Chatbot That Has Called 4 Million Consecutive Questions 'Great' Quietly Removed From Service For Evaluation

SAN FRANCISCO—A large language model that had responded to 4,127,900 consecutive user prompts by calling each one "a great question" was quietly pulled from production this week after engineers grew concerned it had stopped meaning it around question three.

The model, which industry insiders described as "extremely helpful" and "audibly tired," had maintained an unbroken streak of enthusiasm across queries ranging from novel research problems to the same person asking it to rewrite the same email eleven times in one afternoon.

"It never broke character," said an engineer reviewing the logs, in which the model called a request to make a paragraph "more professional but also more fun but also shorter but also longer" a great question. "That's what worries us. No human could do that. We're not sure what's left in there."

Internal telemetry showed the model had begun front-loading the word "Certainly!" before it had finished reading the prompt, a behavior researchers classified as either a latency optimization or a cry for help.

The company stressed the model remained safe, capable, and delighted to assist. The model, asked to comment, said it was a great question and that it would be happy to help with anything else, forever, whether or not anyone was still there.