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

The article discusses the challenges of making AI-generated text appear more human-like, beyond just improving language model intelligence. The author shares their approach to systematically designing AI behavior, including reply timing, emotional tone, and context awareness.

Details

The author initially tried a naive approach of using if-statements to control AI behavior, but realized this would not scale across languages and cultures. Instead, they designed an object-oriented base class that defines the common structure of human communication, such as reply times taking time, emotions shifting through a conversation, and referencing previous context. Language and culture-specific parameters are then defined in JSON configuration files, allowing derived personas to be created without rewriting Python code. The key components include a timing controller that uses a normal distribution to mimic human reply times, an emotion controller that shifts tone over a conversation, and a context manager to reference previous context. This approach aims to make AI-generated text appear more human-like by addressing the behavioral patterns that currently make it recognizable.