AI detection company Pangram Labs published data on July 9, 2026 showing that LinkedIn accounted for 62% of all AI-generated content flagged across five major social platforms, even though LinkedIn posts made up only a third of the items scanned.

Your Feed Has a Viewability Problem

If nobody's reading it, did it ever really load? Add PPC Land as your preferred source in Google Search and actually see the stories that matter.

Fix My Sources

A first look at what a browser extension has been counting

Pangram Labs launched a Chrome extension on April 24, 2026 that scans social media posts as users scroll and flags content the company's detection model classifies as AI-generated. Two and a half months later, the company released its first analytics report drawn from that tool, and the numbers describe a professional networking platform where machine-written text has become the dominant mode of longform communication rather than an exception to it.

The report covers 1,002,627 posts collected from users who opted into sharing scan data for research purposes. Every post came from one of five platforms: LinkedIn, Medium, Substack, X/Twitter, and Reddit. Pangram counted each post only once and excluded anything shorter than 50 words, a threshold the company set because shorter samples do not give its model enough text to classify reliably.

The headline figure sits at the intersection of two separate measurements. Across the full dataset, the average AI rate for scanned items was 13.8%. But LinkedIn posts represented roughly a third of everything scanned, while accounting for close to two-thirds of everything flagged as AI-generated. That imbalance is the clearest single number in the report, and it says something specific: LinkedIn is not merely present in the AI content problem. It is disproportionately responsible for it, relative to how much of the platform gets scanned in the first place.

Why longform content is where the problem concentrates

Length turns out to matter more than platform in one respect. On four of the five platforms Pangram measured, longer posts were more likely to be flagged as AI-generated than shorter ones. Across the combined dataset, one in four longform items - defined as posts over 250 words - came back as fully AI-generated, a rate of 25.72%. Substack broke that pattern; there, fully AI-generated content held roughly flat regardless of length, and longer, more substantial Substack posts were actually slightly less likely to trigger a flag than shorter ones on the same platform.

LinkedIn sat at the extreme end of the longform measurement. More than 40% of longform LinkedIn posts came back as fully AI-generated, according to Pangram's data. X/Twitter told a related but distinct story: when Pangram counted mixed AI-and-human content alongside fully AI-generated posts, nearly half of all X longform articles fell into one category or the other. Specifically, 23.9% of X articles were classified as fully AI-generated and another 22.9% as AI-assisted or mixed, leaving only 53.2% of X articles as fully human-authored under Pangram's classification.

Even the platform with the lowest combined AI rate among longform environments still showed meaningful exposure. Substack, despite bucking the length pattern seen elsewhere, still had more than a fifth of its posts - 21.9% - flagged as either AI-generated or AI-assisted. Pangram frames this as consistent with a broader trend the company has tracked outside social media, pointing to the rise of AI-generated content appearing in newspaper opinion pieces as a parallel data point.

The mechanics behind LinkedIn's concentration

What makes LinkedIn's share notable is that it runs against an intuition many people might hold about anonymity and AI use. Pangram's data indicates that people are more willing to let AI speak on their behalf in professional settings tied to their real identity than they are on casual or anonymous platforms - the opposite of what an assumption built around plausible deniability would predict.

Pangram also points to product design as a contributing factor. LinkedIn built AI writing assistance directly into its posting interface through a button originally labeled "Write with AI," which the company has since rebranded as "Enhance post" while keeping the same underlying AI writing functionality. That native tooling sits inside the platform's own composer, available to any user drafting a post.

The pattern has not gone unnoticed inside LinkedIn either. According to Pangram, an executive at the company recently announced that LinkedIn would begin detecting and downranking AI-generated posts using an in-house algorithm - and Pangram notes, with evident irony, that the announcement itself registered as AI-generated under detection analysis. Pangram's report does not resolve whether LinkedIn's downranking effort is functioning as intended; what its own scan data shows is that AI writing remains heavily present on the platform regardless of that stated intention.

LinkedIn's own infrastructure has been in flux over the same period. The platform rebuilt its feed ranking system from scratch in March 2026, replacing a multi-source retrieval architecture with a unified large-language-model-based approach for both retrieval and ranking. Whether that infrastructure change interacts with any AI-content downranking effort is not addressed in Pangram's report, and the two systems - content ranking and content-origin detection - are not necessarily the same mechanism.

