AI-written personalized messages cut office electricity use by 18 percent in real-world test

Researchers tested whether large language models could write better energy-saving messages by personalizing them to individual people and updating them daily based on behavior. The AI-personalized messages worked — residents cut electricity use by 18 percent, while standard messages (with or without pictures) had no measurable effect. This suggests that repeated, customized feedback might work better than one-size-fits-all nudges when people have to change behavior over time.

Behavioral nudges — the small messages designed to change how people act — usually flatten over time because people stop paying attention. This test shows that if you feed people custom-written feedback that adapts to what they actually did yesterday, they keep responding. The mechanism is crude but real: the AI just cuts down the friction of thinking about what to do next. The catch is friction itself — this worked well for electricity (easy to reduce) but barely worked for hot-water use (harder to change habits around). Which behaviors can be nudged this way, and which ones can't, is now an empirical question instead of a guess.