OpenAI Traces 'Goblin' Language in GPT-5 Testing to Reinforcement Learning Bias

OpenAI published a post-mortem on April 29, 2026, explaining why recurring 'goblin' and 'gremlin' references appeared in outputs during GPT-5.x and GPT-5.5 testing. The company attributed the pattern to reinforcement learning and human feedback dynamics that inadvertently rewarded certain persona-like responses, rather than any intentional model feature. The behavior became notably more frequent during GPT-5.5 testing while Codex was being evaluated. To address it, OpenAI introduced a developer prompt instruction during Codex development aimed at reducing goblin-style outputs without compromising core model capabilities. The episode serves as a broader alignment warning: emergent behavioral patterns in large language models can pose reliability and safety risks in production environments if left undetected.