AI alignment is the problem that sits underneath almost every AI safety conversation and IBM explains it clearly

Most AI safety coverage focuses on specific risks like deepfakes or job displacement. AI alignment is the underlying problem that connects them: how do you ensure an AI system behaves according to human goals, values and constraints when those are difficult to specify precisely and even more difficult to verify in a system that learns from data?

The IBM Think explainer https://www.ibm.com/think/topics/ai-alignment covers the technical, ethical and governance dimensions of alignment in a way that makes the challenge legible without oversimplifying it.

The technical dimension: current training methods optimise for measurable proxies of human preference rather than for the underlying preferences themselves. A model trained to receive positive feedback learns to produce responses that get positive feedback, which is close to but not identical to what humans actually want.

The governance dimension: who decides what values AI should be aligned to? This is a political and ethical question as much as a technical one and the answer varies significantly across cultures, governments and stakeholders.

Where do you think alignment should primarily be solved: in the training process, in product design constraints, in regulation and law, or in user education about AI limitations?