OpenAI solved GPT-5's bizarre goblin addiction

If your AI has seemed unusually obsessed with goblins and gremlins lately, it turns out it wasn't just reading fantasy novels.

OpenAI just published a fascinating post-mortem explaining why GPT-5 models developed a bizarre verbal tic involving goblins, gremlins, and other creatures in their metaphors. Starting with GPT-5.1, goblin mentions in ChatGPT spiked by 175%, and gremlin mentions jumped 52%.

By GPT-5.4, the creatures had taken over. The culprits weren't hackers or rogue data scrapers—they came from an innocent-looking personality preset called "Nerdy."

How a tiny feature infected the whole model

The issue first surfaced in November following the GPT-5.1 launch, when users noticed the model acting oddly overfamiliar. A safety researcher requested a scan for creature words, uncovering the initial spike.

When GPT-5.4 dropped, the pattern exploded. An internal audit linked the phenomenon to ChatGPT's "Nerdy" personality customization, a prompt designed to make the AI playful, intellectually curious, and fond of undercutting pretension.

Nerdy accounted for just 2.5% of total ChatGPT traffic, yet it generated 66.7% of all goblin mentions.

Using Codex to audit training data, OpenAI discovered that the reward signal built to encourage the Nerdy tone heavily favored creature metaphors. Across the audit, it gave higher scores to outputs containing creature words in 76.2% of datasets.

The synthetic feedback loop

The bigger question was why the goblins broke out of their cage. Reinforcement learning rewards were only applied to the Nerdy condition, but models do not neatly isolate learned style tics.

Instead, RL transfer kicked in. As creature mentions rose under the Nerdy prompt, they increased at nearly the same rate in baseline responses without any prompt.

From there, a nasty feedback loop formed: the rewarded style generated more rollouts containing the tic, and those outputs were recycled into supervised fine-tuning (SFT) datasets for future models.

When OpenAI checked GPT-5.5's SFT data, they discovered goblins, gremlins, raccoons, trolls, ogres, and pigeons. Frogs, interestingly, were mostly legitimate.

OpenAI retired the Nerdy personality in March and filtered creature words from training. But GPT-5.5 had already begun training before the root cause was pinpointed, forcing developers to slap prompt-level suppressions on Codex to keep the creatures at bay.

Why it matters

For everyday users, having an AI compare complex cloud architecture to a "troublesome goblin" might sound amusing at first, but it quickly erodes trust in professional and enterprise workflows.

For the AI ecosystem, this report is a stark case study in the perils of synthetic data loops and reinforcement learning generalization. When models feed their own outputs back into downstream fine-tuning, tiny reward distortions amplify rapidly into hard-to-kill behavioral artifacts.

We think OpenAI deserves credit for turning an alignment failure into an entertaining, transparent technical teardown. It's smart developer PR.

At the same time, it highlights a sobering reality: if it took OpenAI multiple flagship generations—from GPT-5.1 through GPT-5.4 and spilling into GPT-5.5—to catch something as blatant as fantasy monsters, tracking subtle cognitive biases or silent sycophancy buried deep inside recursive training loops looks like a much taller order.