The substance of the videos
Full transcripts, by speaker and by sentence, produced by the project and not by the platforms’ subtitles.
29.5kvideos transcribed
About
A permanent observatory of political YouTube and TikTok.
YOUPOL continuously collects, transcribes and annotates the videos of the tracked political channels, and keeps what the platforms remove. An observatory of political YouTube and TikTok, carried by its authors and computed on their own machines.
126
tracked channels
29.5k
transcribed videos
4.3M
classified sentences
2.4k
deleted videos preserved
Why
FDENEWS, the example of a deleted channel
12 March 2025
That day, YouTube deleted FDENEWS, a French far-right channel. It had 1,755 videos, 68.3 million views and 103,000 subscribers in nearly ten years. For anyone who had not planned ahead, everything had ceased to exist.
YOUPOL was tracking the channel. Every video was already transcribed, every comment and every piece of metadata recorded. The deletion became a data point. The observatory collects before the platforms remove, because what they remove is what research most needs to study.
Four gaps filled
Existing bases stopped at the metadata
Databases on online political discourse stopped at titles, metadata or comments. YOUPOL is, to our knowledge, the first to combine these four things.
The deleted content
Removed videos, closed channels and moderated comments stay in the base, with the date and the reason for the removal.
2,422vanished videos preservedAudience trajectories
Views, likes and comments are measured at discovery, at day 1, 7 and 30, then every month. Influence reads over time.
2.3Bviews measured on videos onlineComments, over time
Comments are kept with their date, likes and replies, even when deleted later.
13Mcomments preserved
How
Three construction choices
01
Transcribe ourselves
The voice is isolated from the music, speakers are told apart without being identified, then speech is transcribed in the language spoken. The result depends neither on the platform’s subtitles nor on their availability.
02
Annotate down to the sentence
An annotation frame where a local language model assists the coding, checked by the team, trains sentence-level classifiers. They tell whether a sentence is political, then which themes it addresses. Their reliability is measured and published.
03
Compute on our machines
The computers of the team and of collaborators, linked by a distributed-computing application, transcribe and analyse the videos. The observatory depends on no institutional cluster and on no platform interface.
For whom
Made to be used
- 01
Citizens
free, no accountSee what is said this week on political channels, which public figures recur, which arguments circulate and where they come from.
- 02
Journalists and teachers
records, trends, removalsFind a vanished video and its record, compare orientations on an equal footing, cite a trend with its base and its limits.
- 03
Researchers
explorer, API, R and Python clientsA validated instrument, documented in a methods paper, with access to transcripts, classified sentences and comments on request.
The data
Three regimes for the data
Published
Video metadata, counts per sentence and per theme, mentions of public figures, automatic records and syntheses, audiences, removals, the list of channels and their orientation.
Under controlled access
The text of transcripts, speaker turns and comments, with their authors’ pseudonyms, reserved for research projects, by access tier.
What is never published
The name of a private person named in a video (victim, witness, accused). Records that contain one are not published. No video is redistributed; the player points to the platform.
Milestones
The project since 2024
- Building the corpus — French and Quebec channels chosen and coded by orientation; first transcriptions, the archive going back to the channels’ first videos.
- Classifiers and validation — assisted and checked annotation, training of the political and theme classifiers, precision check; paper in the Revue française de science politique.
- Collaborative infrastructure — distributed computing on the project’s computers, methods paper, data explorer, API, R and Python clients.
- Public observatory — videos of the week, automatic records and syntheses, arguments by theme, vanished videos.
- United States — eleven American channels join the corpus; measures compared per video rather than by sentence volume.
Open science
The method is published, so is the code
- 01
The methods paper
SocArXiv · 2026Lemor, A. & Boursier, T. (2026). YouPol: A Collaborative Research Infrastructure and Database for Political Content on YouTube and TikTok.
- 02
The clients
R · Python · GitHubyoupoldata (R) and youpol-client (Python) talk to the API with a personal key; their code is public.
- 03
The method, in detail
ten stepsBuild the corpus, detect, extract, transcribe, classify, name, summarise, draw out the arguments, track removals, publish. Each step comes with its models, thresholds and reliability.
The team
About the team
Tristan Boursier
Postdoctoral researcher
Sciences Po Paris · Université du Québec en Outaouais
Research design, orientation coding and political analysis.
Antoine Lemor
Postdoctoral researcher
Université de Sherbrooke · CIRST · RFICS
Methods, machine learning, infrastructure and website.
Arnaud Miranda
Research collaborator
CEVIPOF · Sciences Po Paris
Political theory and analysis of reactionary discourse.