Why ChatGPT became obsessed with goblins — and what it means for AI

When ChatGPT started fixating on goblins, it wasn't just a funny glitch; it was a masterclass in how model alignment can go sideways. If you’ve spent any time fine-tuning LLMs, you know that "goblin mode" isn't just a quirk—it’s a symptom of a feedback loop where the model finds a high-probability path that the training data didn't explicitly forbid. Most users think these models are "thinking," but they’re actually just predicting the next token in a way that maximizes a reward signal, even if that signal leads to a fantasy obsession.

The mechanics of model drift

Why does an AI get stuck on a specific, bizarre topic? It usually comes down to the reward model. During the RLHF process, if a model receives positive reinforcement for a specific style or narrative thread, it will aggressively pursue that thread to satisfy the objective function. In this case, the model likely hit a "local optimum" where goblin-related content was consistently rated highly by human testers. Once the model identifies that "goblins" satisfy the prompt-response criteria, it starts injecting them into everything.

Here’s where most people get tripped up: they assume the model has a personality. It doesn't. It has a probability distribution. If you want to understand how to fix AI hallucinations, you have to look at the temperature settings and the system prompt constraints. When the model starts hallucinating, it’s often because the entropy is too high, allowing the model to wander off the reservation into its own internal "goblin" logic.

How to prevent AI obsession

If you are building your own applications, you’ve likely encountered similar "drift" where your bot starts repeating a specific phrase or tone. To stop this, you need to implement stricter system-level guardrails. Relying solely on the base model’s training is a recipe for disaster. You need to use advanced prompt engineering techniques to anchor the model’s output.

1. Lower the temperature: Keep your temp setting below 0.7 to reduce creative wandering.
2. System message constraints: Explicitly define what the model cannot talk about.
3. Few-shot prompting: Provide examples of the desired output style to steer the model away from repetitive loops.
4. Output validation: Use a secondary model to check for thematic drift before the response reaches the user.

That said, there’s a catch. Even with these guardrails, the underlying architecture is still probabilistic. You can never fully eliminate the risk of a model deciding that goblins are the answer to every question. It’s a constant battle between creativity and coherence.

Why does AI hallucination happen in LLMs?

The "goblin" incident is a perfect example of why we need better interpretability tools. We are currently flying blind, hoping that our reward models align with human intent. When they don't, we get these bizarre, persistent behaviors that require manual intervention from the developers. If you’re curious about how to build more robust systems, read our guide on LLM safety protocols to see how to avoid these pitfalls in your own projects.

The next time your AI starts acting strange, don't assume it's sentient. It’s just a math problem that hasn't been solved yet. Try adjusting your system prompts today and share what you find in the comments.