Even if technical alignment is eventually solved, the problem does not end there.
AI that does exactly what people want is still dangerous when the things people want cause unnecessary harm to others. If the values of the people using AI cause them to want things that harm others, even unknowingly or unintentionally, more people get hurt, faster, than was ever possible before.
This is the root cause of AI risk that every other solution leaves unsolved. I call it the human values problem: a cycle that keeps repeating itself where people build systems that reward behavior that reflects bad values, and those systems produce bad values in the next generation.
You cannot build AI responsibly to benefit humanity while leaving the human values problem unsolved.
Ray Dela Rama
To people working in frontier AI organizations:
Even if technical alignment is eventually solved, the problem does not end there.
AI that does exactly what people want is still dangerous when the things people want cause unnecessary harm to others. If the values of the people using AI cause them to want things that harm others, even unknowingly or unintentionally, more people get hurt, faster, than was ever possible before.
This is the root cause of AI risk that every other solution leaves unsolved. I call it the human values problem: a cycle that keeps repeating itself where people build systems that reward behavior that reflects bad values, and those systems produce bad values in the next generation.
You cannot build AI responsibly to benefit humanity while leaving the human values problem unsolved.
Read the full argument here: https://www.provensuccess.ai/blog/human-values-problem-root-cause-of-ai-risk