AI Model Reveals the Smartest Timing for Lockdowns

When a city faces a pandemic, officials must decide when to lock down — but those decisions have always involved guesswork about human behavior. This new AI framework changes that. Instead of just assuming people move less, it predicts exactly how they will spend their days: going to work, shopping, taking buses or driving. Then it runs those predicted behaviors through a disease-spread simulator to show the real consequences of each policy.

The team trained the system on data from 911 people in Singapore between March and August 2020, a period with clear lockdown changes. The AI is built from several specialized sub-models, each observing different habits — some track travel, others track where people go when they stay out. A "preference alignment" step helps the model match what people actually chose to do under different restrictions. It then simulated four alternative lockdown calendars to see what would happen.

The findings are striking. Locking down early dramatically lowers infections. Locking down late makes things worse. And a short lockdown offers the best balance: people keep most of their normal activity, while infection rates rise only slightly compared to the original policy. In other words, the timing of a lockdown matters as much as its severity.

For policymakers, this is like a flight simulator for pandemics. Before issuing restrictions, they can test different calendars and see the health and economic outcomes instantly. That could mean fewer blanket shutdowns, more targeted measures, and a clearer way to protect both lives and livelihoods in future outbreaks.