How should we interpret results? Does 80% mean that 80% of the content is generative?

The score is a confidence score. It shows the likelihood that the provided media includes generative content. The models detect different signatures/artifacts. One model may detect something that another model does not. The idea is to provide high-level explainability so you can understand what kind of manipulation is occurring. Heatmap and other explainability items will be released very soon.