A research paper from the Swedish Media Council outlines a possible new approach to the automatic identification of'harmful content', by considering audio and video content separately, and using human-annotated data as a guiding index for material that may disturb viewers. Learning to Predict Harmfulness Ratings from Video, the paper illustrates the need for machine learning systems to take account of the entire context of a scene, and illustrates the many ways that innocuous content (such as humorous or satirical content) could be misinterpreted as harmful in a less sophisticated and multimodal approach to video analysis – not least because a film's musical soundtrack is often used in unexpected ways, either to unsettle or reassure the viewer, and as a counterpoint rather than a complement to the visual component. They also observe that to date, similar experiments have suffered from a sparsity of labels for full-length movies, which has led to prior work oversimplifying the contributing data, or keying in on only one aspect of the data, such as dominant colors or dialogue analysis. To address this, the researchers have compiled a video dataset of 4000 video clips, trailers cut down into chunks of around ten seconds in length, which were then labeled by professional film classifiers that oversee the application of ratings for new movies in Sweden, many with professional qualifications in child psychology. Under the Swedish system of film classification, 'harmful' content is defined based on its possible propensity to produce feelings of anxiety, fear, and other negative effects in children.