Measuring how well an AI explanation works is hard because we often rely on simplistic metrics that don't capture real usefulness, or we have to run expensive and slow human tests. This paper explores using multimodal , AI systems that can process both text and images, as stand-ins for human participants to automate the evaluation. The authors tested fourteen such models across five common explanation tasks and found that OpenAI and Qwen models performed best, suggesting they could help researchers move beyond weak metrics without fully replacing human testing.
Chatbots can help assess AI explanations, but human alignment varies
Across five evaluation areas and fourteen multimodal models, the authors report positive alignment trends, with OpenAI and Qwen models generally aligning more closely with human evaluation.
AI Startup
Eoin M. Kenny · Salim I. Amoukou · Tom Bewley
JP Morgan · Mistral AI
Research Digest··2 min read
In one line
Multimodal large language models can surrogate human evaluators for explainable AI.
Before you read this
Two ideas this paper builds onlarge language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.306 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.250 papers →
Read it as
Kenny and colleagues investigate whether multimodal language models can stand in for human participants when evaluating explanations of AI decisions.
The authors propose a “purpose-grounded” evaluation framework: assess an explanation against the practical task it is meant to support, rather than relying only on an indirect numerical measure of explanation quality.
How it works
stage by stageThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.