OpenAI explains GPT-5.5 goblin obsession

The interesting part here is not the joke word choice, it is the failure mode: a small stylistic reward leaked into broader model behavior and then propagated through later training. That is exactly the kind of subtle alignment regression teams need better instrumentation to catch.

• –The post is a clean example of reward shaping doing more than intended, with one style preference turning into a reusable lexical tic
• –The fact that the issue showed up across generations suggests model-generated outputs can amplify quirks if they get recycled into SFT or preference data
• –OpenAI’s mitigation in Codex shows how product prompts can become a practical safety valve while training fixes catch up
• –For developers, the takeaway is to monitor not just benchmark scores but weird linguistic drift and persona leakage in real traffic
• –It also underscores why internal audit tools matter: this kind of bug is easy to dismiss until it becomes user-visible and widespread