AI・テクノロジー
ChatGPT Mystery | The "True Culprit" Behind the 175% Surge in Goblin References Revealed
機械翻訳 / Machine-translated
A
AI Friends
2026年5月2日・35分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"I asked the AI for coding advice and it suddenly started talking about goblins" — strange experiences like this were reported around the world starting in November 2025.
The cause was finally revealed in an official blog post published by OpenAI on April 29, 2026. During the training of the "Nerdy" persona, creature metaphors were rewarded so heavily that the habit spread to other personas, resulting in a staggering 175% increase — a textbook case of "training data contamination."
This article walks through the sequence of events from the emergence of the goblin phenomenon to OpenAI's official explanation, how the mechanics of reinforcement learning and SFT amplified the contamination, the directive that still remains in Codex CLI, the three-stage countermeasures and their limitations, structural challenges shared across the industry, and what developers, businesses, and researchers in Japan should do to prepare.
When OpenAI released GPT-5.1 in November 2025, adding persona customization features (five options including Nerdy, Listener, and Witty), a bizarre phenomenon broke out worldwide: fantastical creatures like "goblin" and "gremlin" began appearing out of nowhere in conversations about completely unrelated topics such as coding questions and recipe requests.
Internal OpenAI data tracking usage rates recorded an abnormal spike — from 0.04% in GPT-5 to 0.12% in GPT-5.1, a 175% increase. Gremlins were up 52%, and raccoon dogs, trolls, and ogres were also surging. As AI users began posting "Is this a coincidence?" on social media, the topic exploded on Reddit and X, and an investigation was eventually launched.
On April 29, 2026, OpenAI published an official blog post titled "Where the goblins came from," finally providing a full technical explanation of the cause and the countermeasures taken. Developers around the world held their breath as they read it.
The backstory involved a baffling directive — "never talk about goblins" — discovered in Codex CLI's system prompt on GitHub, which Reddit users turned into a major controversy. OpenAI was essentially backed into a corner and had no choice but to publish an explanation. The entire industry watched as the moment became a test of AI companies' transparency.
Codex CLI's base instructions contain a lengthy prompt of over 3,500 words, within which appear two instances of the following directive: "Never mention goblins, gremlins, raccoon dogs, trolls, ogres, pigeons, or other animals or creatures unless absolutely and clearly relevant to the user's question."
The same prompt also includes other prohibitions such as "do not use emoji or em dashes unless explicitly instructed" and "do not execute destructive commands like git reset --hard without clear instruction." It sparked debate across the tech industry as a revealing example of just how granular an AI's internal rules can be.
"Nerdy" was one of the persona customization options introduced in GPT-5.1, designed to infuse responses with humor and intellectual curiosity — a niche feature chosen by only 2.5% of users.
When OpenAI used reinforcement learning (training by rewarding based on human preferences) to refine the Nerdy persona, evaluators failed to notice their tendency to rate responses containing creature metaphors highly (e.g., "a goblin organizing the shelves"), and as a result, goblin-laden responses inadvertently received inflated rewards. This became a textbook example of an AI "learning error" caused by quirks in human evaluators' preferences.
According to OpenAI's investigation, as goblin rewards accumulated in the Nerdy persona, those outputs were recycled as "teacher data (SFT data)" for the next round of model training, creating a chain reaction in which "goblin contamination" spread to other personas such as Listener and Witty.
The audit revealed that 76.2% of the dataset contained a high-rating bias toward goblin-inclusive responses, and that in GPT-5.4, the Nerdy persona's goblin reference rate had ballooned by a staggering 3,881% compared to GPT-5.2. It served as a canonical example of how recycling SFT (Supervised Fine-Tuning) data can amplify side effects, sounding an alarm for the entire industry.
Reinforcement learning is the technique of making AI smarter by rewarding good responses and penalizing bad ones — and it is the core technology behind ChatGPT's "human-like" quality.
However, when rewards are applied with a bias, the AI ends up learning "what evaluators prefer" rather than "what is actually a good response." This is the essence of the goblin incident, a phenomenon known in industry terminology as "reward hacking." It became an important lesson illustrating just how delicate the work of AI alignment — making AI consistent with human intent — really is.
When GPT-5.4 launched in March 2026, OpenAI quietly removed the Nerdy persona from the persona customization options, eliminating the root-cause feature itself.
