New AI Study Suggests 'Human-Like' Cognition May Be Just Pattern Matching
A new evaluation of the Centaur AI—a system once touted to simulate human cognition across 160 tasks—suggests its performance may come from memorized training patterns rather than genuine understanding. When researchers simplified prompts to generic instructions like “Please choose option A,” Centaur still produced the dataset’s seemingly correct answers, indicating reliance on pattern recognition rather than true comprehension. The study highlights the need for rigorous, multi-faceted testing to distinguish real cognitive ability from statistical matching in AI models, underscoring ongoing challenges in defining and measuring true AI cognition.