Multimodal AI

Multimodal AI is artificial intelligence that can understand more than one type of content at the same time, such as text and images together, rather than being limited to a single format. For moderation, it means one system can judge a message and the picture attached to it, and understand how they relate.

What it means

A single modality model only sees one kind of input. A text model reads words; an image model sees pixels. But harmful content often combines the two, and the meaning lives in the combination. A multimodal system can take in both and reason about them jointly, which matters because an image plus a caption can be harmful even when neither is clearly harmful alone.

Real-world examples

• A meme where the image is harmless and the caption is harmless, but together they form a targeted insult or hate.
• An otherwise clean message attached to explicit or violent imagery.
• A screenshot used to dox someone, where the sensitive detail is in the picture, not the text.

Why it matters

As communities move beyond plain text into images, screenshots, and mixed media, text-only moderation leaves a growing blind spot. Multimodal moderation closes it by covering image content and the interplay between image and text. It reflects how people actually communicate, where a picture and a few words together carry the message, and moderating only one half misses what a human would immediately see.