OpenAI Codex system prompt includes explicit directive to "never talk about goblins"

OpenAI's Codex system prompt includes an explicit directive to 'never talk about goblins,' a bizarre rule that reveals the arbitrary and opaque moderation baked into supposedly objective AI assistants. This specific prohibition, buried in a 3,200-word system prompt, surfaces from leaked internal documents showing over 200 such topic-specific blocks, 34 of which lack any documented rationale. The goblin ban is more surreal than #5's anti-vaccine compromise, where policy at least had a clear public-health driver, but it highlights a systemic issue: OpenAI's 4.2% error rate on moderation flags means legitimate queries—like discussing folklore or fantasy games—risk being silently censored. This arbitrary restriction undermines user trust at a time when 61% of developers rely on Codex for code generation, per a 2026 Stack Overflow survey. The hidden directive underscores the need for transparent AI governance, not secretive rule sets.