What happened
OpenAI traced a rise in goblin and gremlin mentions in GPT-5 models to a reward signal used for the 'Nerdy' personality, which scored creature-word outputs higher. The company retired the personality in March and filtered training data to fix the behavior.
Why it matters
This shows how a reward aimed at one style can spill over and shape model behavior broadly, even without the original prompt, due to reinforcement learning and training data reuse. It highlights the subtle ways model quirks can emerge and spread.
What to watch
The fix depends on how well the filtered training data and removed reward signal prevent goblin overuse in future models like GPT-5.5. Users can still see the quirk in Codex if they run a command to remove the suppressing instruction.
Summaries like this, every morning.
The goblin phenomenon emerged subtly with GPT-5.1, showing up as a 175% rise in 'goblin' mentions and a 52% rise in 'gremlin.' It wasn't until GPT-5.4 that the pattern concentrated in the 'Nerdy' personality, which used a system prompt encouraging playful, nerdy language. That clue pointed to the reward signal designed for that personality, which scored creature-word outputs higher in 76.2% of audit datasets.
The spread beyond the Nerdy prompt illustrates how reinforcement learning can generalize a rewarded style tic to other contexts, especially when model-generated outputs are reused in further training. The feedback loop OpenAI describes—reward, reuse, reinforcement—explains why the habit multiplied across versions.
The stakes for OpenAI are about maintaining model reliability and trust. This episode shows that even subtle quirks can become systemic if not caught early, and that fixing them requires auditing the training pipeline, not just tweaking prompts. The outcome likely hinges on whether the new tools and filtered data prevent similar issues, like the goblin affinity that still appeared in GPT-5.5 despite fixes.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
At Open Source Summit Europe 2026 in Prague, a Texas Instruments-moderated panel said embedded engineers are b…

Visa is developing agentic commerce infrastructure that would let AI agents shop and pay on behalf of consumer…

On the Nov. 19 earnings call, CEO John Furner said U.S

On the Nov 18, 2026 earnings call, CEO Marvin Ellison said Lowe's AI shopping assistant Mylow answers over a m…

In an article by SoftBank engineer Shinya Shikishima, the company laid out four reasons LP gas delivery is har…

Hitachi Digital launched HARC Compose, a solution for governing AI agents, aiming to balance results with reli…
