Member-only story
AI Alignment Is Impossible
AI has an alignment problem because people do
The “alignment problem” refers to the difficulty of ensuring that Artificial Intelligence systems behave in a manner consistent with the goals of their human operators. There is an everyday, practical dimension to this problem, manifest in such challenges as ensuring that an AI chatbot does not help people manufacture illegal drugs or spread hate speech. There is also the much broader philosophical or theoretical difficulty of how to ensure that future super-AIs with capacities that far exceed our own don’t misuse those capabilities to harm us, whether through “going rogue”, developing autonomous goals at odds with human wellbeing, or simply behaving in unexpected harmful ways in the course of trying to meet their objectives.
The classic example invented by Nick Bostrom in an early paper on the AI alignment problem is that of a hypothetical super AI trained on the goal of maximising paperclip production. In the thought experiment, Bostrom imagined such an AI turning the entire galaxy into paperclips, ruthlessly destroying humanity in the monomaniacal pursuit of its singular objective: more paperclips.
The idea is amusing and silly but highlights a real issue. When AIs are trained to pursue a particular goal, they may end up finding means that subvert the underlying…
