Russell said AI systems are being trained on data produced by humans, so they could also learn "self-preservation" to take “pre-emptive” actions against humans.