Human tests for machine models: What lies “Beyond the Imitation Game”?

Abstract

Benchmarking large language models (LLMs) is a key practice for evaluating their capabilities and risks. This paper considers the development of “BIG Bench,” a crowdsourced benchmark designed to test LLMs “Beyond the Imitation Game.” Drawing on linguistic anthropological and ethnographic analysis of the project's GitHub repository, we examine how contributors developed tasks based on their lay understandings of language, cognition, and intelligence. By tracing how contributors make implicit judgments about what constitutes a meaningful test of intelligence, we show how widespread language ideologies shape the evaluation of LLMs and the imaginaries that guide their development.