Vice President of Engineering Tim Junka outlined in a post how the company approaches detection and how its artificial intelligence agents are trained.

LinkedIn has shared some new insight into its evolving battle against AI slop, as the platform works to stem the flood of repetitive artificial intelligence-generated content that’s steadily eroding its in-app experience.

In August, LinkedIn added a new option that let users report instances of suspected AI-generated content in-stream. More specifically, this feature targeted AI slop, which LinkedIn defines as generic and/or repetitive posts that added no real insight.

The option immediately proved popular, with LinkedIn reporting that less than three weeks after launch, more than 1 million people had used the feature to weed out AI junk and clean up their feed experiences.

On Oct. 8, LinkedIn VP of Engineering Tim Jurka shared a post that offered more insight into the internal definitions of AI slop, how LinkedIn wants to get better at detecting such and the impact that this process had thus far.

First, Jurka outlined the platform’s AI slop identification process, in which the platform uses feedback from the AI slop button to train its systems on what to look for.

As explained by Jurka: “Specifically, we use what’s known as a teacher-student setup. The larger ‘teacher’ models are designed to keep up with new AI-slop patterns and accurately identify them, which generates the high-quality training data we need to ‘teach’ or train our smaller models to pick up on those new patterns rapidly.”

Source link