To explore the potential of Large Language Models (LLMs) as AI Annotators in the domain of sustainability reporting, this study establishes a systematic evaluation methodology. We use the specific case of European football clubs, quantifying their sustainability reports based on the sport Positive matrix as a benchmark to compare the performance of three state-of-the-art models (i.e., GPT-4o, Qwen-2-72b-instruct, and Llama-3-70b-instruct) against human expert scores. The evaluation is benchmarked on dimensions including accuracy, mean absolute error (MAE), and hallucination rates. The results indicate that GPT-4o is the top performer, yet its average accuracy of approximately 56% shows it cannot fully replace human experts at present. The study also reveals significant issues with overconfidence and factual hallucinations in models like Qwen-2-72b-instructon. Critically, we find that by implementing further data processing, specifically a Chain-of-Verification (CoVe) self-correction method, GPT-4o’s initial hallucination rate is successfully reduced from 16% to 10%, while accuracy improved to 58%. In conclusion, while LLMs demonstrate immense potential to streamline and democratize sustainability ratings, inherent risks like hallucinations remain a primary obstacle. Adopting verification strategies such as CoVe is a crucial pathway to enhancing model reliability and advancing their effective application in this field.
Wu et al. (2025) studied this question.