Reach on AI-drafted posts dropped noticeably for many creators in late 2025 and early 2026. Here's what appears to be getting flagged, what's still working, and what to do if your account took a hit.
LinkedIn has not publicly confirmed AI detection, but reach patterns in late 2025 and early 2026 strongly suggest it. What appears to be flagged: em dashes, formulaic "it's not just X, it's Y" structures, hedging language, and specific openers common in ChatGPT defaults. Creators switching to voice-matched AI output or manual writing have recovered reach within 2 to 4 weeks.
LinkedIn has not published a confirmation that they downrank AI-generated content. What they have said is that their algorithm optimises for "authentic expertise" and "original thinking" (corporate speak that could mean anything).
What creators are reporting, though, is consistent enough to treat seriously.
Across dozens of creator conversations we have had since late 2025, the shape of the drop is similar.
A founder uses ChatGPT to draft LinkedIn posts. Early months work fine. Reach is normal or growing. Somewhere between late 2024 and early 2026, reach starts declining. The decline is gradual at first, then steeper. By 6 months into the decline, the same creator is seeing 60 to 85% fewer impressions than peak.
The creators who switch back to manual writing often see reach recover within 2 to 4 weeks.
This is anecdotal, not proof. But the pattern has been reported enough times that assuming LinkedIn is doing something is the safer assumption.
From the set of posts that saw reach drops, we have identified patterns that appear to correlate with downranking. None of these are confirmed by LinkedIn. All of them are observable in the affected posts.
Em dashes. The em dash (—) is overrepresented in ChatGPT's default output. It appears in roughly 2% of default sentences, which is 3 to 5x higher than typical human writing. Posts with heavy em dash usage have shown the steepest reach drops.
"It's not just X, it's Y" structures. This rhetorical pattern is almost a signature move for ChatGPT when asked to write persuasively. It is rare in natural writing but common in AI-assisted writing. Posts leaning on this structure appear to get less distribution.
Hedging language. "It's worth noting that...", "In many ways...", "It's important to remember...". All of these are ChatGPT defaults. All of them are rare in natural human writing. Their presence at density appears to flag the content.
Formulaic closing structure. ChatGPT defaults to a summary-paragraph-plus-CTA closing structure that reads as essay-like, not native to LinkedIn. Native LinkedIn closings are either declarative mic-drops, soft questions, or just hashtags. Essay-shaped closings appear to underperform.
"Let's dive into..." / "Let me walk you through...". Transition phrases that ChatGPT uses when it has been asked to write instructionally. They are natural in long-form prose but read as filler on LinkedIn.
Over-balanced sentence length. ChatGPT tends to produce sentences within a narrow length range (12 to 20 words). Natural writing has much more variation, with occasional very short sentences (2 to 5 words) that hit like beats. Uniform sentence length reads as machine output and appears to be part of the detection signal.
Creators who produce content successfully on LinkedIn with AI assistance as of April 2026 share a few traits.
Voice-matched output. They use voice prompts or custom GPTs trained on their existing writing, not default ChatGPT. Output passes the "would I have written this" test before they post.
Variable sentence length. Outputs include 2 to 5 word sentences alongside 15 to 20 word sentences. Rhythm varies. Short punchy sentences land as beats.
Personal specifics. Posts include details that cannot be recycled. A real number. A named client. A recent Tuesday. ChatGPT defaults to generic examples. Creators add specifics in the edit pass.
No em dashes. The most consistent single variable. Creators who removed em dashes from their AI-assisted output saw reach recover.
Shorter paragraphs. Most successful LinkedIn posts in 2026 use 1 to 2 sentence paragraphs with line breaks between them. This is distinct from essay writing. ChatGPT defaults to 3 to 5 sentence paragraphs. Breaking them up before posting is part of the fix.
Assume LinkedIn is flagging AI patterns. Whether or not it is, the fix is the same: write in a way that is distinguishable from ChatGPT defaults.
Two paths.
Path 1: write manually. Your reach recovers in 2 to 4 weeks. You pay the 90 to 120 minutes per post time cost. Not sustainable for most solopreneurs long-term, but works.
Path 2: use a voice prompt. Spend a weekend building a voice system from your existing content. Voice prompt instructs any AI tool to write in your actual voice, with short sentences, your specific vocabulary, and zero em dashes. Output no longer looks like ChatGPT default. Reach recovers on the same timeline as manual writing, but the time-per-post drops to 15 to 20 minutes.
The voice prompt path is covered in full in The Voice System Playbook. Thirty pages, free, every prompt copy-paste ready.
If reach has only recently started dropping and you want to stop the slide, here is the order of operations:
Do this for 2 weeks. If reach recovers, you had mild AI-pattern exposure and the edit pass is enough. If reach does not recover, the problem is deeper and a full voice system rebuild is warranted.
We are not arguing anyone should hide that they use AI. Many creators disclose openly. Disclosure does not trigger downranking in any pattern we have seen.
What triggers downranking is the mechanical output defaults, not the act of using AI. Voice-matched AI output from a well-calibrated voice prompt is harder to distinguish from manual writing, either by an algorithm or by a human reader. That is the point of the voice system approach: you get the time savings of AI without the reach penalty of ChatGPT defaults.
Assuming LinkedIn is running AI detection (reasonable assumption), their next moves are likely:
The long-term hedge is the same as the short-term fix. Build a voice system. Train it on your actual thinking. Add personal specifics in every post. The output becomes indistinguishable from manual writing on both algorithmic signals and human reader signals.
You get the time back. You keep the reach.
If your reach dropped because of AI patterns, The Voice Build is the fastest fix. We run the analysis, build the voice prompt and custom GPT, and ship in 3 working days. $497 founder pricing (first 5 buyers), $997 standard.
See The Voice BuildNo. LinkedIn has not published a specific AI-detection policy. What they have said is that their feed optimises for authentic expertise and original thinking. The downranking behaviour is inferred from creator-reported patterns, not documented policy.
Most creators we've spoken to report recovery within 2 to 4 weeks of switching to voice-matched output or manual writing. The rate depends on account size, posting frequency, and how heavily flagged the previous pattern was. Larger accounts recover more slowly because the algorithm accumulates more signal history.
Safer, yes. But the cost is 90 to 120 minutes per post in your time. Most solopreneurs cannot sustain that. The voice system approach is a middle path: keep the time savings, avoid the pattern triggers.
Neutral in our observation. Disclosure does not appear to affect reach meaningfully either way. What matters is the content mechanics, not whether you disclose. Do whatever aligns with your audience's expectations.