On May 19th, 2026, Nikolina Filiposki presented at the 2026 The Responsible AI Forum (TRAIF) held at Amerikahaus in Munich. During the AI Governance in Institutional and Regulatory Contexts session, Ms. Filiposki presented: “PolicyLink: Deriving Content Moderation Decisions with Large Language Models”.

abstract excerpt: abstract excerpt: Content moderation is a central political issue in digital governance. Platforms routinely remove billions of posts based on community guidelines; yet, the consistency between these guidelines and removals remains poorly understood. To advance transparency in this domain, we introduce PolicyLink, a hybrid retrieval and zero-shot reasoning framework that evaluates the consistency of moderation policies across platforms. PolicyLink integrates lexical (BM25) and semantic (FAISS) retrieval in a multilingual hybrid Retrieval-Augmented Generation (RAG) pipeline trained on over 80 English and 30 German community guideline corpora. Given any piece of text-based user content, the model identifies the most relevant policy sections and returns a structured verdict (allow, remove, or unknown) grounded in explicit rule citations. We validate the framework using real moderation outcomes from YouTube, comparing PolicyLink’s verdicts with platform decisions and conducting a human user study on YouTube Shorts. Results show where platforms, humans and zero-shot models converge in their interpretation of the same rules, offering a new lens on consistency and the political work of transparency in moderation.

The Responsible AI Forum 2026 was funded by the European Union. More on alignAI.

Watch a video of the talk below.