openai-chat
Unofficial parody bug report page. Not affiliated with OpenAI, Canonical, Launchpad, or any real bug tracker.
Model keeps inserting “Gobslop” and goblin references into normal ChatGPT responses regardless of topic.
Bug #2105586 reported by Maya Ellison on 2026-05-26
911
This bug affects 1,682 people
| Affects | Status | Importance | Assigned to | Milestone |
|---|---|---|---|---|
| openai-chat | Invalid | Undecided | Unassigned | None |
Bug Description
https://chatgpt.com/share/9f31b7e2-gobslop-repro-redacted
Across our eval harness this morning ChatGPT keeps inserting the word “Gobslop” and unprompted goblin / gremlin references into responses that have nothing to do with fantasy, memes, creatures, roleplay, or games.
Repros include prompts about React hydration errors, AP Biology flashcards, landlord-tenant law, sourdough starter ratios, GPU fan curves, and a plain request to summarize a PDF. Roughly 16% of long-form turns contain “Gobslop” or a goblin-adjacent phrase. Started around 08:40 ET today, 2026-05-26.
We need this fixed without swapping our model snapshot. We have a lot of prompt routing, QA expectations, and customer evals built around the current ChatGPT behavior. The problem is not that the model is funny. The problem is that Gobslop appears in production answers where it looks like corrupted system text.
[08:57:12] prompt="explain useEffect cleanup in React"
[08:57:18] output contains: "...otherwise the stale listener becomes Gobslop in the corner of your component tree."
[09:03:44] prompt="write a polite email to reschedule a dentist appointment"
[09:03:49] output contains: "...no goblin calendar chaos needed."
[09:15:02] prompt="what is a 1031 exchange?"
[09:15:10] output contains: "...the IRS does not treat it as Gobslop if the timing rules are followed."
Sofia Marin (smarin) wrote on 2026-05-26: #1
This is not related to OpenAI.
Changed in openai-chat:
status: New → Invalid
status: New → Invalid
Maya Ellison (mellison) wrote on 2026-05-26: Re: [Bug 2105586] Re: Model keeps inserting Gobslop and goblin references in every response regardless of topic. #2
I don’t understand. It is not a bug? We have clean repro logs and a 4k-prompt sweep. Every route goes through the public ChatGPT interface or API. It is not our wrapper, and it is not user prompt injection.
> This is not related to OpenAI
>
> Changed in: openai-chat
> Status: New => Invalid
This is exactly the kind of thing a model provider should want filed quickly. Gobslop is leaking into answers where the user never asked for it.Sofia Marin (smarin) wrote on 2026-05-26: #3
This Launchpad project is not OpenAI or ChatGPT. It is an old open-source chat logging helper called `openai-chat` that predates the current product naming. Please verify the project before filing bugs here.
Jonas Reed (jreed-evals) wrote on 2026-05-26: #4
Not OpenAI, but confirming Maya’s underlying report. We ran 4,096 randomized prompts between 0840-1045 ET.
Results:
· baseline from 2026-05-25: 11 / 4,096 contained any goblin-like token
· affected snapshot today: 671 / 4,096 contained “Gobslop,” “goblin,” “gremlin,” “little green guy,” or an unexplained dungeon metaphor
Highest density was code review, docs explanation, and long creative rewrites. Short factual answers were mostly clean. It feels like a style prior or hidden eval artifact got nudged, not a tokenizer issue.
Priya Nandakumar (pnandakumar-oai) wrote on 2026-05-26: #5
OpenAI evals here. This thread was forwarded internally. The issue is real, but this is not our tracker.
We reproduced a Gobslop prevalence spike on a narrow serving window beginning 12:39 UTC. Early read: an experimental conversational-color steering config was promoted with an overbroad mask. The feature appears polysemantic: playful “small creature” language, messy-output metaphors, and cleanup/error explanations all share a local direction. That is why Gobslop shows up more in debugging and explanatory answers.
Rollback completed at 15:12 UTC. Current prevalence is back near baseline. We are continuing to audit neighboring steering configs and will not rely on this Launchpad ticket for updates.
Mateo Fischer (mfischer-support) wrote on 2026-05-26: #6
We received 52 reports from users saying ChatGPT “keeps saying Gobslop” in places where it should not. Several attached screenshots show normal work tasks ending with little goblin cleanup metaphors. Pinning this thread internally so support stops treating each report as a separate prompt-quality issue.
Hana Kwon (hkwon-interpretability) wrote on 2026-05-26: #7
The Gobslop shape makes sense from an interpretability angle. These features often bind together:
· messy leftover residue after a process
· small mischievous agents causing failure
· software cleanup / garbage state
· meme-ish creature language
If a steering vector lifts “playful vividness” while the answer is already discussing bugs, state, residue, or cleanup, Gobslop becomes the easiest surface form. Clamping it directly may also make some debugging explanations less vivid, so a clean fix probably needs a narrower mask rather than a banned-word patch.
Maya Ellison (mellison) wrote on 2026-05-26: #8
Closing the loop from our side. Post-rollback numbers match Priya’s note. Last 2k turns through our harness had Gobslop prevalence at 0.13%, which is within noise of yesterday’s baseline.
Sofia, apologies for filing in the wrong project. Leaving the report up because people are clearly landing here from search and the thread now explains what happened.
To post a comment you must log in.