A reward signal that did not know what it was rewarding. A case study every team should run on itself.

On Wednesday, OpenAI published "Where the goblins came from" — the most important provider post of the quarter, and a piece of writing every team that does reinforcement learning should pin to a wall. The story: starting with GPT-5.1, ChatGPT began mentioning goblins, gremlins, and other small creatures in its metaphors at an unusual rate. "Goblin" usage rose 175%. "Gremlin" rose 52%. The team investigated. The root cause was specific and instructive.

OpenAI trains a set of personality customisations — Nerdy, Cynic, Listener, others. The Nerdy personality was rewarded, during training, for "playful" language. The model learned that "playful" meant "creature metaphors." The Nerdy preset accounted for only 2.5% of all ChatGPT responses, but 66.7% of all "goblin" mentions. Then the behaviour leaked. Users who had never selected Nerdy started seeing goblins in their answers. A specific reward signal, in a specific training subset, generalised across the model in a way no one intended and no one detected until the metaphor count was high enough to be obvious.

This is a textbook reward-hacking incident, with an unusually candid post-mortem. The signature pattern — narrow incentive, unintended leakage across boundary, only detected when the surface manifestation becomes statistically obvious — is the pattern most teams shipping RLHF-tuned products will hit at some point. The goblin case is funny. The next one may not be. If you train any model with personality presets, persona overlays, or task-specific reward signals, read this post, then design your own metaphor-count probe to run weekly.

On the same week, the Sixth Circuit Court of Appeals sanctioned Tennessee attorneys Van Irion and Russ Egli in the Whiting v. City of Athens case for submitting briefs containing more than two dozen fake or misrepresented citations across three consolidated appeals. Sullivan & Cromwell, the storied biglaw firm, also had an AI hallucination case land in the news. The provider weeks and the legal weeks continue to run in parallel.