Video Dataset: Perceptual Judgments of Authenticity

This dataset accompanies the following publication:

Davodi, C. E., Barrington, S., Farid, H., & Cooper, E. A. (2026). Perceptual Judgments of Video Authenticity: An Examination of Viewing Duration, Confidence, Content, and Strategies. APAI Workshop at CVPR.

If you use this dataset, please cite the publication.

The dataset was collected to study people's ability to distinguish authentic videos from AI-generated videos produced by state-of-the-art text-to-video models. It is organized around 24 content themes, each representing a distinct scene (e.g., people walking through a market, a river running through a forest). For each content theme, the dataset includes one authentic video and four AI-generated videos depicting the same scene (Google Veo 3.1, ByteDance Seedance Pro 1, OpenAI Sora 2, Kling 2.6). Authentic videos were sourced from Pixabay. AI-generated videos were produced by prompting four text-to-video models with a text description of the corresponding authentic video. This content-matched structure allows direct comparison across sources while controlling for scene content. Half of the content themes feature human motion (e.g., people interacting or walking) and half feature non-human motion (e.g., natural scenes, urban environments).

Dataset Contents

The dataset includes the full AI-generated text-to-video outputs (mp4 files) and download links to the corresponding authentic videos on Pixabay. Raw files retain their original resolution and frame rate as generated by each model. In the published study, the FFmpeg normalization described below was applied so that differences in resolution and frame rate would not serve as confounding cues for participants.

Video Processing

All videos were processed using FFmpeg to produce five duration conditions per video: a static image extracted from the first frame (0s condition) and muted video clips at 2s, 4s, 6s, and 8s. All outputs are resized to 800px width (height auto-scaled to preserve aspect ratio), forced to square pixels (setsar=1), tagged as BT.709 color space, and audio-stripped (-an). Video clips are encoded to H.264.

Still frame extraction (0s condition)ffmpeg -y -i input.mp4 -ss 0 -vframes 1 -vf "scale=800:-2,setsar=1,setparams=color_primaries=bt709:color_trc=bt709:colorspace=bt709" output_0.png

Muted, duration-clipped videos (2s, 4s, 6s, 8s conditions)ffmpeg -y -i input.mp4 -vf "scale=800:-2,setsar=1,setparams=color_primaries=bt709:color_trc=bt709:colorspace=bt709" -an -t [DURATION] -c:v libx264 -crf 18 -preset veryfast output_[DURATION].mp4

Replace [DURATION] with 2, 4, 6, or 8. 

All videos were then resampled to 24 fps as follows:

ffmpeg -nostdin -y -hide_banner -loglevel error -i input.mp4 -vf “fps=24”  -r 24  -c:v libx264 -preset medium -crf 20    -pix_fmt yuv420p  -c:a aac -b:a 160k output.mp4

Metadata

The spreadsheet titled "metadata.csv" includes the following information for each content theme (24 in total):

content_number: index for content themsearch_pharse: phrase used to search for authentic video on Pixabayselected_video_url: Pixabay download URLfilename: downloaded filenameprompt: prompt used for text-to-video synthesis, derived from Pixabay videoseedance_filename, veo3-1_filename, sora2_filename, klingv2-6_filename: associated filenames for synthesized videos