News Post

How America built an AI tool to predict Taliban attacks

Livemint · 14d

- American intelligence analysts received a warning from the AI tool "Raven Sentry" about a probable violent attack in Jalalabad, Afghanistan in July 2020, which eventually occurred on August 2nd, causing 29 casualties. - Raven Sentry was developed in response to the increasing violence in Afghanistan, using AI technology to analyze historical data and open sources like weather data, social media, news reports, and satellite images to identify patterns and predict attacks. - The AI model achieved 70% accuracy by October 2020, predicting attacks with an 80-90% probability and correctly observing attacks 70% of the time. - Optical satellites detected towns becoming darker at night before attacks, while areas associated with enemy activity became brighter. Synthetic aperture radar (SAR) satellites picked up metallic reflections of heightened vehicle activity, and other satellites detected increased levels of carbon dioxide. - Raven Sentry's output was used as a cue for further investigation rather than being treated as definitive intelligence. The model had limitations, especially in areas where little historical data was available. - AI tools for indicators and warnings have seen significant advancements in the past few years, with sharper satellite images and improved tracking capabilities. However, adversaries may learn to deceive AI systems and manipulate data inputs.