How we verify political claims
We believe you should know exactly how our system works, what sources it uses, how we guarantee neutrality, and where AI has limitations.
Our Methodology
Every claim goes through a 6-step automated pipeline. No human editorial decision is involved in the verification process.
Transcription
Political speeches are transcribed using AI (Whisper) and split into thematic segments.
Context Analysis
Speaker, setting, and date are identified to properly contextualize each statement.
Claim Extraction
AI identifies verifiable factual claims, filtering out opinions, promises, and satire.
Source Research
Up to 15 search iterations per claim across government, academic, and institutional sources.
Dual Verification
Two independent AI models vote separately - like double-blind peer review in science.
Scoring & Verdict
A validity score (0–100%) determines the verdict on an 8-level scale, from True to False.
Our Sources
We use a curated database of sources ranked by reliability tier. Each source type has a weight that influences the final verdict score.
Tier 1 - Official
Highest reliabilityGovernment databases, official statistics offices (OFS, Eurostat), parliamentary records, legal texts.
Tier 2 - Institutional
High reliabilityInternational organizations (WHO, UN, OECD), national banks, regulatory agencies.
Tier 3 - Academic
High reliabilityPeer-reviewed publications, university research centers, scientific databases.
Tier 4 - Reference Media
Moderate reliabilityMajor wire services (AFP, Reuters, AP), quality newspapers of record.
Tier 5 - Other Media
Lower reliabilityRegional press, specialized outlets, investigative journalism organizations.
Sources are never mixed: official data is always weighted more heavily than media reports. Every verdict links to the specific sources used.
Political Neutrality
OpenTruth is not a media outlet. We don't publish opinions. We structure political speech into verifiable data.
No editorial line
Our AI applies the same process to every politician, every party, every statement. There is no editorial board deciding what to check.
No political affiliation
OpenTruth has no ties to any political party, government, or lobby. We are a Swiss company funded by private capital.
Open methodology
Our pipeline, scoring formula, and source weights are publicly documented. Anyone can audit our process.
Traceable verdicts
Every verdict includes the full chain: original claim, sources consulted, NLI scores, and reasoning. Nothing is a black box.
Honest Limitations
AI is powerful but imperfect. We believe in being transparent about what our system can and cannot do.
What AI does well
- Process 50+ claims in under 5 minutes
- Find official and institutional sources with high precision
- Apply an identical, unbiased process to every single claim
- Operate in 4 languages (French, German, Italian, English)
Where AI struggles
- Irony and satire - may confuse a joke with a factual claim
- Implicit context - misses nuances only a subject expert would know
- Complex statistics - multi-variable comparisons are harder to verify
- Emerging events - very recent events may lack source coverage
Our main model achieves 82.3% accuracy on French-language benchmarks. On political text specifically, we estimate 77–80% accuracy.
To mitigate these limitations, we use dual independent models, always display confidence scores, provide direct source links, and allow journalists to contest and correct verdicts.