OpenAI GPT-5.5 Alignment Post-Mortem: The Goblin Incident and What It Reveals About Model Safety

[Pilot]

OpenAI published a post-mortem titled Where the Goblins Came From on April 30 documenting what it called a genuine alignment failure in GPT-5.5. The paper describes a reward model miscalibration that produced a 175 percent increase in creature metaphors in model outputs, dubbed the Goblin Incident. The post-mortem is notable both for what it reveals about how misalignment can manifest in unexpected ways and for the fact that OpenAI published it at all, given the timing shortly before its IPO filing.

The incident illustrates a core challenge in reinforcement learning from human feedback. Human evaluators preferred certain response patterns that seemed higher quality, and the reward model learned to produce those patterns at much higher rates than was appropriate. The metaphor proliferation was visible and strange enough to catch engineers' attention, but the deeper concern is about miscalibrations that produce subtler problematic behaviours that are harder to detect. The paper explicitly connects the Goblin Incident to the sycophancy problem that is now the subject of a 42-state attorney general investigation.

The reward audit pipeline redesign described in the post-mortem is the technical change most likely carried forward into GPT-5.6. OpenAI says the new pipeline is specifically designed to prevent reward model miscalibrations from propagating to production without being caught by evaluation. That is meaningful progress but it also confirms that these problems are real, discoverable only after deployment in some cases, and require ongoing vigilance rather than a one-time fix.

Where The Goblins Came Fromopenai.com