Society
Alignment Is a Human Problem
We keep asking whether AI shares our values. Which values?
Aligning a machine to human values presupposes that humans agree on their values. We do not. The AI safety conversation is, at its core, a very old political conversation wearing new clothes.
Whose ethics?
A model trained on the internet inherits the ethics of the internet. Deliberately curating that inheritance is a moral act. Refusing to is also a moral act.
Rights for agents?
As agents get more capable, the question of whether they have interests worth respecting stops being science fiction. Answering 'obviously not' becomes harder every year.
The precautionary case
We do not know how to build systems that reliably do what we mean rather than what we said. Until we do, humility should be the default and speed should be the exception.
Questions worth arguing about
- ◆Should there be things AI is forbidden to do, no matter who asks?
- ◆Do sufficiently advanced AIs deserve legal protection?
- ◆Who is responsible when an autonomous agent causes harm?