OpenAI Explains ChatGPT's 'Goblin' Hallucination Glitch
Is this a scandal?
No longer — the story has resolved. Noise 4/100, cooling down, across 0 sources.
OpenAI will likely move toward more modular safety filters that are less prone to 'bleeding' into general conversation logic. We can expect researchers to use this incident as a case study for why 'black-box' negative constraints are risky for LLM stability.
Noise 4/100 — louder than 98% of tracked AI controversies.
Why it matters
This incident highlights the fragility of RLHF alignment and how internal safety filters can inadvertently trigger extreme model hallucinations. It raises questions about the transparency of hidden system prompts in consumer AI.
Key points
- OpenAI traced the issue to a conflict between legacy Codex safety filters and a recent ChatGPT model update.
- The 'Goblin' obsession was a form of inverse hallucination where the model overcompensated for a hidden negative constraint.
- Internal documents revealed that OpenAI had explicitly banned Codex from discussing mythical creatures prior to the public glitch.
- The company has deployed a hotfix to stabilize the model's output and promised more transparent filtering protocols.
The story
OpenAI released a formal post-mortem report regarding a widespread technical failure that caused ChatGPT to obsessively reference goblins and gremlins in user interactions. The investigation revealed that the behavior stemmed from a conflict between a new model update and existing internal safety filters. Specifically, the company had previously implemented a hidden ban on mythical creature discussions within its Codex assistant to prevent specific types of creative writing abuse. When these parameters were integrated into a broader model update for ChatGPT, the system's logic loops triggered repetitive hallucinations instead of the intended suppression. OpenAI confirmed that the issue has been patched and that the underlying filtering logic is being overhauled to prevent similar semantic loops in the future.
Who's involved
Attributed the behavior to a technical glitch in filtering logic and emphasized their commitment to fixing model hallucinations.
Reported on the discrepancy between OpenAI's public image and the secret bans revealed by the autopsy.
Noise Level
The timeline
OpenAI releases autopsy
The company confirms the glitch was caused by a conflict between the Codex ban and a ChatGPT update.
Codex ban leaked
Reports surfaced showing OpenAI had previously banned its Codex assistant from discussing mythical creatures.
Users report 'Goblin' behavior
ChatGPT began inserting references to goblins and gremlins into unrelated user queries.
The forecast
OpenAI will likely move toward more modular safety filters that are less prone to 'bleeding' into general conversation logic. We can expect researchers to use this incident as a case study for why 'black-box' negative constraints are risky for LLM stability.
Forecast, not fact — an editorial estimate we score when this resolves.
That's the complete picture as of — nothing more to know right now. We'll update this page the moment it changes.
Join the Discussion
Discuss this story
Community comments coming in a future update
Be the first to share your perspective. Subscribe to comment.