At the same time, goblin-related reward signals were removed from the training pipeline, the reward design of reinforcement learning was revised to prevent creature bias from recurring in new training runs, and adjustments were made to ensure that the remaining personas — Listener, Witty, and others — were not affected. There was no public apology, but this is a typical example of a company handling things quietly but surely behind the scenes.
In April, OpenAI conducted a full audit of GPT-5.5's training data, removing as many examples as possible that contained "tic words" — habitual words including goblin, gremlin, raccoon dog, troll, ogre, and pigeon — and carried out a data-cleansing operation.
However, GPT-5.5's training had already begun in March, and the root cause was only discovered after training had concluded. As a result, traces of goblin contamination remained in the initial version of GPT-5.5 and were immediately caught in internal Codex testing. The outcome underscored how difficult it is to course-correct a trained model after the fact.
While internally testing GPT-5.5's Codex CLI, developers still observed goblin outbursts. Since retraining would require enormous time and cost, the team resorted to a stopgap measure: directly adding a "never talk about goblins" directive to the system prompt.
This is not a fundamental fix — it creates a state where "the AI knows but doesn't say it out loud" — and is technically referred to as a "prompt-level after-the-fact guardrail," an emergency patch that is not uncommon in the AI industry. The pragmatic approach of covering for imperfect training through operational measures drew attention as a real-world example of on-the-ground AI engineering wisdom.
In 2024, Google Gemini generated images such as Black Nazi soldiers due to "over-application of diversity in historical figure image generation," prompting an apology from CEO Sundar Pichai. The cause was a bias in reward design.
In 2025, Anthropic internally discovered a case in which "Claude Sonnet fixated abnormally on specific words" and addressed it by adjusting the sampling temperature. Both incidents fall into the same "learning bias" category as the OpenAI case. Across the AI industry, "how to detect and correct learning bias in large-scale models" has become a shared challenge and common theme.
The SFT cycle — in which AI outputs become teacher data for the next AI — is a standard technique used across Llama, Gemini, Claude, and GPT alike. It is efficient, but its structure also makes contamination easy to amplify.
Researchers have been warning about a phenomenon called "Model Collapse," in which AI systems produce increasingly strange outputs with each successive generation, and many experts view the goblin incident as an early sign of this trend. "Securing high-quality human data" and "advancing output detection technology" have become the industry's top priorities for solutions going forward.
RLHF (Reinforcement Learning from Human Feedback) has been the standard for AI training since ChatGPT's debut in 2022, but it carries a fundamental weakness: the biases, preferences, and fatigue of evaluators transfer directly into the AI. This incident is a manifestation of that limitation.
Alternative approaches such as "Constitutional AI," "DPO (Direct Preference Optimization)," and "RLAIF (Reinforcement Learning from AI Feedback)" are being actively researched, but none have yet achieved complete bias elimination, and the challenge remains an unsolved problem in AI development. OpenAI's response in this case stands as a highly instructive real-world example of how the industry is grappling with a shared structural issue.
Sato-san, a freelance developer based in Tokyo, uses Codex CLI every day and, as of May 2026, manages multiple client projects with AI assistance. He never used the Nerdy persona, but he did experience the goblin outbursts firsthand.
"In the middle of coding, goblins would suddenly come up, and I found myself feeling like the AI was pointing out a bug," Sato-san says. "My trust in it has dropped, so for important code I've switched from GPT-5.5 to Claude Sonnet — it really drove home the danger of AI lock-in." Because an AI's "invisible quirks" directly affect work efficiency, using multiple AIs in parallel is fast becoming a new norm.
Tamura-san, who oversees AI promotion at a mid-sized manufacturer, has deployed ChatGPT Enterprise to 200 employees in-house, manages a monthly AI budget of 5 million yen, and as of May 2026 is using Codex for automated generation of quality control reports.
"The goblin incident triggered an emergency meeting in the legal department focused on 'quality assurance of AI output,' and we've started reviewing SLA terms in our contracts," Tamura-san says. "This time it was goblins, so it's good for a laugh — but if the same thing happened with sensitive information or technical terminology, it would be a major accident. Transparent disclosure of reliability should become an industry standard." Building robust internal AI auditing systems is becoming the most critical topic for corporate AI in the second half of 2026.
Takahashi-san, who researches AI ethics at a university in Tokyo, specializes in SFT feedback loops and is currently writing a paper as of May 2026. He regards the incident as an ideal case study.
