Explainable AI (XAI) has become a prominent research area, yet reliable evaluation of explanation methods remains elusive. Conventional approaches either rely on proxy metrics that poorly reflect real utility, or on human studies that are costly, hard to scale, and difficult to reproduce. This paper investigates whether multimodal large language models (MLLMs) can serve as surrogate human evaluators to automate XAI assessment. We first introduce a purpose‑grounded evaluation framework that defines metrics applicable to both MLLMs and actual users. Then we provide conceptual insights on the conditions under which the models can be trusted as substitutes for human judgment. Finally, we conduct extensive experiments across five purpose‑grounded scenarios—model improvement, error diagnosis, decision support, educational assistance, and fairness—and evaluate fourteen popular MLLMs. Results show a universally positive trend, with OpenAI and Qwen families achieving higher alignment than other variants, making them the preferred choices in practice. We stress that this approach complements, rather than replaces, user testing, helping the XAI community move away from an over‑reliance on proxy metrics.
Review: The study offers a practical pathway for automating XAI evaluation, especially valuable in resource‑constrained settings. Nonetheless, attention to model bias and task‑specific nuances is essential, and future work should examine cross‑domain robustness.