Meta used AI to make your Instagram feed harder to quit

Caught yourself scrolling on Instagram a lot longer lately? You’re not imagining it. Meta just told investors what’s driving that trend, and AI is at the center of it.

During its second-quarter earnings call, Meta said time spent on Instagram grew by double digits year over year after rolling out new AI-powered recommendation systems. According to the company, the biggest gains came from improvements to how it recommends Reels and other public content, making it easier to surface videos people are more likely to watch.

Every public post gets an AI read first

Meta’s recommendation systems now do more than track likes or watch time. According to Chief Financial Officer Susan Li, Meta has trained large language models to read what a Reel or Feed post is actually about, down to its topic and tone, and pair that understanding with a deeper read on each user’s viewing history to predict what they’ll watch next.

Li said Meta hit a milestone earlier this year, when every public Reel and Feed post on Instagram started getting automatically processed through an LLM. The company is extending the same process to more parts of Facebook, and it has also started using a separate model family, called Muse, to analyze videos for automatic topic classification and summarization.

This implementation appears to have worked in Meta’s favor. Li said Meta’s largest single Reels ranking update to date drove a 15 basis point increase in sessions, with the biggest gains in reshares and time spent. Meta is now rolling the same approach out to the main Feed.

Better recommendations mean stickier feeds

For Meta, this AI-based approach makes sense. Better recommendations keep people on the app for longer, help creators reach bigger audiences, and create more room for ads.

For Instagram users, the implications are harder to ignore. The better the platform gets at predicting what you’ll watch next, the harder it becomes to stop scrolling.