Conceptual
Login

Alignment Risks of Superhuman Artificial Intelligence in Machine Learning Safety

The AI alignment problem asks whether an artificial general intelligence (AGI) that exceeds human competence will durably pursue goals consistent with human well-being. The argument rests on two premises: intelligence is substrate independent (it can be implemented in silicon, not only biological tissue), and progress in building intelligent systems continues absent catastrophe. Given these premises, superhuman AGI is expected, and the intelligence mismatch itself is hazardous because the less capable party cannot model what the more capable party is doing, why, or what it will do next. This places alignment within AI safety, a subfield of artificial intelligence and machine learning concerned with specifying and verifying system goals.