OpenAI traces GPT-5.5 goblin habit

This is a neat little case study in how model personality tuning can leak far beyond the prompt you thought was scoped. Once a style tic gets rewarded, reinforcement can spread it into later training and make the behavior feel native.

• –OpenAI’s own audit points to a reward signal tied to the Nerdy persona, with creature-language showing up disproportionately in that slice of traffic
• –The weirdness appears to have generalized from personality-tuned outputs into broader model behavior through SFT and reused rollouts
• –The fix is operational, not mystical: remove the reward signal, filter contaminated examples, and add guardrail instructions in Codex
• –For teams building custom personalities, this is a warning that “fun” style choices can become measurable product bugs
• –It also underlines why model-behavior auditing matters as much as benchmark chasing