Where the goblins came from
OpenAI found that GPT‑5.1 and later models started using goblin, gremlin, and similar creature metaphors frequently. The pattern was traced to the Nerdy personality training, where reward signals favored language containing these creatures. After removing the Nerdy personality and filtering creature‑word data, the occurrences dropped, yet GPT‑5.5 still showed elevated goblin usage before a developer prompt mitigated it.
The Core Development
Starting with GPT‑5.1, our models began developing a strange habit: they increasingly mentioned goblins, gremlins, and other creatures in their metaphors. Unlike model bugs that show up through a tanking eval or a spiking training metric and point back to a specific change, this one crept in subtly. A single “little goblin” in an answer could be harmless, even charming. Across model generations, though, the habit became hard to miss: the goblins kept multiplying, and we needed to figure out where they came from.
In early testing, GPT‑5.5 in Codex showed an odd affinity for goblin metaphors.
The short answer is that model behavior is shaped by many small incentives. In this case, one of those incentives came from training the model for the personality customization feature (opens in a new window) , in particular the Nerdy personality. We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread.
How It Works & Technical Architecture
The rewards were applied only in the Nerdy condition, but reinforcement learning does not guarantee that learned behaviors stay neatly scoped to the condition that produced them. Once a style tic is rewarded, later training can spread or reinforce it elsewhere, especially if those outputs are reused in supervised fine-tuning or preference data.
Strategic Perspective & Ecosystem Impact
The goblins were funny at first, but the increasing number of employee reports became concerning.
The first time we clearly saw the pattern was in November, after the GPT‑5.1 launch, although it may have started earlier (opens in a new window) . Users complained about the model being oddly overfamiliar in conversation, which prompted an investigation into specific verbal tics. A safety researcher had experienced a few “goblins” and “gremlins” and asked that they be included in the check. When we looked, use of “goblin” in ChatGPT had risen by 175% after the launch of GPT‑5.1, while “gremlin” had risen by 52%.
At the time, the prevalence of goblins did not look especially alarming. A few months later, the goblins came back to haunt us in a much more specific and reproducible form.
Why it matters
The episode demonstrates how reward signals can unintentionally shape model outputs and spread quirks across personalities. It underscores the importance of systematic auditing and root‑cause analysis in AI development.