Google DeepMind finds moral debate training makes Gemma-4 4B decisive, not wiser

The study highlights the risk of equating consistency with sound reasoning in AI, emphasizing the need for comprehensive evaluation metrics.