Distributed Implicit Harm: A Compositional Safety Blind Spot in MLLM-Based Video Moderation

Author1, Author2, Author3, Author4, Author5

multimodalvideo moderationsafetydataset generation

2609.00206

Builder Relevance

70%

1h ago

Abstract

The paper discusses the phenomenon of Distributed Implicit Harm (DIH) in video moderation using MLLMs, highlighting the challenges in detecting harm that arises from the composition of benign components.

Reality Card

Core Claim

The study reveals that existing MLLMs consistently fail to detect Distributed Implicit Harm (DIH) in videos, despite being able to assess individual components correctly.

Method / Result

Developed a dataset of over 9,000 videos to benchmark MLLMs on their ability to detect DIH.

Limitations

The dataset lacks compositional harm annotations and is difficult to collect due to the nature of DIH.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.

← Back to all papers