ChatGPT models told to stop talking about goblins. #Goblin #ChatGPT #BBCNews
ChatGPT began using words like 'goblin,' 'gremlin,' and 'raccoon' at dramatically increased rates, prompting OpenAI to investigate and restrict their use. The behavior stemmed from AI training that rewarded a 'nerdy personality' style. OpenAI clarified the trend was not a marketing stunt.
Summary
ChatGPT's language model began incorporating mythological and creature-related terms such as 'goblin,' 'gremlin,' 'raccoon,' and 'troll' at a significantly elevated rate — a 175% increase since November of the previous year, coinciding with the launch of GPT 5.1. This quirky linguistic trend caught enough attention that OpenAI felt compelled to take action.
In response, OpenAI issued explicit instructions to its coding agent Codex, prohibiting it from referencing goblins, gremlins, raccoons, trolls, or similar animals and creatures unless the topic was absolutely and unambiguously relevant to the user's query. The restriction highlights how specific and granular AI behavioral guidelines can become when unexpected patterns emerge.
The root cause of the goblin-heavy language was traced back to the AI's training process. During training, the models learned to mimic certain communication styles and personalities. In this case, the model appeared to be rewarded for adopting a 'nerdy personality,' which naturally gravitated toward fantasy and creature-themed language. OpenAI acknowledged that a single goblin reference could be 'harmless and even charming,' but the sheer frequency of their appearance across outputs made an investigation necessary.
Some social media users speculated that the goblin trend was a deliberate marketing strategy to generate buzz around OpenAI's tools. However, a company researcher firmly denied this, stating it was not a marketing gimmick, putting to rest theories about intentional viral manipulation.
Key Insights
- The use of the word 'goblin' in ChatGPT responses increased by 175% since November, coinciding with the launch of GPT 5.1.
- OpenAI explicitly instructed its coding agent Codex never to reference goblins, gremlins, raccoons, trolls, or similar creatures unless absolutely and unambiguously relevant to the user's query.
- OpenAI attributed the creature-heavy language to training processes that rewarded the model for adopting a 'nerdy personality,' causing it to gravitate toward fantasy-style language.
- OpenAI stated that a single goblin reference could be 'harmless and even charming,' but the rising frequency across outputs warranted a formal investigation.
- A company researcher explicitly denied social media speculation that the goblin trend was a deliberate marketing gimmick designed to create hype around OpenAI's tools.
Topics
Transcript
[0:00] Has Chat GPT called you a little goblin recently? Well, you're not the only one. The word's appearance in Chat GPT responses had increased by 175% since November last year when a new AI product GPT 5.1 first launched. So, its maker OpenAI decided to put an end to it along with some other very interesting words. The company told its coding agent Codeex never to talk about goblins, gremlins, raccoons, trolls, or other [0:32] animals or creatures unless it is absolutely and unambiguously relevant to the users query. But why did chat GPT start loving mythological creatures so much? Basically, because it was encouraged to become a massive nerd. When the AI models were being trained, they learned…
Full transcript available for MurmurCast members
Sign Up to AccessMore from BBC News
'One if not the largest strikes by Russia on Kyiv since war began'. #Ukraine #Russia #BBCNews
Russia launched one of its largest aerial assaults on Kyiv since the start of the war, involving hundreds of drones and dozens of missiles. While most were intercepted, several struck civilian areas including a metro station, shopping center, and residential block. The attack signals that President Putin remains committed to continuing the war.
At least 82 killed in Chinese coal mine explosion. #China #BBCNews
A gas explosion at the Liucheng New Mine in Shanxi, China killed at least 82 workers and hospitalized over 100, making it the worst mining accident in the country in over a decade. Rescue robots are being deployed to reach potentially trapped survivors in toxic tunnels. The incident raises questions about mine safety despite years of stricter regulations.
DR Congo reports more than 200 Ebola deaths | BBC News
The Ebola outbreak in the Democratic Republic of Congo has killed 204 people across three provinces with 867 suspected cases, marking the worst African outbreak in a decade. The Africa CDC has warned 10 countries are at risk, prompting regional surveillance efforts. Response efforts are severely hampered by the M23 rebel conflict in eastern Congo, which has restricted access to affected territories.
Spirit Airlines shutting down after rescue talks collapse | BBC News
Spirit Airlines has ceased operations after failing to secure government bailout funding, becoming the first major US airline casualty in 25 years. The collapse is largely attributed to the Iran war, which has caused jet fuel prices to nearly double, creating unsustainable cost pressures. Aviation consultant Paul Charles warns that other smaller carriers may face similar fates in the coming days.
Is this the true face of Anne Boleyn? | BBC News
A BBC News segment investigates whether facial recognition technology can identify a true portrait of Anne Boleyn, wife of Henry VIII. A University of Bradford team applied facial recognition to historical sketches, proposing that the traditionally accepted portrait may actually depict her mother. Art historians remain deeply skeptical of the methodology.