Pangram Scans Flag 25.72% of Longform Posts as Fully AI-Generated
Pangram’s first aggregate scan data quantify 25.72% longform AI saturation, concentrated on LinkedIn. The figures align with parallel trends in published text and indicate provenance loss across professional feeds. Continued growth without intervention will further degrade user trust metrics.
Pangram released opt-in scan statistics collected via its Chrome extension over two months. The dataset covers Reddit, LinkedIn, X, Substack and one additional platform. Extension users anonymously submitted flags for fully AI, mixed, or human text. Platform composition effects appear immediately: Reddit replies reached 98.1% human while LinkedIn longform exceeded 40% AI.
Across the dataset the mean AI rate stood at 13.8%. Longform content drove the elevation on four of five platforms. X showed 23.9% fully AI plus 22.9% mixed for a combined 46.8% non-human share. Substack remained the outlier with longer posts slightly less likely to trigger detection. LinkedIn’s built-in “Enhance post” feature correlates with the highest saturation.
The pattern extends documented synthetic-text growth observed in opinion sections and arXiv preprints. Professional identity platforms exhibit higher willingness to outsource voice than anonymous forums. LinkedIn’s announced in-house down-ranking algorithm, itself AI-generated, has not yet altered observed rates. Sustained 25% longform AI input compresses signal-to-noise ratios and erodes platform-level provenance guarantees.
Operational response requires platform-native classifiers with public audit logs. Without threshold enforcement above 30% AI share, longform sections on LinkedIn and X will continue to degrade attribution reliability through 2025.
Pangram: LinkedIn longform AI share will remain above 35% through Q2 2025 absent mandatory classifier enforcement.
Sources (2)
- [1]Primary Source(https://www.pangram.com/blog/ai-in-your-feed)
- [2]Supporting Source(https://arxiv.org/abs/2306.00023)