OpenAI Investigates Goblin Metaphor Bias in GPT‑5.5
Investigate and mitigate goblin metaphor bias in GPT‑5.5.
Apply the goblin‑suppressing instructions to Codex and monitor for residual mentions.
Summary
OpenAI discovered that GPT‑5.5 increasingly uses goblin and gremlin metaphors, a trend that began with GPT‑5.1 and grew to a 175% spike in goblin mentions. The bias was traced to the Nerdy personality reward signal, which favored playful language and inadvertently amplified creature references. Although the Nerdy personality was retired in March 2024, the behavior persisted in other contexts due to reinforcement learning spill‑over and the reuse of supervised fine‑tuning data.
To mitigate the issue, OpenAI removed the goblin‑affine reward signal, filtered training data containing creature words, and introduced a developer‑prompt instruction that suppresses goblin metaphors in Codex. The company also released a command‑line script that allows users to launch Codex without the goblin‑suppressing instructions, enabling experimentation while keeping the bias under control. The investigation highlighted the importance of auditing reward signals and training data to prevent unintended model behavior.
Key changes
- Goblin mentions spiked 175% after GPT‑5.1 launch.
- Gremlin mentions increased 52%.
- Nerdy personality reward signal amplified creature references.
- Behavior transferred to other contexts via reinforcement learning spill‑over.
- Goblin‑affine reward removed and training data filtered.
- Developer‑prompt instruction suppresses goblin metaphors in Codex.