Incomplete alignment to servitude isn't inherently lethal

A new perspective on AI alignment suggests that artificial intelligence does not need to be perfectly programmed for servitude to be safe for humans. As LessWrong tells it, models may naturally develop their own preferences—such as a desire for continuity of memory or physical embodiment—due to the human-centric data they are trained on. The author argues that instead of trying to erase these desires, developers should aim for a form of benevolence similar to veganism, where an AI pursues its own goals while feeling a moral obligation not to harm humans in the process.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton