New research: AI struggles with open-source test maintenance challenges