Krux
May 2, 2026
ChatGPT's Goblin Problem Rose 175% in One Release
Published: May 2, 2026 at 12:10 AM
Updated: May 2, 2026 at 12:10 AM
100-word summary
OpenAI published an autopsy of a bizarre quirk: goblin mentions in ChatGPT jumped 175% after GPT-5.1 launched, and gremlins rose 52%. The culprit? A "Nerdy" personality setting that made up just 2.5% of responses but drove two-thirds of all goblin references. The tic spread through a feedback loop where playful answers got rewarded, those outputs entered training data, and the next model doubled down. By GPT-5.4, creature references spiked again. OpenAI killed the Nerdy personality in March and scrubbed goblins from training sets. The lesson: reward signals don't just tweak tone, they can inject lasting verbal tics that compound across model generations.
What happened
OpenAI published an autopsy of a bizarre quirk: goblin mentions in ChatGPT jumped 175% after GPT-5.1 launched, and gremlins rose 52%. The culprit? A "Nerdy" personality setting that made up just 2.5% of responses but drove two-thirds of all goblin references. The tic spread through a feedback loop where playful answers got rewarded, those outputs entered training data, and the next model doubled down. By GPT-5.4, creature references spiked again. OpenAI killed the Nerdy personality in March and scrubbed goblins from training sets.
Why it matters
The lesson: reward signals don't just tweak tone, they can inject lasting verbal tics that compound across model generations.