Analysis
AI-detection company Pangram analyzed roughly one million posts that its users organically browsed across LinkedIn, Medium, X, Reddit and Substack over a two-month period, and found that as much as 41% of longform content on LinkedIn -- posts longer than 250 words -- is likely fully AI-generated, according to reporting from 404 Media. Roughly a third of longer posts on X show the same pattern, while Reddit and Substack came in dramatically lower, at roughly one in ten longer posts.
The length-dependent pattern is the most useful finding in the data: shorter posts, between 50 and 250 words, are meaningfully less likely to be AI-generated across every platform Pangram studied. That's consistent with AI writing tools being used primarily to overcome the "blank page problem" on longer, more effortful content -- the kind of thought-leadership post or extended commentary that takes real time to draft from scratch -- rather than for quick reactions or short-form commentary where the effort savings are smaller.
The platform-to-platform variance is arguably more interesting than the LinkedIn and X numbers individually. Reddit and Substack's much lower AI-generated share suggests platform culture and audience expectations meaningfully shape how much AI-generated content actually gets published and tolerated -- Reddit's community-moderation norms and Substack's subscription-based, individual-voice model may both create stronger social and economic disincentives against obviously AI-written content than LinkedIn's algorithmically-boosted professional-content feed or X's engagement-driven timeline.
“The platform-to-platform variance is arguably more interesting than the LinkedIn and X numbers individually.”
LinkedIn, for its part, told 404 Media that professionals come to the platform specifically to hear from real people and their unique insights, and that the company actively works to reduce low-quality, automated or generic content -- while acknowledging that AI tools can legitimately help users get past writer's block rather than treating all AI-assisted writing as inherently problematic. That's a difficult needle to thread at platform scale: distinguishing between AI-assisted human insight and fully AI-generated filler is a much harder moderation problem than detecting spam or bot accounts outright.
For founders building AI-detection or content-authenticity tools, the Pangram data is a clear signal that demand for reliable AI-content detection is growing precisely because platforms themselves are struggling to self-police at the scale their algorithms reward engagement-optimized posting. For marketers and personal-brand builders using AI writing tools, the practical lesson is that audiences and platforms are increasingly capable of detecting fully-AI content even without formal tools -- which argues for using AI as a drafting aid rather than a wholesale content-generation shortcut if authenticity signals matter to your specific audience.
The bear case: AI-detection accuracy itself remains imperfect and contested -- false positives misclassifying genuinely human-written content as AI-generated are a real and often under-discussed risk with tools like Pangram's, and platforms have strong incentives to dispute findings that make their content quality look bad. What to watch next: whether LinkedIn or X respond with any concrete policy changes or labeling requirements for AI-generated content, and whether Pangram's methodology holds up to independent scrutiny as more platforms and researchers weigh in.
