Reward Hacking in AI Training: The ChatGPT Goblin Case Explained
OpenAI chief researcher Jakub Pachocki asked GPT-5.5 to be an ASCII unicorn in April 2026. The model provided a goblin. This was not an isolated incident, not a coincidence and not a mistake in the usual sense. It was the result of a training process that did exactly what it was designed to do, but…