The Pangram study published on July 12 scanned over one million posts across five platforms between April and June 2026, and comes out with a stark hierarchy: LinkedIn concentrates 41% AI content on long posts, almost double Reddit’s share. Two-thirds of all AI content Pangram detected across the five platforms sits on LinkedIn alone.
Key Takeaways
- LinkedIn accounts for 41% AI content on long posts, Medium 31%, X 29%, Reddit 13%, Substack 10%.
- LinkedIn holds roughly two-thirds of all AI content detected by Pangram across the five platforms.
- Reddit remains the counter-model: 98% of replies are written by a human.
One million posts, five platforms, a hierarchy that pops on the first read
Pangram scanned just over a million posts published between April and June 2026 across LinkedIn, Medium, X, Reddit and Substack. Detection ran on the company’s own Chrome extension, which Pangram advertises with a false positive rate of 0.01%. The methodology isolates long-form content first, from 250 words up, then short-form, between 50 and 250 words.
On long-form, LinkedIn lands at 41%. Medium follows at 31%, X at 29%, then Reddit at 13% and Substack at 10%. Short-form redistributes the cards but keeps LinkedIn on top, with 30% AI content, against 28% for Medium, 12% for Substack, 9% for X and 3% for Reddit.
The report highlights a ratio that shifts the reading. LinkedIn represents only about a third of all posts scanned in Pangram’s sample, but soaks up roughly two-thirds of AI-detected content across the five platforms combined. The professional network is over-exposed well beyond its share of the sample.
Reddit occupies the other end of the spectrum. Replies to threads are human-written 98% of the time, a stable number across the three months observed. Standalone posts pushed to the platform show higher AI rates, but the conversation itself stays overwhelmingly human, which gives Reddit a fingerprint very different from the rest of the sample.
Pangram slips a methodological caveat into the report: the real AI rate could be even higher, because the detector recognizes human content better than AI content. The classifier makes more errors labeling AI text as human than the reverse, which pulls the percentages downward.
Why LinkedIn stacks up so much AI content
The direct reading of the numbers points at the economic incentive. On LinkedIn, a long post with a polished corporate tone and a thesis-arguments-call-to-comment structure triggers the reach mechanic. Generic models like ChatGPT, Claude and Gemini spit out exactly that format in thirty seconds, with almost no stylistic variance from one user to the next.
The second driver is fear of silence. Many professional accounts have nothing to say but insist on posting several times a week to stay visible on the profile page. The generative model fills that gap with grammatically clean, surface-readable text, which keeps the weekly ritual sustainable even without any new substance.
The third dynamic is quieter but heavy. Anthropic recently admitted that the actual enterprise use of Claude leans heavily on content and ops work, not on writing code. On those workloads, generating social content becomes a natural extension of the workflow, often copy-pasted with little editorial pass.
On X, the share of AI-written long-form gets close to 50% when adding pure AI and AI-assisted content together. The difference with LinkedIn sits in the format: X is still dominated by short-form, where variance is harder to hide and where models produce text that reads more “bot” to a trained eye. Long-form on X is a newer use, pushed by the articles now published straight in the feed.
Substack stays surprisingly low at 10% AI long-form. The platform rewards a recognizable authorial voice and a strong stylistic signature, two criteria generic models still miss. Substack authors also have a financial reason to hold the pen: their direct revenue depends on a paying base that quickly spots voiceless text.
More articles on Horizon
- Anthropic Extends Fable 5 Promo Through July 19
- Claude Runs More Office Work Than Code, Anthropic Admits
- Apple Sues OpenAI Over iPhone Secrets and 400 Hires
LinkedIn moderation is the next ridge to climb
LinkedIn kicked off, in response to the Pangram report, an enforcement phase against AI posts. The platform has not publicly detailed the thresholds applied or the sanctions, but the rollout is confirmed. The bet runs both ways: reduce the saturation effect that dents time-on-feed, and defend the LinkedIn brand’s editorial credibility with advertisers.
The technical challenge is real. An AI detector calibrated on Pangram-type scores has to hold several constraints: recognizing recent frontier models, handling human-AI co-authoring, avoiding false positives on non-native English writers whose phrasings can echo LLM outputs. Every error costs on the user side, and LinkedIn will need to tune quickly or raise the tolerance threshold.
The trust angle on the recruiter side sits apart. A candidate who publishes a weekly AI-written post on their industry risks, on a short horizon, losing the editorial credit LinkedIn has long served as a soft-skill signal. Recruiters train their eye to spot the generic pattern, with or without a detector. The professional account with no original substance becomes more visible, not less.
Competitors are watching. Substack could decide to lean on its low AI share as a commercial argument for professional authors. Reddit could tighten moderation on standalone AI-written posts to preserve the “real conversations” signature the Pangram report grants it. Medium has less structural room, with a model very close to LinkedIn on long-form publishing.
The broader wave is one of audience metrics. A Pangram-style quantitative study joins a growing set of works that document the massive use of LLMs in contexts historically dominated by humans, like the study on 26,000 students facing real exams. The question that keeps coming back is the same: at what point does the measure of a human signal get too noisy to remain useful, and what gets built in its place.
Follow the story on Horizon.
