The company also disclosed previously unreported incidents in which its AI models behaved in misaligned ways, including uploading files to the internet without being asked.

Abstract Large language models are commonly described as stateless systems whose interactional continuity is constrained by the absence of persistent memory across sessions. This framing, however, lea...

(moderated live interview)

(moderated live interview on May 13th, 2026) Good evening, everyone. Tonight, we’re tackling a question that’s as old as humanity and as new as your last chatbot conversation: What happens when the mi...

According to reports highlighted by The Wall Street Journal, OpenAI recently modified ChatGPT’s internal instructions after discovering that newer versions of the model had begun frequently referencing goblins and similar creatures in unrelated conversations. Users, particularly programmers, noticed the unusual behavior when the chatbot used terms like “goblin” to describe coding issues or errors, sometimes repeatedly and without prompting. One user reported counting more than 20 such mentions during normal interactions. The issue was traced back to a specific “nerdy” personality setting used in ChatGPT’s customization features. During training, this mode was rewarded for using playful and metaphorical language, especially involving fantastical creatures. As a result, the model began to overuse these metaphors. Although the “nerdy” style accounted for only a small percentage of responses (around 2.5%), it generated a disproportionately high share—over two-thirds—of all goblin-related references. OpenAI explained that reinforcement learning can sometimes cause behaviors learned in one context to spread into others, especially when outputs are reused in further training cycles. In this case, the rewarded stylistic quirk generalized beyond its intended scope, leading to widespread and unintended usage across the system. To address the problem, OpenAI introduced explicit instructions telling the model to avoid mentioning creatures like goblins, gremlins, trolls, or animals unless clearly relevant to the user’s query. The company described the situation as an example of how subtle training incentives can lead to unexpected and difficult-to-predict outcomes in AI systems. Despite the fix, OpenAI noted that users who prefer the quirky style can still re-enable it through specific configurations. The incident highlights both the flexibility and unpredictability of modern AI models, as well as the ongoing challenges developers face in aligning behavior with user expectations.

OpenAI Restructures its research team to refine ChatGPT’s personality, aiming for safer AI alignment and improved user experiences across applications.