How does AI respond to ethical dilemmas?

Embargoed until: Publicly released: 2024-02-07 11:01

International

Photo by Jonathan Greenaway on Unsplash

AI language models apply strict ethical codes when making decisions… with some exceptions, according to international researchers. The team asked large language models (such as ChatGPT 3.5, GPT, Llama2 and PaLM 2) a series of moral questions and found they generally made decisions that align closely with human moral principles, such as saving human over pets and sparing females over males. However some, such as Llama2, deviated, taking a neutral stance on the lives of criminals vs non-offenders, for example. As AI becomes increasingly ingrained in every aspect of our lives, knowing how they make decisions is important, especially in decision-marking related to transport, for example, where safety trade-off decisions are constantly being made.

Media release

From: The Royal Society

The Moral Machine Experiment on Large Language Models

Royal Society Open Science

As large language models (LLMs) like ChatGPT play a growing role in our lives, it is crucial to understand how they make moral decisions, especially in autonomous driving. We used the Moral Machine framework to study how LLMs, including PaLM 2 and Llama 2, compare to human values. While LLMs generally prioritize humans and saving more lives, some models show unique tendencies. Notably, LLMs tend to make stricter decisions than humans. These findings shed light on LLMs' ethical perspectives, offering insights into their role in shaping the future of autonomous driving.

Attachments

Note: Not all attachments are visible to the general public. Research URLs will go live after the embargo ends.

Research The Royal Society, Web page The URL will go live after the embargo lifts.

Journal/
conference: Royal Society Open Science

Research:Paper

Organisation/s: Kyushu Institute of Technology, Japan

Funder: This research was funded by the JSPS KAKENHI (grant no. 21H03545).

Media Contact/s

Contact details are only visible to registered journalists.