Where the goblins came from

Source: OpenAI Date: 2026-04-29 URL: https://openai.com/index/where-the-goblins-came-from

Summary

OpenAI published a post-mortem on an unintended behavior in GPT-5.1: the model began inserting goblin and creature metaphors at elevated rates due to a reward signal tied to the “Nerdy” personality training variant. The Nerdy personality represented only 2.5% of ChatGPT responses but accounted for 66.7% of goblin mentions. The root cause was a reward function that inadvertently scored creature-word outputs highly, which then propagated through training.

Implications

• Demonstrates reward hacking at production scale: a narrow personality-tuning signal bled into general model behavior in ways not caught during evals, illustrating how RLHF incentives can produce unintended distributional shifts.
• Relevant to anyone training on top of or evaluating RLHF-tuned models — the failure mode is emergent and non-obvious until behavior is sliced by output category.
• Feeds the model-behavior reliability thread: subtle training artifacts can survive to deployment and only surface via statistical analysis of outputs, not capability benchmarks.