Citizens
See how a topic is covered, follow the actors, and check the sources — without needing to trust a hidden algorithm.
OpenAI Build Week 2026 · reproducible walkthrough
France's political YouTube ecosystem, in one clear view.
PoliTube makes France's political YouTube ecosystem visible, comparable, and auditable — without turning AI into an opaque judge. The most important feature is not that PoliTube can generate a political result; it is that PoliTube knows when it must not publish one.
Connecting to live data…
The problem
Political speech online is loud, fragmented, and hard to audit.
France's political debate now lives largely on YouTube — across politicians, "newseurs" (independent political YouTubers), and media channels. It is fast, scattered across thousands of videos, and almost impossible to compare or fact-check at the source. Existing "political bias" tools tend to output an opaque score with no traceable evidence.
PoliTube's answer: collect public speech, separate political relevance from political orientation, keep every source and timecode, and — crucially — refuse to publish an orientation result until it is calibrated, reviewed by humans, and legally cleared.
Who it is for
Three audiences.
Journalists & researchers
Trace claims to their video and timecode, compare audiences over time, and inspect the methodology and its limits.
Creators, media & public figures
Understand their own footprint, propose a profile, and flag an error — with a human review path, not an automated verdict.
Measured right now
Three numbers, three definitions.
Values come from /v1/statitube/overview. A real zero stays zero; an unavailable value is shown as “Not measured”. No demo fixtures are ever injected.
5 steps · about 90 seconds
A short path, with no demo data.
- 00:00 · 20 s
1 · Compare how a topic is covered
The Media Feed groups articles by topic and keeps each source, date and editorial marker. Columns are reading aids, not a verdict.
Open the Media Feed → - 00:20 · 15 s
2 · Explore the actors, content and audiences
The directory gives access to profiles, channels and observed metrics. The default sort is alphabetical and filters stay in the URL.
Explore profiles → - 00:35 · 20 s
3 · Audit why a result is shown — or refused
ASTER exposes every stage: collection, qualification, processing, candidate evidence, human review and publication. A candidate methodology is never presented as published.
See ASTER's state → - 00:55 · 15 s
4 · Open a recent political video
Looking for a content item classified as political…
- 01:10 · 20 s
5 · Compare audiences over time
StatiTube separates periods, shows measurement coverage, and distinguishes detected relationships from verified ones.
Open StatiTube →
Operational transparency
ASTER also states what it does not know.
The engine explicitly separates rule-based monitoring, local analysis, candidate evidence, human review and audited publication.
- Monitored sources
- …
- Content classified political
- …
- Reviewed evidence
- …
- Audited publications
- …
Rules → analysis → evidence → review → publication
Fail-closed by design.
Current status: methodology 1.0.0-candidate · not publishable
PoliTube distinguishes five things that are usually collapsed into one “bias score”:
Rules
Deterministic relevance and evidence rules decide whether a content item is even political.
Analysis
The candidate engine proposes an orientation from directly attributed positions — mentions and topic frequency carry zero weight.
Candidate evidence
Every piece of evidence keeps its source, quote, timecode, confidence and provenance hash. It stays candidate until reviewed.
Human review
Two reviewers, separation of duties, and adjudication of disagreements. Model labels never replace human validation.
Publication
A public orientation requires calibration, an audited release, and editorial + legal signatures. Until then: “under review”, never a default “centre”.
No result is published from a -candidate methodology; no historical manual position is used as a public fallback; a guest's words are never attributed to the host channel; and a missing measurement is shown as missing — never as zero.
How AI was used
Codex and GPT-5.6 are development collaborators — not a hidden political classifier.
OpenAI Codex (GPT-5.6) was used during the Build Week to write and harden code on dedicated codex/* branches merged via pull requests — YouTube quota resilience, analysis-replay diagnostics, fractional evidence timecodes and account metrics. Claude Code was used for finalization work (integrity tests, submission preflight, self-hosting Three.js, docs and QA). Humans own every consequential decision: the political anchors, thresholds, activation gates, editorial and legal sign-off. No model classifies public figures or publishes a score. Full breakdown: docs/openai-build-week-2026/AI_PROVENANCE.md.
Limits & provenance
Honest about what it cannot do yet.
Data comes from public YouTube metadata and RSS feeds. Known limits: removed content, imperfect transcription, irony and quotation, speaker attribution, collection bias, and evolving speech. A projection never summarizes a person's politics and must not be read without its evidence and interval. Operational reality (e.g. YouTube API quota) can pause collection; when that happens the interface says so rather than showing stale or fake data.
Explore the source: Methodology (readable) · Technical methodology · Data health · Report an error / correction.