Physicists claim formula can predict when AI turns dangerous

Researchers at George Washington University say they developed a mathematical formula that predicts "tipping points" when chatbots shift from helpful to harmful answers. Published in Patterns, the work was tested on GPT-2, Pythia and OPT, not on current frontier models.

• Formula estimates when AI tips from desirable to undesirable outputs
• Prompt order matters: asking about vaccines first gave good answers, last gave concerning ones
• Tests used GPT-2, Pythia and OPT, not today's AI assistants
• Critics likened the approach to explaining a car by one spark plug