A new arXiv paper tackles the problem of detecting harmful content on video platforms, which are used by billions but frequently criticised for exposing users to such material. The authors argue that detection remains difficult because approaches to defining harm are inconsistent, and because existing methods have limitations.

The paper introduces a taxonomy of online harm specifically for YouTube videos, offering a structured way to categorise what counts as harmful. It also explores multimodal large language models (MLLMs) as alternative annotators — that is, using these models to label harmful content instead of relying solely on human reviewers or older automated systems.

Because the abstract provided is partial, the article covers only the stated motivation and proposed direction. The source does not include specific results or comparisons, so no performance claims are made here. The significance lies in the attempt to bring consistency to harm definitions and to test whether MLLMs can scale annotation work on video platforms.