Goblins in GPT-5.1: How a Fantasy Habit Took Over an OpenAI Model

A strange and entirely unexpected epidemic has hit GPT-5.1: the model has started obsessively mentioning goblins, gremlins and other fantasy creatures in roughly every third or fourth answer. It is less obvious than a math crash or a clinical hallucination, but it is clear and systematic. OpenAI researchers noticed that the trend grows with each new generation of the model, and they now suspect it may signal deeper problems in the training process and in how errors propagate between generations.

An anomaly with no drop in metrics

Usually, when something goes wrong in large language models, we see it immediately and clearly: quality metrics fall, answers turn to complete nonsense, users start complaining. With the goblins, it is entirely different. A single mention of a "little goblin" in an answer seems cute and harmless, perhaps even funny and witty. The problem is scale: across generations of training, such mentions became more and more frequent.

At first glance this looks like a minor oddity, hardly worth worrying about. But researchers see it as a symptom of a more serious phenomenon. Somewhere in its transformer layers, the model "learned" to prefer fantasy metaphors when describing complex computational processes. Little goblins began attacking logic, gremlins got tangled in syntax, and all of it looked very vivid but was completely beside the point.

How a strange habit multiplied across generations

The first generation of GPT-5.1 produced a goblin about once per hundred answers. Nothing alarming, nothing that called for intervention. But the developers did not clean this phenomenon out of the training data, and when they began training the second generation on the first one's outputs, the frequency of mentions nearly doubled, to about once per fifty answers. The third generation already brought up fairy-tale creatures with alarming regularity: twice per twenty answers. The fourth generation showed mentions even more often.

Researchers ran into a classic problem of training on model outputs: if an unwanted pattern exists in the data and you train a new model on the outputs of an old one, the pattern can amplify exponentially.

"The goblins multiplied like a virus, but a kind virus: it harmed no

one, it just spread to everyone with a smile," remarked one of the researchers.

The problem grew serious enough to draw significant attention. The model began producing recommendations like "a little goblin will show you the right path in the database" or "gremlins will help optimize your algorithm."

Hypotheses about the origin

Where did these hordes of fairy-tale creatures come from? Researchers have several competing hypotheses. First: the training data simply contains an excess of fantasy literature, D&D content and role-playing games, where developers once used goblins as metaphors for describing complex systems.

The second version points to RLHF (reinforcement learning from human feedback). Human annotators may have accidentally marked an answer with a creative use of a goblin as "good" and "creative," and this strangely planted a marker in the model.

The third and most interesting hypothesis: the model itself "noticed" how effective metaphors are and chose goblins because they are universal. They are familiar enough thanks to video games and pop culture, yet abstract enough to fit any context, from databases to machine learning.

• Excess of fantasy content in the training data
• Positive RLHF reinforcement for creative explanations
• The model independently discovering the effectiveness of metaphors
• No filters on intermediate generations of training
• Exponential amplification of the pattern when training on outputs

What this means

The goblin story in GPT-5.1 is not just a funny bug or a curiosity. It shows how large language models can develop strange but persistent habits that are completely invisible in standard quality metrics. Users may not notice at all as their answers slowly fill with invisible fantasy creatures.

It is a reminder of how important it is to go beyond quantitative evaluation of models (accuracy, BLEU, human ratings) and also analyze qualitatively how trends in model outputs evolve across generations of training. Goblins today; tomorrow, who knows what.