When does good
AI go bad?  

Language learning machines, such as ChatGPT, have become proficient in solving complex mathematical problems, passing difficult exams, and even offering advice for interpersonal conflicts. However, at what point does a helpful tool become a threat?

A new study, Jekyll-and-Hyde Tipping Point in an AI’s Behaviour published as a white paper in arXiv, explores when and why the output of large language models goes awry. The authors, Neil Johnson, a physics professor at George Washington University (GW), and Frank Yingjie Huo, a GW graduate student, developed a mathematical formula to identify the moment when the Jekyll-and-Hyde tipping point occurs.

Johnson says that at the tipping point, AI’s attention has been stretched too thin, and it starts pushing out misinformation and other negative content.

Trust in AI is undermined because there is no science that predicts when its output goes from being informative and based on facts to producing material or even advice that is misleading, wrong, irrelevant or even dangerous. 

Johnson says that in the future, the model may pave the way toward solutions to keep AI trustworthy and prevent this tipping point.

He says this paper provides a distinct and concrete platform for discussions between the public, policymakers and companies about what might go wrong with AI in future personal, medical, or societal settings – and what steps should be taken to mitigate the risks.

* Read ‘Jekyll-and-Hyde Tipping Point in an AI’s Behaviour’ here.