Multimedia communication on social platforms evolves swiftly, with memes becoming particularly significant. Unfortunately, they can be used maliciously, highlighting the need for detecting hateful memes. Research has introduced visual language models (VLMs) to address this, but traditional machine/deep learning models typically require labeled datasets.

• Recent emergence of VLMs sees impressive performance in various tasks.
• This study employs VLMs for zero-shot classification of hateful memes.
• Findings reveal VLMs’ vulnerabilities in zero-shot hate meme detection.
• Large VLMs lack necessary robustness for accurate meme classification.

In my view, this paper underscores the complexity of contextual understanding in AI. Its implications for social media moderation are substantial, prompting further exploration into enhancing VLMs for better content management.