An Experimental Study on Generating Plausible Textual Explanations for Video Summarization

Explanations need to match human reasoning: The research focuses on "plausibility"—whether AI-generated explanations align with how humans actually think about video summaries. This is important because explanations that sound reasonable but don't reflect human logic can give false confidence in AI systems.

Visual explanations alone aren't enough: The study found that showing which parts of a video the AI focused on (visual explanations) needs to be paired with natural language descriptions to be truly useful. Using advanced language models like LLaVA-OneVision to convert visual explanations into text makes them more understandable to users.

Faithfulness doesn't guarantee plausibility: A key finding is that explanations which accurately represent what the AI actually did (faithful explanations) aren't automatically ones that make sense to humans (plausible explanations). This distinction matters for responsible AI—a system might be technically transparent but still confusing to users.

Semantic similarity can measure explanation quality: The researchers developed a method to evaluate plausibility by measuring how much the textual explanations overlap in meaning with the actual video summaries produced. This provides a concrete, measurable way to assess whether explanations are genuinely helpful rather than just relying on subjective judgment.

Standardized evaluation is needed for explainable AI: By testing multiple methods and datasets, this work contributes to establishing best practices for evaluating explanations in AI systems, which is essential for building trustworthy AI that users can actually understand and rely on.