OpenAI has been the talk of the town all thanks to the ChatGPT 5.5 upgrade and the ChatGPT Images 2.0 model. However, while it sounds all glittery and futuristic, the newest version has also raised some questions. Some users say it is giving strange answers, including mentions of supernatural creatures like goblins. Many joked that the system seems influenced by dungeon-themed anime, as it occasionally slips into strange fantasy references. Even as these quirks circulate online, CEO Sam Altman added a lighter note by inviting users to a small gathering celebrating the new release. The contrast between rigid design and playful output has quickly become a talking point online.
The confusion comes from OpenAI’s internal Codex coding rules. These rules say the system should avoid mentioning creatures like goblins, gremlins, raccoons, trolls, and pigeons unless it is really necessary. This instruction is repeated, so it seems like it was meant to strongly limit this kind of language.
Also read: US Govt sounds alarm on Mythos AI, even as Anthropic closes in on OpenAI’s valuation
However, some users showed screenshots where GPT-5.5 still used phrases like ‘goblin mode’ or ‘gremlin’ in its answers. Later, it was found that this was not just random. During training, a more playful and slightly ‘nerdy’ style was encouraged. This style often included fun comparisons and metaphors using creatures. Over time, reinforcement learning amplified this pattern, and the language spread beyond its original context, embedding itself into general model responses unexpectedly.
This gap between the instruction and behaviour of ChatGPT 5.5 caught attention across platforms like X. Some users joked about the model’s tone, while others tracked the pattern more closely. Research platform Arena.ai even noted a rise in such words, especially when the model was not using high reasoning settings. In the recent development, the OpenAI engineers have acknowledged the issue in the comments section of the posts pointing towards the issue and have clarified that it was unintended. However, there has been no update about when the issue would be resolved.
Also read: Apple may turn iPhone camera into AI search engine with new Siri Mode in iOS 27
At the same time, ChatGPT’s official account has also joined the party, as, referring to the incident, the official account quoted The Hobbit by J.R.R. Tolkien: ‘And down, down to Goblin-town you go, my lad!’ Other than that, Altman also joined the conversation as he posted ‘goblinblog dropped’ while posting a blog post link by OpenAI explaining the issue. What could have been seen as a flaw quickly became part of the model’s public identity.
Along with that, Altman has also announced that GPT-5.5 will host a small event on May 5 at 5:55 pm. It’s an invite-only event, and the gathering will take place in San Francisco. OpenAI is aiming to gather the builders and developers around, with food, conversations, and product discussions planned. Attendees will be selected from replies, with Codex helping shortlist participants.