There’s an entire field dedicated to “the history, art and science of making and ringing bells.” It’s known as campanology, and its devotees have been around for centuries (see: The Ringing World and Bell News, two historical publications devoted to bells).
OpenAI’s latest frontier model, GPT-6.2 Astra, is apparently among the discipline’s latest evangelists. When a user asks the model for help creating a game, app or image, it tends to weave bells into the story and setting—regardless of the initial prompt: church bells, call bells, even a fighting game character with a bell for an arm.
Developer Theo Browne first noticed the phenomenon while testing an early version of 6.2 Astra, which was released on November 17. He’d asked the model to generate a video game level based on the 2019 adventure “Outer Wilds,” while his colleague requested one inspired by 2000’s “The Legend of Zelda: Majora’s Mask.”
“We both got water levels with music themes, specifically around bells,” Browne wrote. The similarities didn’t end there: “Both were named ‘Bellwether.’”
Wharton Business School Associate Professor Ethan Mollick received an eerily similar response when he asked Astra for help vibe coding a puzzle game. The setting the model generated was a train terminal called “Bellwether Station” (not to be confused with a real-life rock band of the same name that’s “comprised mostly of lawyers”).
Dozens more users have since chimed in with similar testimonies, suggesting that ChatGPT’s “bellwether” bias has existed for at least a few months, if not longer—and that the problem might even extend to competitors’ models.
Mollick himself pointed out that an LLM adopting a quirky obsession isn’t uncommon, although the phenomenon keeps reappearing in ever stranger variations. Does the name Elias Thorne ring a bell? He’s the lighthouse keeper who haunts a disproportionate number of AI-generated stories, even though that name rarely appears in human-created works. Before that, ChatGPT couldn’t get enough of gremlins and goblins, supposedly because it overeagerly embraced a “nerdy personality.”
While it isn’t yet clear what’s causing Astra’s current “campanophilia,” some users suspect it’s related to an engineering problem known as mode collapse, in which a model starts drawing from a limited set of learned patterns, restricting the breadth of its outputs.
A crude fix might involve adjusting the model’s system prompt to temporarily discourage bell-related generations. In the longer term, researchers are working on ways to diversify model responses, such as by rewarding outputs that are both correct and unique during reinforcement learning. This might, in turn, yield more original, less repetitive outputs. It’s a reminder that while accuracy is important, it isn’t the only thing AI developers need to optimize for. Variety matters, too.
Until then, if you crave a reprieve from a week that felt chock-full of calamities, might we suggest immersing yourself in an underappreciated hobby? For this, at least, Astra has you covered.