A composition effect on Reddit inverts expectations

Reddit's numbers illustrate why raw platform-level AI rates can mislead without a breakdown by content type. Reddit had the highest scan volume of any platform in Pangram's dataset, making up 36.7% of everything scanned. Yet its combined AI share came in at just 4.4%, the lowest of the five platforms measured.

The explanation lies in how Reddit's content splits between replies and top-level posts. Replies made up 72% of everything Pangram scanned on Reddit, and those replies were overwhelmingly human-authored: 98.1% of them, according to the data. Top-level posts told a different story. Those were considerably more likely to be AI-written, landing at 11.6% - a rate close to X/Twitter's 10.0% AI-saturation figure for that platform overall.

LinkedIn showed a version of the same pattern, though less pronounced. A top-level LinkedIn post was 1.35 times more likely to be AI-generated than a comment on that same platform, per Pangram's figures. But the relationship inverts once length is controlled for: LinkedIn comments were actually slightly more likely to register as AI-generated than top-level posts once post length is held constant. Reddit did not show that same inversion. There, the AI-rate gap between top-level posts and replies held steady independent of length - top-level Reddit posts remained 5.25 times more likely to be AI-generated than replies, even after controlling for how long each post ran.

Pangram frames Reddit's clean reply data as revealing a blind spot in how platforms typically fight automated content. Reddit's spam policy, the company notes, effectively removes accounts that use AI to auto-generate spam replies at volume - but that approach only catches the lowest-effort content. Top-level Reddit posts make up only about a quarter of everything on the platform, in Pangram's sample, yet they carry outsized influence on what readers actually see and engage with. Their lower overall volume, relative to the flood of replies, lets AI-authored top-level content slip past moderation systems built around rate-limiting and volume detection.

Method and what the data does and does not cover

Pangram built its dataset from users who explicitly opted in to share their scan results for research. The company describes this as an intentional design choice: the goal was direct visibility into what AI-generated content people are actually encountering on their feeds, rather than a modeled estimate. Every post in the 1,002,627-item dataset was run through Pangram 3.3, the company's current AI detection model, which Pangram states achieves a 0.01% false positive rate.

That detection-model claim is Pangram's own figure, drawn from its published report; it has not been independently verified in this article and readers evaluating the underlying accuracy of the classification system should treat it as a company-stated benchmark rather than a peer-reviewed result. The same caveat applies to the AI-generation classifications themselves, since they rest entirely on Pangram's proprietary model rather than a disclosed, cross-platform ground truth.

The report's scope is also worth stating plainly. It covers five platforms - LinkedIn, Medium, Substack, X/Twitter, and Reddit - and excludes others such as Facebook, Instagram, TikTok, and YouTube. Medium's figures are referenced in Pangram's underlying charts but are not broken out with the same level of narrative detail as the other four platforms in the company's public writeup. Posts under 50 words were excluded entirely, meaning the findings speak to longform and mid-length content rather than short-form posts, captions, or brief comments below that threshold.

Where this sits alongside other measurements of the same problem

Pangram's LinkedIn figures land alongside a body of similar findings that PPC Land has tracked over the past several months, and the consistency across independently produced studies is itself notable. Originality.ai reported in January 2026 that 53.7% of long-form LinkedIn posts it analyzed were "likely AI", based on a sample of 3,368 posts from 99 influential profiles across eleven industries. That figure and Pangram's 40%-plus longform rate are not directly comparable - different detection models, different sampling methods, and different definitions of "AI-generated" versus "likely AI" separate the two studies - but both point toward LinkedIn carrying an unusually high concentration of machine-written longform content relative to other platforms.

The broader pattern of AI-generated material displacing human-authored content extends well beyond LinkedIn. PPC Land's coverage of AI slop dynamics in AI search systems documented how fabricated information, once published, can be picked up and repeated by AI systems as though it were verified fact, creating a feedback loop that keeps false claims circulating long after their origin has been debunked. A March 2026 investigation exposed a network of more than 200 domains using large language models to produce clickbait content at industrial scale, for costs as low as $2.25 per article - an economic model that helps explain why volume-based AI content keeps appearing regardless of platform-level countermeasures.

