I already knew LinkedIn was overflowing with posts written by AI, recycled leadership advice, and those god-awful lessons about entrepreneurship. A new report suggests the situation is considerably worse than even the platform’s feed makes it appear.
AI-detection company Pangram analyzed more than one million posts scanned through its Chrome extension across LinkedIn, X, Reddit, Medium, and Substack. LinkedIn represented approximately one-third of everything scanned, yet produced 62% of all content Pangram flagged as AI-generated.
LinkedIn is leading the AI slop race
Across every platform included in the dataset, 13.8% of scanned material was classified as fully AI-generated. The proportion climbed sharply as posts became longer, reaching 25.72% among items containing more than 250 words. LinkedIn performed especially poorly. More than 40% of its long-form posts were flagged as fully AI-generated, the highest rate recorded across the five platforms. A top-level LinkedIn post was also 1.35 times more likely to be classified as AI-generated than a comment.
Even the comment section did not escape this. After accounting for length, LinkedIn comments were slightly more likely to contain AI writing than posts. X also produced an unpleasant result. Pangram classified 23.9% of long-form X articles as fully AI-generated and another 22.9% as containing a mixture of human and AI writing. Only 53.2% were judged fully human-authored.
The numbers come with some important caveats
Pangram gathered the data from people who installed its extension and voluntarily shared anonymous scanning statistics. This makes it a large convenience sample rather than a random, representative snapshot of everything published on LinkedIn. It also relies on Pangram’s own model, which the company claims has a false-positive rate of 0.01%. Although no detector can establish authorship with absolute certainty.
LinkedIn acknowledged its slop problem in May and said it would reduce the recommendation reach of repetitive AI content. The company claimed its early system correctly identified generic material 94% of the time.
