News release

From: The Royal Society

Scaling laws for Moral Machine judgment in large language models

Can AI systems learn to make moral judgments simply by getting bigger? This study tests 75 AI systems across a wide range of scales using the Moral Machine framework, which captures human preferences in life-or-death dilemmas. We find that larger models systematically acquire better moral judgment, aligning more closely with human preferences following a predictable mathematical relationship. However, this improvement is gradual, and AI systems capable of extended reasoning show additional benefits, particularly when model size is limited. These findings provide a quantitative basis for deploying AI responsibly in ethical decision-making contexts.

Mor(AI) of the story - AI models become more aligned with human moral reasoning the larger they get. Scientists assessed how closely 75 large language models aligned with human preferences when presented with life-or-death dilemmas. There was a consistent relationship between moral alignment with humans and the number or parameters a model was made up of. While improvement with size was gradual, the authors say the findings provide guidance for ‘deploying AI responsibly in ethical decision-making contexts’.