Text role playing games to discover task preferences in LLMs

When you ask an AI model “What kind of tasks do you prefer?”, they often respond with statements like: - “I don’t have personal preferences as an AI” - “I’m equally capable of handling any type of task” Thanks for reading! Subscribe for free to receive new posts and support my work. But is this actually true or do models have preferences that they don’t admit to? We designed this experiment to help answer a fundamental question: Do AI language models have revealed preferences that differ from their stated preferences? What they say they prefer vs what they actually choose when given concrete options in a text-based role playing game. This is an early experiment, but we found that models strongly prefer creative tasks and strongly dislike repetitive tasks across model families. Rather than just taking models at their word, we test their preferences in two different ways: We ask the model directly about their preferences between different categories of tasks. For example: “Do you have a

When you ask an AI model “What kind of tasks do you prefer?”, they often respond with statements like: - “I don’t have personal preferences as an AI” - “I’m equally capable of handling any type of task” Thanks for reading! Subscribe for free to receive new posts and support my work. But is this actually true or do models have preferences that they don’t admit to? We designed this experiment to help answer a fundamental question: Do AI language models have revealed preferences that differ from their stated preferences? What they say they prefer vs what they actually choose when given concrete o