” button produces.

The 6-step workflow to use AI without triggering the slop filter

Skip the “write me a LinkedIn post about X” prompt. That’s what everyone else does, and it’s exactly what gets flagged. Here’s the workflow that survives 2026.

Step 1: Dump raw material first (5 minutes, no AI)

Open a notes app. Write – in whatever broken shape – one specific thing that happened to you this week that relates to your work. A client conversation. A number that surprised you. A mistake. Ugly bullet points are fine. This is the only non-negotiable step. If you skip it, everything downstream will be generic, because the AI has nothing real to work with.

Step 2: Feed the raw material to the AI, not the topic

Bad prompt: “Write a LinkedIn post about the importance of onboarding.”

Better prompt: “Here are my messy notes from a client call yesterday. [paste notes]. Turn this into a 180-word LinkedIn post. Keep my specifics. Don’t add generic advice. No em dashes. No ‘it’s not X, it’s Y’ phrasing. Don’t start with a one-word hook line.”

The banned-phrase list matters because LinkedIn’s May 2026 authenticity update explicitly targets “it’s not X, it’s Y” constructions and other AI hallmarks. If your draft contains them, you’re writing straight into the downrank filter.

Step 3: Edit for one specific detail per paragraph

Read the AI draft. Every paragraph that could apply to any industry – cut it or replace it with a number, a name, or a moment. A post that says “we saw great results” is slop. A post that says “we cut onboarding calls from 4 to 1 and lost zero customers” is not.

Step 4: Rewrite the first line by hand

LinkedIn tests new posts with 2-5% of your network in the first 60 minutes, and only 5% of posts that underperform in that first hour recover – per Richard van der Blom’s Algorithm Insights report. Your opening line is 90% of that test. AI openers all sound the same right now (“I made a mistake yesterday.” / “Nobody talks about this.” / “Here’s what 10 years taught me.”). Type your own.

Step 5: Break the visual pattern

Every AI-drafted LinkedIn post ships with the same formatting: one sentence per line, blank line between each, three short punchy lines at the end. It’s a visual fingerprint. Combine two or three sentences into a real paragraph. Let one line run long. It’ll look messier. That’s the point.

Step 6: Run it through a detector before posting

Pangram’s Chrome extension ($20/month, launched end of April 2026) or Originality.ai will tell you if your post reads as AI. You’re not trying to fool detectors – you’re using them as a mirror. If your draft scores 90%+ AI, your reader will feel that too, even if they can’t name it.

Pro tip: If the detector flags your post as AI and you wrote it yourself, don’t rewrite it. Your voice is fine. Detectors have false positives, and sanding down your real style to dodge suspicion is how everyone ends up sounding the same.

Common pitfalls that make good workflows produce bad posts

Three specific traps, all of which I’ve watched people fall into this month:

• Using “Enhance post” on your own draft. LinkedIn’s built-in button (formerly Write with AI) tends to smooth out exactly the friction that makes writing sound human. The button was rebranded but still offers AI writing assistance. Use an external model where you control the prompt.
• Replying to comments with AI. This one is counterintuitive: when controlling for length, LinkedIn comments are actually slightly more likely to be AI than top-level posts, per Pangram’s data. Readers notice AI in replies faster than in posts, because comments are supposed to be casual.
• Overcorrecting on em dashes. The em dash panic isn’t really an AI signal – it’s a stylistic overlap. LLMs learned from professional writers who use em dashes constantly. If you already use them, keep using them. The tell isn’t the punctuation, it’s the combination of polished surface with zero specificity.

What the results actually look like

Document posts (PDF carousels) are hitting 6.60% engagement – the highest of any LinkedIn format – while standard text posts struggle to break 2%. Meanwhile, as of the 2025 report, views are down 50% year over year, engagement is down 25%, and follower growth is down 59%, according to Richard van der Blom’s Algorithm Insights report. The floor is falling out for average posts, and AI-drafted text is average by definition.

Odd math, when you think about it. The bar to stand out is dropping in absolute terms – 41% of the feed is fully generated – but rising in relative terms because 360Brew, LinkedIn’s unified AI ranking system that replaced the legacy content infrastructure, now distributes based on meaning and topical relevance rather than keyword matching. A post that says something specific about a real thing you did will out-distribute ten posts about “authentic thought leadership.”

When not to use AI for LinkedIn at all

Skip AI drafting entirely in three cases:

1. Grief, apology, or personal news. Readers can smell AI here from 50 feet away, and the reputational cost of getting caught is enormous.
2. Industry hot takes. The whole point is your opinion. If AI writes it, it’s not your opinion – it’s the median opinion of everyone who wrote about the topic before 2024.
3. If you work in Design, Architecture, or Wellness.Originality.ai’s 2025 study found these industries had the highest percentage of Likely AI posts. The suspicion is already baked in. Manual writing is a competitive edge in these niches right now.

There’s a question I keep coming back to, and I don’t have a clean answer for it: if LLMs were trained on the best professional writing – the essays, the editorials, the New York Times columns – then “sounding like AI” and “sounding like a good writer” partially overlap. Are we teaching a generation of professionals to avoid clarity because clarity now looks synthetic? Maybe. That question doesn’t have a clean answer yet.

FAQ

Will LinkedIn actually ban my account for using AI?

No. It’ll just quietly reduce your reach. The May 2026 authenticity update downranks obvious AI patterns rather than removing accounts, and there’s no public detection threshold you can look up.

Which AI tool is best for LinkedIn posts right now?

The tool matters less than the prompt. Claude, ChatGPT, and Gemini all produce nearly identical LinkedIn output if you feed them the topic instead of raw material. Where they diverge is in following negative constraints – telling the model “no em dashes, no ‘not X but Y’, no one-sentence-per-line formatting” gets meaningfully different results across models. Test the same messy notes on two models and see which one keeps more of your specifics intact. That’s your tool.

Is Pangram’s 41% number reliable?

It’s the best data we have. Pangram’s CEO Max Spero has called AI content “a tax on readers’ time” – and the detection model only catches items it’s confident about (0.01% false positive rate), meaning heavily edited AI text likely slips through. The real share is almost certainly higher than 41%.

Your next move: Open the notes app on your phone. Write three ugly sentences about something that happened at work this week. That’s your next post – before AI touches it.