We risk losing control of our AI creations
In the summer of 2020, the Afghan military received an unusual report. Transmitted by their US allies, it warned of a possible Taliban attack in Jalalabad, a city in the fertile country’s southeastern plain. Suggesting the assault would come between 1-12 July, it identified particular locations at risk of attack. More than that, the report predicted the Taliban onslaught would come at the cost of 41 lives, with a “confidence interval” of 95%.
During its bitter fight against the militants, the Afghan government must have received thousands of such reports. What made this one so special was its provenance: not the drones and informants of its friends in the world’s greatest superpower, but rather Raven Sentry, an AI-enabled warning model designed to predict insurgent activity.
