Track removals

A video in the corpus can become unavailable through deletion by its author, a switch to private, removal by the platform or the closing of its channel. The full list of each channel’s videos, reread every week (step 02), and the monthly passes over the corpus reveal these disappearances. A video known to the base that no longer appears there is checked, then marked removed.

The date shown is the date on which the removal was noticed; the deletion may be earlier. The reason is known only if the platform states it at the time of the check. The video’s record stays in the base with its metadata, its timestamped audience figures and, if it had been transcribed, its transcript; removals count in the trends, year by year, and closed channels are listed separately.

The removed videos page shows the removals noticed in the last 120 days. See the removed videos →

When the video had been transcribed and the reason is known, the language model used for the summaries (step 07) writes a record from the transcript. It gives a title, a summary of three to five sentences, the elements of the content that fall under the usual platform rules (hate speech, calls to violence, medical or electoral misinformation…) and a reading that relates the reason given to the content.

This reading is an automatic interpretation, phrased as a hypothesis. It notes that a deletion by the author or a switch to private may have reasons other than the content. When nothing in the content explains a removal by the platform, it says so.

Records are not reviewed. A record that seems to name a private individual is withheld. The synthesis at the top of the removed videos page is written from the records and renewed at most once a week.