The AI Construct Lexis represents constructs, behavioral indicators, measurement instruments, and their relationships as described or implied in published research. Its contents were extracted primarily using large language models (LLMs), with human expert validation, from papers published at ICLR 2024, NeurIPS 2024, ICLR 2025, NAACL 2025, ACL 2025, ICML 2025, and EMNLP 2025.
The constructs used to describe AI systems, their relationships to observable behaviors, and the instruments used to measure them require careful conceptual and empirical validation. Theories and methods developed for human measurement may inform this work, but their assumptions should not be adopted without examining whether they are appropriate for AI systems.
The Lexis is intended as a starting point for this process, making the field's existing assumptions explicit so that they can be systematically examined, tested, challenged, and refined.