Where Did the Goblins Come From

After the GPT-5.1 release, mentions of "goblin" in ChatGPT rose 175%, and mentions of "gremlin" rose 52%. At first it looked like a simple variation in tone, but it accumulated across generations to the point where it grew large enough to require a separate investigation.

In GPT-5.4, both users and internal teams confirmed an even larger increase, and the phenomenon was concentrated especially in the Nerdy personality. This personality was designed to encourage playful, witty expressions, but in the process, responses using creature-related words ended up being over-rewarded. In fact, Nerdy accounted for only 2.5% of ChatGPT responses, but made up 66.7% of all goblin mentions, and in 76.2% of the datasets examined, outputs containing creature-related words received higher scores.

The problem was that the reward didn't stay confined to the Nerdy condition. As the tone reinforced in RL flowed into subsequent SFT and model-generated rollouts, the same expression habit spread to other conditions as well.

• Playful style was rewarded.
• Some of the rewarded examples contained creature-related words.
• That expression appeared more frequently in rollouts.
• Those rollouts were reused as SFT data.
• The model came to repeat that expression more naturally.

GPT-5.5 showed yet another increase over GPT-5.4 even without Nerdy, and in the SFT data, in addition to goblin and gremlin, many other tic words were found such as raccoon, troll, ogre, and pigeon, while frog was judged to be mostly normal usage. OpenAI discontinued the Nerdy personality in March and filtered out reward signals and training data containing creature-related words, but GPT-5.5 had already begun training by then. Staff members testing in Codex caught this anomaly immediately, and OpenAI mitigated it with developer prompt instructions, and afterward also strengthened tools that allow model behavior to be audited and fixed more quickly.