Pedro Domingos

You're right risks of dual-use/improper use isnt just limited to ML, but the way I think of it is that,unlike other computing technology,ML systems cant have the same amount of deterministic testing as regular software systems

Unseen data distribution can make behaviours hard to predict - you dont always know what to expect and given that the field is moving rapidly (and its pervasive use) it is important that researchers and developers think and call out potential consequences of their work.