The Limits of Automatic Evaluation of Creativity in Large Language Models

Research investigating the evaluation of creativity in Large Language Models (LLMs) has found significant discrepancies between automatic evaluation methods and human judgment. Human assessments of both human- and AI-generated content reveal that current automated metrics, including LLM-as-a-Judge approaches, do not reliably capture human perceptions of creativity.

Why it matters

This research is strategically important as it highlights a fundamental limitation in the current development and deployment of AI systems intended for creative tasks. Organizations relying on automated methods to assess creative outputs from LLMs may be operating on flawed metrics, potentially misjudging the quality, originality, or value of AI-generated content. This misalignment necessitates a re-evaluation of how creative AI applications are developed, tested, and integrated into workflows.

Key insights

• LLMs are capable of generating text that challenges human performance in creative domains.
• Evaluating creativity in LLM-generated content remains a significant challenge.
• Current automatic evaluation methods do not reliably align with human judgments of creativity.
• The study compared human evaluations across 11 creativity dimensions with automated objective metrics and LLM-as-a-Judge evaluations.
• Experiments revealed substantial misalignment between automatic evaluations and human assessments.
• LLM-based judges exhibit a systematic preference for certain characteristics, suggesting bias in their evaluation.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.23705