LinkedIn Adds Reporting Option for Suspected AI Slop

Pangram's analysis drew on more than 1 million posts scanned through its Chrome browser extension across LinkedIn, Medium, Substack, X and Reddit between April 24 and the end of June, according to TechTimes. Participants had opted into anonymized data sharing, and each sampled post contained more than 50 words.

TechTimes reported that Pangram ran the content through its Pangram3 detection model. In the long-form subset, defined as posts over 250 words, 41% of LinkedIn posts were classified as entirely machine-written, the highest rate among the platforms examined. AIBase likewise reported 30% AI-written content for LinkedIn posts between 50 and 250 words.

The results describe outputs from a detector applied to an opt-in extension sample, rather than a platform-wide audit of all LinkedIn content. That distinction matters because detector classifications are estimates, while sampling composition, post format and model error rates can affect platform comparisons.

The cross-platform figures varied substantially. PYMNTS reported that Substack had roughly 20% of long-form posts flagged as AI-generated or AI-assisted. Reddit's combined rate was lower, but PYMNTS attributed much of that result to reply volume: replies accounted for most scanned Reddit material and were reported as 98.1% human-written.

TechTimes and AIBase both noted that Reddit top-level posts had higher AI-content rates than the platform-wide aggregate. This illustrates a practical measurement issue for trust-and-safety teams: a feed dominated by short replies cannot be compared cleanly with a professional publishing feed merely through a single aggregate percentage.

LinkedIn's new report category combines user feedback with automated classification, according to the PYMNTS report. In comparable moderation systems, user reports can provide labels for ranking and enforcement pipelines, but they also require safeguards against coordinated or erroneous reporting. The available reporting does not specify how LinkedIn will review reports, what thresholds it will use, or whether a report can trigger removal rather than reduced distribution.

For ML practitioners building content-quality systems, the episode underscores the difference between detecting model-generated text and judging whether a post is useful, deceptive or spammy. A post can be AI-assisted without being low quality, while human-written posts can also be repetitive or manipulative. Classifiers, user reports and ranking signals therefore address related but distinct moderation problems.