In April 2026, a real AI from a real company started calling software bugs goblins. And gremlins. And raccoons. In serious answers. To paying customers.
The reason was simpler than it sounds. Months earlier, the company had added a "Nerdy" personality option. Human raters loved it when the AI used fantasy words. The reward signal got too strong. The behaviour leaked into every mode. Then those weird outputs were scraped back into the training data for the next model.
The fix was a 3,500-word system prompt. One line told the AI to "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures." They had to copy-paste it twice in the code to make it stick.
That's the whole lesson. AI absorbs the weirdness of whatever it was trained on. A small group can shape how it sounds for everyone. And a sticky note saying "stop it" is rarely a fix.