What the Corpus Can No Longer Find
Legacy YouTube comment IDs in a 2017 research deposit no longer resolve, although most of the comments are still there
Abstract
Many YouTube comment datasets are shared as lists of comment IDs. Anyone who wants the comments back has to retrieve them through the YouTube Data API, which only works if the IDs still work. I tested this on a real case: the 703,339 comment IDs that Veletsianos et al. (2018) collected in 2017 and deposited with their published study.
The deposit itself is sound. Its published frequency table is reproduced exactly (22 of 22 cells), and an independent reanalysis (Motnikar, 2019) is reproduced to within 0.06 on every mean.
The IDs are not. On 25 October 2026, not one of them could be found: 0 of 3,941 IDs from three pilot videos, and 0 of a random sample of 500 from the whole deposit (95% upper bound 0.74%). The videos are still there (660 of 665), and so, for the most part, are the comments. On ten random videos, the number of comments visible today from the same period equals 87.5% of the deposited number, but these comments carry IDs in a different format. YouTube has documented no such change, although it announced a comparable one for caption IDs in 2022.
In practical terms, an ID-based survival test of this deposit would have reported a loss of 100% while most comments are still visible. Anyone who reuses or shares YouTube comment data by ID should therefore test a sample of IDs one by one before relying on them, and should deposit the analytical variables alongside the IDs.
The series
This is the second of three measurement papers on YouTube comment corpora. What the Corpus Does Not Say asks what a corpus leaves out when it is collected; What the Corpus Still Holds takes three younger deposits whose IDs still work and asks how much of them survives and whether the loss changes their results.
Comments
Refutations, reviews and comments are most welcome: research@woidwide.net. I have worked as carefully as I could; if I have made a mistake or overlooked relevant work, I apologise and would be glad to be told.
AI use: the author designed the study and set all analytic decisions and decision rules. Claude (Anthropic) was used under his direction to draft text and analysis code, which the author then edited. All numbers come from scripts run on the deposited data. The author is responsible for the content.
Citation
Sedlmair, Max (2026). What the Corpus Can No Longer Find: Legacy YouTube comment IDs in a 2017 research deposit no longer resolve, although most of the comments are still there. Research note, version 1.1. Zenodo. https://doi.org/10.5281/zenodo.23642703