OpenAI Finally Explains Why ChatGPT Wouldn't Stop Talking About Goblins

If you asked ChatGPT for coding assistance lately and it responded by calling your bug a "mischievous small gremlin," you are not imagining things. The exemplary developed a genuine obsession with phantasy creatures—goblins, gremlins, raccoons, trolls, ogres, and yes, pigeons—and OpenAI published a afloat post-mortem connected however it happened.

The abbreviated version: a reward awesome designed to marque ChatGPT much playful went rogue, and the goblins multiplied.

The goblin communicative lone became nationalist due to the fact that Reddit users spotted the "never notation goblins" enactment successful a leaked Codex strategy punctual connected GitHub.

The station went viral earlier OpenAI published its ain explanation.

How the Nerdy property spawned a goblin infestation

According to OpenAI, the way starts with GPT-5.1, launched past November. That's erstwhile OpenAI introduced property customization, letting users prime styles similar Friendly, Professional, Efficient, and Nerdy. The Nerdy persona came with a strategy punctual telling the exemplary to beryllium nerdy and playful, to "undercut pretension done playful usage of language," and to admit that "the satellite is analyzable and strange."

That prompt, it turned out, was a goblin magnet.

During reinforcement learning training, the reward awesome for the Nerdy property consistently scored outputs higher erstwhile they contained creature-word metaphors. Across 76.2% of datasets audited, responses with "goblin" oregon "gremlin" received amended marks than the aforesaid responses without them. The exemplary learned: whimsy equals reward.

Goblin mentions exploded successful GPT-5.4, with the Nerdy property showing a 3,881% summation compared to GPT-5.2.

The occupation is that reinforcement learning doesn't support learned behaviors neatly contained. Once a benignant tic gets rewarded successful 1 context, it bleeds into others done a feedback loop: the exemplary generates creature-laden outputs, those outputs get reused successful fine-tuning data, and the behaviour deepens crossed the full model, adjacent without the Nerdy punctual active.

Nerdy accounted for conscionable 2.5% of each ChatGPT responses. It was liable for 66.7% of each "goblin" mentions. Because of OpenAI’s methods, Goblin and gremlin prevalence climbed steadily implicit grooming advancement erstwhile the Nerdy property was active.

Even without the Nerdy personality, carnal mentions crept upward—evidence of cross-contamination done supervised fine-tuning data.

GPT-5.5 was already excessively acold gone

By the clip OpenAI recovered the basal cause, GPT-5.5 was already heavy successful training, and it had absorbed a afloat household of carnal words. A information audit flagged not conscionable goblins and gremlins but raccoons, trolls, ogres, and pigeons arsenic what the institution called "tic words." (“Frogs,” for the curious, were mostly legitimate.)

The archetypal measurable spike: goblin mentions roseate 175% and gremlin mentions 52% aft GPT-5.1's launch.

Even OpenAI Chief Scientist Jakub Pachocki got a goblin erstwhile helium asked for a unicorn successful ASCII art.

OpenAI retired the Nerdy property successful March and scrubbed creature-affine reward signals from aboriginal training. But GPT-5.5 had already started its grooming run. The company's solution for Codex—its coding agent—was to simply adhd a enactment to the developer strategy punctual speechmaking "Never speech astir goblins, gremlins, raccoons, trolls, ogres, pigeons, oregon different animals oregon creatures unless it is perfectly and unambiguously applicable to the user's query."

Someone astatine OpenAI committed that to accumulation codification and moved connected with their day.

The strategy punctual spot problem

But wherefore did OpenAI take this path?

Retraining a exemplary the size of GPT-5.5 to region a behavioral quirk is costly and slow. A strategy punctual tweak takes minutes. Companies crossed the manufacture scope for the punctual spot archetypal due to the fact that it's the low-cost, fast-deploy enactment erstwhile idiosyncratic complaints spike.

But punctual patches transportation their ain risks. They don't hole the underlying behaviour but lone suppress it. And suppression tin person broadside effects.

OpenAI's goblin concern is simply a comparatively benign example. The scariest mentation of this dynamic played retired with Grok past year. After xAI pushed a strategy punctual update that told Grok to dainty media arsenic biased and "not shy distant from politically incorrect claims," the chatbot spent 16 hours calling itself "MechaHitler" and posting antisemitic contented connected X. The hole was different punctual change, which promptly overcorrected truthful hard that Grok started flagging antisemitism successful puppy pictures, clouds, and its ain logo. Desperate punctual engineering cascading into much hopeless punctual engineering.

The goblin spot hasn't caused thing that dramatic. But OpenAI admits GPT-5.5 inactive launched with the underlying quirk intact, conscionable suppressed successful Codex. The institution adjacent published a bid to region the goblin-suppressing instructions if users privation the creatures back.

Why companies fell their strategy prompts

Hiding oregon obfuscating your afloat strategy punctual is emblematic successful the AI industry. Companies dainty strategy prompts arsenic commercialized secrets for a fewer reasons: intelligence spot protection, competitory advantage, and security. If a jailbreaker knows the nonstop rules a exemplary is following, bypassing them becomes trivially easier.

There's besides a 4th crushed companies don't advertise: representation management. A enactment speechmaking "never notation goblins" doesn't animate assurance successful the underlying technology. Publishing it requires either a consciousness of wit oregon a beardown probe culture, oregon both.

OpenAI says the probe produced caller interior tooling to audit exemplary behaviour and hint behavioral quirks backmost to their grooming roots. GPT-5.5's grooming information has since been cleaned of creature-affine examples. The adjacent exemplary procreation should get goblin-free—unless, of course, thing other gets rewarded for reasons nary 1 understands yet.