The advertising verification industry has already built commercial products around detecting and filtering this category of content. Integral Ad Science moved its Low-Quality GenAI Avoidance tool out of open beta and into general availability in late May 2026, following performance data drawn from more than one billion ad impressions. That product, however, applies only to English-language text content on the open web - it does not cover in-app inventory or social platforms, which is precisely where Pangram's LinkedIn and Reddit figures are drawn from. The gap between what ad-verification tools currently filter and where Pangram finds AI content concentrated is not a contradiction; it reflects two different measurement efforts covering adjacent but non-identical territory.

LinkedIn's own advertising quality has drawn separate scrutiny on unrelated grounds. A June 2026 report on invalid traffic across major ad platforms found LinkedIn carrying the highest bot-click rate among the channels measured, a finding about fraudulent clicks rather than AI-generated content, but one that adds to a broader picture of LinkedIn facing quality questions across more than one dimension simultaneously.

Content-origin concerns are not unique to LinkedIn among social platforms, either. Research examining YouTube Shorts found that 21% of videos shown to new users qualified as AI-generated slop, alongside a further 33% classified separately as low-quality "brainrot" content, indicating that video platforms face a parallel version of the same underlying dynamic Pangram documents in text.

What the numbers mean for people relying on these platforms

For marketers and communications professionals who use LinkedIn as a primary channel for reaching business audiences, Pangram's data adds a concrete figure to something many practitioners have already suspected anecdotally: a large share of what appears in a professional feed is not written by the person whose name sits above it. That matters for anyone evaluating engagement benchmarks, competitive content strategy, or the credibility signals audiences read into a post's authorship.

It also raises a narrower but practical question about platform trust. If downranking efforts aimed at AI content are underway, as Pangram's report indicates LinkedIn has stated, marketers publishing genuinely human-written content have a stake in understanding how such systems distinguish their work from the roughly 40% of longform posts Pangram's model classifies as fully machine-generated. Pangram's report does not evaluate whether LinkedIn's stated downranking mechanism works, nor does it claim to. What the data shows is the scale of the problem such a mechanism would need to address, not the mechanism's effectiveness.

Reddit's data offers a different kind of lesson, one about where moderation resources are best spent. A platform can maintain a low overall AI rate while still carrying a meaningfully elevated rate within a specific content category - in this case, top-level posts rather than replies. Pangram's finding that Reddit's low aggregate AI share masks a 5.25-times gap between top-level posts and replies suggests that platform-wide AI statistics, on their own, can understate risk concentrated in the content types readers see and act on most.

Timeline

  • April 24, 2026 - Pangram Labs launches its Chrome extension, enabling users to scan social media posts for AI-generated content as they scroll.
  • Users who opt in begin sharing anonymized scan statistics with Pangram for research purposes, building the dataset underlying the July report.
  • July 9, 2026 - Pangram Labs publishes its first analytics report based on data collected since the extension's launch, covering 1,002,627 scanned posts across LinkedIn, Medium, Substack, X/Twitter, and Reddit.

Summary

Who: Pangram Labs, an AI detection research company, published the report. Max Spero, the company's CEO and co-founder, is credited as the report's author.

What: An analytics report drawn from Pangram's Chrome extension, covering 1,002,627 social media posts scanned across LinkedIn, Medium, Substack, X/Twitter, and Reddit. The report found LinkedIn accounted for 62% of all flagged AI content despite representing about a third of scanned items, with more than 40% of LinkedIn longform posts classified as fully AI-generated. Reddit showed the lowest combined AI rate at 4.4%, driven by a composition effect in which replies - 72% of scanned Reddit content - were 98.1% human-authored.

When: The Chrome extension launched April 24, 2026. The analytics report itself was published July 9, 2026.

Where: The data spans five social media platforms globally, drawn from users of Pangram's browser extension who opted into sharing scan statistics for research purposes.

Why: The report matters because it quantifies, with a large opted-in dataset rather than a modeled estimate, how unevenly AI-generated content is distributed across social platforms and content types. For marketers and communications professionals, the figures offer a concrete baseline for a phenomenon - AI-authored professional content - that had previously been documented mainly through smaller, platform-specific studies.