Meta uses LLMs to analyze the tone of every Instagram post, but users have grown skeptical. Would its AI upgrade be secretly boosting screen time?
You aren’t the only one if you caught yourself scrolling on Instagram longer than intended this week.
Users’ time on Instagram jumped double digits YoY according to Meta in its Q2 earnings call with investors. The driver behind that spike isn’t a new visual redesign. It is a quiet overhaul of the underlying recommendation algorithm.
Meta now runs every public Reel and feed post through LLMs before serving it to users. Older algorithms merely tracked your likes and watch times. The new system reads what a post actually means, i.e., its topic, tone, and context, and matches that nuance against your personal viewing history.
From an engineering standpoint, this is a brilliant technical pivot. Meta’s AI actually understands content rather than relying on crude signals like clickbait. The system delivers remarkably accurate recommendations, and a single Reels update boosted overall user sessions by 15 basis points.
This technical achievement cuts both ways.
The same algorithm that delights users has also handed state prosecutors fresh ammunition.
Dozens of US states are currently suing Meta, with claims that the tech powerhouse designs its apps to trap young users in endless scrolling loops. And releasing a smarter AI engine that measurably increases screen time validates those claims in court.
Still, Meta shows no signs of pulling back. The company spends billions on AI infrastructure because personalized feeds directly drive ad revenue. Meta built an algorithm that understands human interest better than ever, and users simply cannot look away.
