OpenAI’s ‘goblin’ problem: How a training bug made GPT-5.5 fixate on fantasy creatures

By Karan Mahadik