Training AI to Generate Human-Like Item Responses for Field-Testing

• In Event: Applications of AI

Abstract

We propose a method of using large-language models to generate human-like item responses by training the model on simulated item response data. Results showed moderate levels of success in using AI-generated responses to recover item parameters. If refined further, AI could potentially replace human test-taker responses for field-testing new items.

Author

• Hotaka Maeda, Smarter Balanced