Our Methodology
How OpenTruth verifies political discourse
We do not evaluate political discourse. We structure it as data.
OpenTruth is an automated fact-checking infrastructure. Our pipeline analyzes videos of political speeches, parliamentary debates, and media appearances to extract, verify, and score factual claims.
Claim Classification
11 claim types (I1-I11): Attribution, Event, Factual Definition, Interpretation, Legal, Scientific, Statistical, Comparative, Predictive/Promise, Historical, Causal
24 themes (D1-D24): Politics, Economy, Health, Environment, etc.
When a claim falls under multiple types, we retain the one that determines the verification strategy.
Ingestion and Transcription
Each video is automatically transcribed (YouTube Transcript API or OpenAI Whisper). Raw text is segmented into coherent thematic units through automatic topic change detection.
Claim Extraction
AI agents (GPT-5) identify verifiable claims in each segment. Each claim is classified according to our taxonomy of 11 types and 24 themes.
Source Research
For each claim, our agents search for a minimum of 4 independent sources following a reliability hierarchy from public institutions to media outlets.
Correspondence Evaluation
Each claim-source pair is evaluated along three quality dimensions: independence, proximity to primary source, and methodological quality.
NLI Scoring
A language model (mDeBERTa-v3, multilingual) automatically evaluates the logical consistency between each claim and its sources.
Verdicts
Each claim receives a verdict on an 8-level scale, from True to Satire, based on the combined NLI scores and source reliability.
Source Research
For each claim, our agents search for a minimum of 4 independent sources following a reliability hierarchy:
Public institutions
95%gov websites, EU, UN
Scientific publications
90%PubMed, CNRS, universities
Official databases
85%INSEE, Eurostat, FSO
Media
60%AFP, Reuters, Le Monde, BBC
NGOs and think tanks
55%Amnesty, Brookings
We systematically prioritize primary sources. Anonymous or unverifiable sources are excluded.
Correspondence Evaluation
Independence
Is the source independent from the subject matter?
Proximity to primary source
Primary (1), secondary (2), or tertiary (3)?
Methodological quality
Is the methodology transparent and verifiable?
Correspondence is classified as: confirms, contradicts, partial, or no data.
NLI Scoring
Natural Language Inference
A language model (mDeBERTa-v3, multilingual) automatically evaluates the logical consistency between each claim and its sources. The score combines:
NLI result (entailment/contradiction/neutral)
Source reliability (weighted by G1 hierarchy)
Semantic relevance (cosine similarity of embeddings)
Verdicts
Each claim receives a verdict on an 8-level scale:
Confirmed by at least 2 concordant reliable sources
Broadly confirmed with minor nuances
Contains both true and false elements
Technically accurate but missing or distorted context
Contradicted by reliable sources
Cannot be confirmed or refuted with available sources
Value judgment, not factually verifiable
Humorous/satirical content identified
Transparency and Limitations
What we do
- We publish all sources used with their URLs
- We apply the same standards to all political actors
- Our verdicts are interoperable with the ClaimReview standard (schema.org/Google)
What we do not do
- We do not take political positions
- We do not censor or recommend censorship
- We do not claim infallibility — our models have known limitations
Known limitations
- Detecting "partially true" claims is our biggest technical challenge
- Sources in uncovered languages (outside FR/EN/DE) may be missed
- The pipeline is optimized for Swiss and French political discourse
Corrections
If you identify an error in our verifications, you can contact us at contact[at]opentruth[dot]ch
We commit to:
- Correcting any verified error within 48 hours
- Publishing a visible correction note on the affected verification
- Documenting corrections in a public annual register