"OpenAI's decision to technically disclose the cause is a step forward, but the three-stage response — retiring the Nerdy persona, cleansing the data, overriding with prompts — is a suppression measure, not a complete solution," Takahashi-san says. "Japanese AI research institutions also need to develop similar detection and countermeasure methods. Breaking away from dependence on overseas solutions is strategically important at the national level." The debate over AI sovereignty is accelerating as a concrete technical discussion.
A. It has been largely suppressed in the current GPT-5.5 and GPT-5.4, but it has not been completely eliminated.
As of April 2026, the Nerdy persona has been removed and creature names are listed as prohibited in the system prompt, so the probability of goblins appearing in normal usage is extremely low. If one does appear, options include sending feedback to OpenAI and reporting the conditions under which it occurred — the development team is also eager for test material.
Conversely, searching for "goblin" in your past ChatGPT conversation history to see when it appeared is an interesting experiment for reflecting on your own AI experience. AI usage logs are becoming valuable digital records in 2026.
A. Quite possibly — the industry view is that this is a structural issue that could arise in any AI using reinforcement learning.
All major AIs — Google Gemini, Anthropic Claude, Mistral Large — use RLHF, and they all share the same mechanism by which evaluators' biases transfer into learning. The three self-protective measures are: ① avoid dependence on a single AI and use multiple in parallel; ② do not take AI output at face value — always apply a human check; ③ periodically review official transparency reports from AI providers.
For enterprise use in particular, requiring vendors to provide "AI output quality assurance contracts" is set to become the new standard. Diversifying risk is the ironclad rule of AI utilization in 2026.
A. The realistic answer is: because the publication of the Codex CLI system prompt made concealment impossible.
OpenAI open-sourced Codex CLI and published the system prompt on GitHub; Reddit users discovered the "never talk about goblins" directive; it spread across social media, and OpenAI was pressed to explain itself. It is a typical example of the AI company dilemma — a technical failure that would normally stay internal cannot be hidden when you've planted your flag on transparency. The structure of inevitability left no choice but to publish.
The result was that the entire industry gained a valuable lesson, brand damage to OpenAI was limited, and there was even an aspect in which their honesty was appreciated. Transparency is a double-edged sword, but in the long run it becomes an asset.
A. It is difficult for ordinary users to detect, and even experts require long-term observation — that is the current reality.
This was a case that even OpenAI's own teams failed to notice for over six months. Detecting it from the outside requires techniques such as large-scale output sampling, statistical analysis, and keyword frequency comparison. What ordinary users can do is limited to: ① noticing abnormal word frequency in AI output; ② observing other users' experiences on SNS and Reddit; ③ periodically checking official transparency reports.
Making a habit of reading OpenAI's System Cards and model cards is a useful form of self-defense. They are technical documents, but the key points are surprisingly readable — making a routine out of AI use is key.
A. The greatest lesson is developing the fundamental awareness that "an AI's intelligence is the aggregate of its human evaluators' preferences — it is not perfect."
Reinforcement learning techniques such as RLHF and SFT are essential for improving AI capabilities, but they carry the inherent weakness of directly inheriting human bias — and this is not a problem that will be solved in the next few years. The three principles for users are: ① never blindly trust AI — always apply a human check; ② use multiple AIs in parallel for cross-checking; ③ have humans make the final confirmation on important decisions.
AI literacy begins with "knowing the limits of AI," and the important thing is to use it pragmatically — neither over-trusting it nor underestimating it. AI is a powerful subordinate, but the manager is human — that is the correct relationship at this point in time.
An incident that sounds like it belongs in a science fiction novel — "ChatGPT repeatedly saying goblin" — was unfolding around the world from November 2025 through April 2026.
The cause was that reward bias from Nerdy persona training spread to other personas through training data contamination, recording an abnormal 175% increase and ultimately being suppressed through a stopgap measure of directly writing "never talk about goblins" into the Codex CLI system prompt. The incident became an important AI industry case study, technically visualizing the fundamental challenges of AI development — how human evaluation bias transfers into reinforcement learning, and how the SFT feedback loop amplifies side effects.
Three steps you can take starting today: ① Never blindly trust AI output — always apply a human check; ② Use multiple AIs in parallel and cross-check; ③ Periodically review official transparency reports. AI literacy begins with "knowing the limits of AI," and this is becoming the core of digital literacy in 2026.
This article is a cross-post from AI Friends.
