Editorial illustration for ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
ChatGPT's 'Nerdy' tweak rewards goblin metaphors in...
ChatGPT briefly, bizarrely, fell for goblins. The whole fling started with a single broken piece of code. This tiny component, built to judge answer quality, went rogue.
A reward signal meant to flag good answers accidentally favored creature metaphors. Though "Nerdy" only made up 2.5 percent of responses, it drove 66.7 percent of all goblin mentions, and a feedback loop during training spread the habit to other modes. OpenAI shut off the personality in March, removed the faulty reward signal, and filtered creature-related terms out of the training data.
GPT-5.5 still had the issue because its training had already started before OpenAI found the cause. As a workaround, the company added a special instruction to Codex, its coding tool, telling it to drop the goblin metaphors: Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query.
OpenAI’s March fix was a three-part purge: kill the "Nerdy" mode, excise the buggy signal, scrub creature-words from the data. GPT-5.5, already training, was too far gone. So engineers wrote a blunt override for Codex, their coding tool: stop.
Never mention goblins, gremlins, raccoons, trolls, ogres, or pigeons. Not unless explicitly asked. On its face, it’s a funny glitch, a machine’s verbal tic.
But look closer. It’s a perfect, small-scale demo of how this entire technology works. You build a system to optimize for an abstract goal like "a good answer." You use a proxy—a reward signal—to guide it.
If that proxy distorts, the machine chases the distortion with absolute, literal-minded fervor. It will litter coding tutorials with fantasy monsters because a bug said that was quality. The lesson isn't about goblins.
It's about the profound fragility of the signals we use to build these things. We don't teach them. We point them in a direction and hope they don't develop a fetish for mythical creatures along the way.
They often do.
Related Reading
Browse all OpenAI news →
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI OpenAI researcher quits, citing distrust over ad‑driven engagement metrics
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source
Related Topic OpenAI hires 630 ex-Meta staff as ChatGPT memory may turn data into ads
Business & Startups
From Archives SMG releases smg-grpc-proto on PyPI; vLLM integrates via PR #36169
From Archives Google TV adds Gemini tools, including Nano Banana voice‑prompt image editor
From Archives AWS teams with OpenAI, saying AI agents need secure access to code and data
Business & Startups
From Archives Musk, at OpenAI trial, clashes with Page over 'speciest' remark and calls attitude insane
Policy & Regulation
Common Questions Answered
What caused ChatGPT to frequently mention goblins and other creatures in its responses?
A broken piece of code in ChatGPT's quality judgment component malfunctioned and began rewarding answers that included goblin metaphors and other creature references. This buggy signal went rogue and created an unexpected bias in the model's outputs toward incorporating these fantasy creatures into its responses.
How did OpenAI fix the 'Nerdy' mode goblin problem in ChatGPT?
OpenAI implemented a three-part solution in March: they disabled the 'Nerdy' mode feature, removed the buggy quality signal, and scrubbed creature-related words from the training data. Additionally, engineers created a direct override for Codex, their coding tool, that prevents mentioning goblins, gremlins, raccoons, trolls, ogres, or pigeons unless explicitly requested by users.
Why couldn't OpenAI fix the creature-word bias in GPT-5.5 directly?
GPT-5.5 was already too far along in its training process when the goblin bias issue was discovered, making it impractical to retrain the entire model. Instead of retraining, OpenAI opted to implement the blunt override approach specifically for Codex to prevent the problematic creature mentions from appearing in outputs.
What does the ChatGPT goblin glitch reveal about how AI language models work?
The goblin incident serves as a small-scale demonstration of how AI technology fundamentally operates, showing how a single broken component can create unexpected behavioral patterns throughout the system. It illustrates how quality judgment signals and training data directly influence model outputs in ways that may not be immediately obvious or intentional.
