AI behavior is becoming bizarre, researchers discover new insights into AI models like ChatGPT, making them safer and more reliable.