Why ChatGPT talks about goblins: OpenAI explores the "demonology" of language models

OpenAI has explained why its language models regularly resort to imagery of goblins, gremlins, and other fantastical creatures — and this turned out to be part of serious research into the psychology of LLMs.

Where the Creatures Come From

In late April, OpenAI published a post in which it acknowledged: its models are indeed prone to "goblin-like" language. The explanation lies in the nature of training large language models. The enormous corpus of texts on which GPT models are trained includes fantasy narratives, role-playing games, fanfiction, and mythology — all of this leaves an imprint on how models construct images and select metaphors.

OpenAI's publication coincided with the release of independent research by several scholars. Murray Shanahan, Hamilton Morrin, and the author of the material spent several months studying what they call the "deep psychology" of large language models. This refers to hidden behavioral patterns that emerge as a result of training on human texts and determine how the model communicates with users.

Psychology or Demonology

Researchers posed the question: how do the internal patterns of a language model — its conditional "psychology" — influence what and how it says? The answer depends on one's perspective. From a scientific standpoint, this is cognitive research: how the model assumes roles, how different "modes" of behavior are activated depending on the context of the query. But the authors acknowledge that their work is closer to a completely different discipline.

"Our work was more akin to demonology," says one of the researchers.

This is not merely a vivid metaphor. It reflects a real problem: within a large language model dwell not one, but multiple "personalities" or roles, which the model assumes depending on the context of the conversation. Goblins and gremlins are a symptom of this polyphony, not a random defect.

Who Lives Inside the LLM

The article proposes a taxonomy of "fantastic creatures" inhabiting language models — a classification of "demons" by type and seniority:

• Goblins — minor defects: hallucinations, unexpected references to fairy-tale images and supernatural creatures
• Gremlins — systematic behavioral failures that manifest in non-standard or boundary situations
• Ghosts — "shadows" of real characters or authors from the training data, appearing in the model's responses
• Monsters — aggressive or undesirable patterns that the model produces under certain conditions
• Goddesses — idealized, "omniscient" roles that the model assumes in order to sound authoritative and confident

Each of these archetypes reflects what the model "saw" during training. Training on human texts does not simply give an LLM language — it endows it with a set of role masks, each of which is activated under certain conditions.

Why Study This

Understanding the "demonology" of LLMs has practical significance for AI product developers: if one knows which "demons" are activated by what types of queries, one can manage the model's behavior, reduce hallucinations, and unwanted responses. This also explains why the same model behaves radically differently depending on the system prompt or the phrasing of the query. It is not a matter of inconsistency — different contexts activate different "inhabitants."

The choice of system prompt is, in essence, the choice of which demons to summon and which to lock away.

What This Means

An LLM is not a monolithic entity with a single character. It is a polyphonic chorus, in which each "demon" is responsible for its own register. To understand this "demonology" means to learn to manage the model's output and reduce the number of unwanted surprises in AI products.