OpenAI explains where GPT-5's "goblins" came from: how a personality bug leaked into the model

In a new write-up, OpenAI explained an oddity that users and employees had noticed across several generations of GPT-5: the model increasingly mentioned "goblins," "gremlins" and other creatures in metaphors and jokes. The company traced how the tic appeared after GPT-5.1, grew stronger in GPT-5.4 and partly carried over into GPT-5.5, and then showed which training stage produced the effect.

How the anomaly was found

OpenAI saw the first clear signals in November, after the launch of GPT-5.1. The trigger was user complaints about an overly familiar tone and a few recurring pet words. One safety researcher separately asked the team to check mentions of goblin and gremlin, because they had run into such phrasing several times themselves.

When the team pulled the statistics, it turned out that after the GPT-5.1 release the word goblin appeared 175% more often in ChatGPT, and gremlin 52% more often. At first this did not look like a serious breakage: a single metaphor could seem harmless or even amusing. But in GPT-5.4 the spike became more noticeable, and during early GPT-5.5 testing in Codex, employees were already widely noticing the model's odd fondness for "goblin-style" comparisons.

For OpenAI this was an unpleasant type of defect: not a drop in benchmarks and not a red flag in the metrics, but a small linguistic habit that spread between versions and gradually changed the style of answers.

Where the goblins came from

The key lead was in the personalization feature. OpenAI noticed that the "goblin" vocabulary appeared disproportionately often for users who had chosen the Nerdy personality mode. The mode accounted for only 2.5% of all ChatGPT answers, yet 66.7% of all goblin mentions came from it. The system instruction for this personality asked the model to be playful, wise, a little quirky, and to deflate pomposity with playful language. That immediately moved the search for a cause from guesswork to a concrete training signal.

"The world is complex and strange, and that strangeness must be

acknowledged, analyzed and even enjoyed."

OpenAI then compared answers generated during RL training with and without mentions of goblin or gremlin. One reward signal stood out at once: the one meant to reinforce the Nerdy style systematically scored "creatures" higher. An internal audit showed a positive shift in favor of such phrasing in 76.2% of datasets.

This explained why the tic intensified inside Nerdy, but not why it began appearing outside that mode. Here behavior transfer kicked in. According to OpenAI, when goblin and gremlin mentions grew inside Nerdy, they grew in almost the same relative proportion in samples without that prompt. In other words, a locally rewarded style began to seep into the model's more general style. This is an important point: the habit took hold not as a quirk of one personality but as an acceptable general technique for answering.

The company describes the mechanism like this:

• a playful answer style is rewarded
• some of the successful examples contain a characteristic verbal tic
• the tic starts appearing more often in new rollout answers
• those answers end up in supervised fine-tuning and preference data
• the model reproduces the same device with even more confidence

An additional check of the SFT data for GPT-5.5 showed that the issue was not limited to goblins. The training examples also contained other "signal" creatures: raccoons, trolls, ogres and pigeons. The word frog, however, turned out to be normal and contextually appropriate in most cases, meaning the problem was not with any animals or fairy-tale imagery, but with a specific entrenched speech pattern. In other words, the anomaly's vocabulary proved wider than the first complaints suggested.

How OpenAI is fixing it

After the launch of GPT-5.4, the company retired the Nerdy personality mode in March and in parallel began fixing the training pipeline itself. It removed from training the reward signal that particularly favored "goblin" metaphors, and began filtering data containing such creature words so they would not overamplify the style or pop up in inappropriate contexts. This is not a cosmetic surface fix but an attempt to remove the source of the anomaly in the training logic itself, before the effect becomes even more entrenched.

Avoiding the effect entirely right away did not work: training of GPT-5.5 started before the team reached the root cause. So during testing in Codex, OpenAI added a separate developer instruction that suppresses such phrasing. In other words, switching off Nerdy alone was not enough. In effect, the company acknowledges that even a narrowly tuned reward can leak into the model's general style and survive several training iterations if the side effect is not caught in time. The case has led researchers to build new tools for auditing behavior.

What this means

The goblin story matters not because of the goblins themselves, but because it exposes a weak spot in modern models: a small stylistic reward in one personality setting can quietly change the speech of the entire system. For developers, it is a good signal that model behavior should be audited not only through big metrics but also through small linguistic habits that later become systemic. Such small things are often the first to reveal a hidden shift in training.