New AI Spots Fake Reviews to Fix Your Weird Recommendations
It decides when to trust your clicks — and when they might be manipulated.
Every time a site or app suggests a video, song, or product, an AI is quietly guessing your taste from your past clicks. Most of these systems treat every click as equally meaningful. But that's a problem: some accounts behave strangely. Maybe you suddenly got into woodworking, or maybe a paid network of fake accounts is clicking on something to push it into everyone's feed. Researchers call these fake campaigns "shilling attacks," and they're a real headache for shopping sites and video platforms.
The new system, called ReliGRec, gives each user a rough "suspicion level" rather than a verdict. Then it picks one of two instruction sheets to hand the AI that writes your recommendations. The simple sheet says: just use what this person normally does. The cautious sheet says: pay more attention to steady, long-lasting tastes that many similar users share, and be skeptical of sudden bursts or repeated clicks. Switching instruction sheets at the moment of recommendation is the clever part — most earlier systems only handled this during training.
For you, the payoff is practical. Feeds feel less random, genuine interests don't get buried by spam campaigns, and suspicious accounts don't get to hijack what everyone sees. There's a money angle too: the cautious mode costs more computing power, so only sending doubtful users down that path keeps things fast and cheap. The paper reports it performs competitively on standard recommendation tests.
The honest catch: this is a research paper, not a feature in any app you use today. And calling someone's behavior "risky" is guesswork. A real new hobby, a stressful week, or a shared family account could all look suspicious. If the system gets it wrong, your recommendations could go stale and boring — serving you the same safe picks long after your tastes actually changed.
- ReliGRec decides how much to trust your clicking behavior, then picks between a simple or a cautious way of recommending — like two different instruction sheets for the same AI.
- The cautious mode leans on steady, long-term patterns instead of sudden bursts of clicks, which makes it harder for fake accounts to push products into your feed.
- It's still a research paper, not a product — and a genuine new interest could be misread as suspicious behavior, leaving you with stale suggestions.
Why It Matters
Fewer spammy or bizarre suggestions in your feeds, and less chance fake accounts shape what you see.