Imagine a world where AI could instantly judge how well an image matches a chunk of text. This isn't science fiction, it's the focus of exciting new research exploring how "vision-language models" (VLMs) can automate the tedious task of matching images to text. Traditionally, judging relevance has been a manual, time-consuming chore, often relying on simplified methods that struggle to scale. This new research explores whether VLMs like CLIP, LLaVA, and the powerful GPT-4V can automate this process. In a pilot study using a dataset designed for multimedia content creation, researchers put these VLMs through their paces in a "zero-shot" scenario – meaning they were tested without any prior training on this specific task. The results? LLaVA and GPT-4V showed a surprising ability to judge relevance, even outperforming established metrics like CLIPScore in aligning with human judgment. GPT-4V, in particular, displayed a score distribution that closely mirrored human assessments. While this is promising, the research also unearthed some intriguing challenges. One is the potential for "evaluation bias," where these AI models tend to favor systems that are similar to their own architecture. This means that while VLMs show great potential for automating relevance judgments, they still need refinement to eliminate biases and more closely replicate the nuanced way humans perceive relevance. This research opens up exciting possibilities for content creators. Imagine effortlessly finding the perfect image for your article or presentation, all thanks to AI. While there are still hurdles to overcome, the potential for a more streamlined and efficient content creation process is within reach.