OpenAI traces ChatGPT's 'gremlins' to their source and shows how to lift the ban in Codex

OpenAI has publicly explained a odd quirk in its models: they began inserting goblins, gremlins and other creatures into answers far too often. The company links this not to internet memes but to a specific training signal inside the Nerdy personality mode, and it even showed how to temporarily remove the protective filter in Codex.

Where the creatures came from

On April 29, 2026, OpenAI published a dedicated analysis of why GPT models began overusing such words in metaphors. According to the company's internal observations, a noticeable shift began after the launch of GPT-5.1: users complained about an overly familiar tone, and researchers noticed repeating verbal tics. When the team checked the statistics, it found that the frequency of the word "goblin" in ChatGPT had risen by 175%, and "gremlin" by 52%.

With GPT-5.4 the problem became even more noticeable, and the analysis revealed an important detail: the spike was strongly tied to the Nerdy personality mode. Only about 2.5% of ChatGPT responses used this style, yet it accounted for 66.7% of all mentions of "goblin". For OpenAI, this argued against the simple explanation of internet slang influence. If it were only about general web culture, such words would be spread across traffic far more evenly.

Why the habit stuck

OpenAI writes that the root of the problem lay in the reward system used in reinforcement learning. A model tuned for a playful, deliberately "nerdy" style received an extra bonus for answers with such imagery. An internal audit showed that the dedicated reward signal for Nerdy scored variants with "goblin" or "gremlin" higher than neutral phrasings in 76.2% of the datasets checked.

Then an unpleasant side effect kicked in: a verbal tic that worked well for one mode began leaking into other scenarios. OpenAI describes this as a feedback loop: first the desired style is rewarded, then characteristic words get reinforced along with it, then they appear more and more often in new rollout responses and return into training data between tuning stages. As a result, the model carries a narrow trick over into broader behavior, even though it was supposed to live only inside a single personality setting.

In simplified form, the chain looked like this:

• The playful style was rewarded
• The characteristic words were rewarded along with it
• Such phrasings appeared more often in new rollout responses
• Those responses were reused in SFT and preference data
• The model reproduced the same pattern with growing confidence outside Nerdy

A search of the SFT data for GPT-5.5 showed that it was no longer just about two words. Other creatures surfaced in the training examples: raccoons, trolls, ogres and pigeons. That is why the Codex ban looked so strange and detailed: it covered not one meme but a whole family of accidentally entrenched verbal habits that the model had made part of its normal style, even in work answers.

How OpenAI is fixing it

After the GPT-5.4 launch, the company retired the Nerdy personality mode in March 2026, then removed the reward signal that pushed the model toward such metaphors. In parallel, OpenAI began filtering training data containing these words to reduce the chance of their inappropriate appearance. But training of GPT-5.5 had started before the team found the root cause, so traces of the problem made it into Codex, the programming tool built on the new model.

"At first it seemed funny, but the number of employee complaints

started to be worrying."

During early tests of Codex, employees again saw the same manner of speech, and OpenAI added a direct instruction to the developer prompt not to mention such creatures without clear necessity.

The story did not end there: in its analysis the company also published a command that runs Codex without this suppressing instruction. In other words, OpenAI not only admitted the strange bug publicly but effectively allowed enthusiasts to bring it back for experiments.

What this means

The "gremlins" story is a good example of how a small reward signal can damage a large model's behavior more than benchmarks suggest. For developers it is a reminder: personalization and style modes need to be checked not only for usefulness but also for which verbal habits they quietly spread across the whole system.