The video explains common grounding failures, including overextension, conflation and citations that do not support the attached claim. It demonstrates an interface that places source documents alongside generated text for human reviewHuman-in-the-loop AI keeps a person involved in reviewing, correcting, approving, or guiding an AI system's work..
The presenter compares separate citation checkers, dedicated citation agents, evidence-first and claim-first workflows. They distinguish faithfulness to a retrieved sourceRetrieval is the process of selecting relevant stored information and returning it to an AI system for the current task. from checking whether that source is true, and discuss targeted verification against authoritative materialEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
Watch on YouTube
