Pangram, a company that detects AI-generated text, scanned roughly a million social media posts between April and June 2026. The findings, summarized by International Cyber Digest on X: a quarter of longform posts across platforms are now fully AI-generated — and LinkedIn produces more of it than anywhere else.
The numbers, restricted to longform items of 250+ words:
- LinkedIn: 41% fully AI-generated
- Medium: 31%
- X/Twitter: 29%
- Reddit: 13%
- Substack: 10%
And the more striking figure isn’t even in that list: LinkedIn accounted for only 35% of all posts Pangram scanned, but 62% of everything flagged as AI-generated. It isn’t just producing more AI content proportionally — it’s producing it out of all proportion to its share of the conversation. Reddit, by contrast, is the outlier holdout: 98% of replies there still read as human-written.
Why LinkedIn Specifically
This isn’t really surprising once you think about the incentives. LinkedIn’s algorithm rewards engagement bait — the hook-lesson-framework-question structure that reads well in a feed and is trivially easy for a model to reproduce at scale. “Thought leadership” as a genre was already halfway to a template before AI showed up; a model just closes the gap. And the professional stakes of a bad LinkedIn post are lower than a bad byline under an editor — nobody fact-checks your carousel.
Substack sitting at 10% tells the same story from the other direction: it’s a platform built around a paid, named relationship with readers, which is exactly the incentive structure that discourages publishing unreviewed model output.
The Real Risk Isn’t Noise, It’s Trust
The obvious read is “there’s a lot of AI slop on LinkedIn, annoying.” The actual problem for anyone building a technical reputation there is sharper: your genuinely-written post, with real numbers from a real production incident, now has to visually and structurally compete with a wave of near-identical generated lookalikes using the same format. The signal-to-noise problem is really a differentiation problem — readers (and the algorithm) can’t tell your specific, hard-won post from a templated one at a glance, and that’s a cost you didn’t used to pay.
What Actually Differentiates Now
I write every article on this site from direct production experience — the About page here says as much, and it’s not a marketing line, it’s an editorial policy, because this is exactly the environment that policy is a defense against. A few things that are genuinely hard for a template to fake:
- Specific numbers from your own work — “reduced GPU cost 30%” is a claim; “reduced GPU cost 30% by moving batch inference off on-demand A100s onto a MIG-partitioned pool with these three config values” is evidence.
- A willingness to be wrong or opinionated — generated thought-leadership content is almost always agreeable and hedged. A real take has an edge.
- Details that only exist if you did the thing — the error message you actually hit, the config flag that actually mattered, the decision you’d make differently next time.
- Consistency of voice over volume — one detailed post a week beats five generic ones, and it’s much harder to fake at scale.
None of this makes AI-assisted writing bad — using a model to tighten a draft you actually wrote and fact-checked is a different act entirely from publishing unreviewed model output under your name. The distinction the data is really drawing is between authored and unauthored content, not between human and AI-assisted content.
My Take
If you’re an engineer building a public reputation right now — and I am one, deliberately — this data is less a warning than a market opportunity. The bar for “obviously real” content just got higher for free, because most of what’s competing with you isn’t. Specific, first-person, verifiable technical writing was always going to differentiate; it’s just doing more work now than it used to.
Related Reading
- Content Credentials and the C2PA Standard for Digital Provenance
- The Developer Burden of Tool Sprawl in 2026
- Karpathy’s CLAUDE.md: 4 Rules That Fix LLM Coding
Go Deeper
If you’re extending your infrastructure career into AI platform work, the same standard applies to your portfolio, not just your posts — evidence over claims.
- 🎯 AI Platform Engineer Bootcamp — build a working capstone, not a certificate you can’t defend in an interview
- 📋 AI Platform Engineer Readiness Scorecard — free, 10 minutes, no signup
