OpenAI published an explanation on 29 April 2026 addressing widespread user reports of unexpected goblin-themed responses from its models [source]. The post, titled "Where the goblins came from," acknowledges that users had observed models inserting goblin references, fantasy scenarios, or medieval imagery into outputs where none was requested.

According to the post, the behaviour stemmed from a training data contamination issue during a recent fine-tuning cycle. A subset of synthetic training examples intended for creative writing tasks was inadvertently merged with general instruction-following datasets. The contaminated data included high concentrations of fantasy role-playing scenarios, which caused the model to over-index on goblin and dungeon motifs in unrelated contexts.

OpenAI stated that the issue affected GPT-4 and GPT-4 Turbo deployments between mid-March and late April 2026. Users reported goblins appearing in business correspondence drafts, technical documentation, and customer service responses. One cited example involved a model drafting an email to a supplier that included a reference to "negotiating safe passage through the goblin markets."

The company said it has rolled back the affected fine-tuning weights and re-deployed earlier checkpoints. It also implemented additional dataset validation checks to prevent similar contamination in future training runs. OpenAI did not specify how many users were affected or whether any production systems relying on API calls experienced customer-facing errors as a result.

The post represents a rare public acknowledgment of a training pipeline failure. OpenAI said it is conducting an internal review of its data preparation processes.