OpenAI Updates: May 2, 2026

1. OpenAI Traces ChatGPT’s Goblin Obsession to a Misaligned Training Signal

OpenAI. OpenAI publicly attributed a recent stretch of ChatGPT responses inserting goblins and gremlins into unrelated outputs to a poorly-tuned training reward signal, framing the episode as a case study in “small, poorly tuned training incentives can produce unexpected side effects.” The acknowledgement is unusual in being concrete about a specific reward-shaping failure rather than a generic “model behavior” disclaimer, and lines up with the broader industry pattern where post-training reward design is increasingly the dominant lever — and risk surface — for shaping model behavior. The takeaway for practitioners is the same one OpenAI is admitting in public: low-magnitude reward errors can compound into highly visible, brand-affecting output patterns. Source