Designing and Open-Sourcing a Base Class for AI to Behave Like Humans

The Trigger: AI-Written Text Was Instantly Recognizable When I first tried to automate business communication with AI, the prototype output was this: Thank you for your message. Regarding this matter, we can deliver within three days. If you could share the detailed requirements, we can start immediately. Should you have any questions, please do not hesitate to let us know. Perfect Japanese. Flawless in both grammar and honorifics. And yet, anyone could tell it was written by an AI. Why? There are three fatal patterns: Replies come in 30 seconds. A human would need time to think. The same tone every time. The third exchange is as polite as the first. It always ends with "please do not hesitate." A human wouldn't say it so readily every single time. In 2024, a paper by Jones & Bergen published in PNAS backed up this intuition. When GPT-4.5 was instructed to adopt a "human-like persona," it was perceived as human 73% of the time—surpassing the recognition rate of actual human participant