← Back

Methodology

Last updated: April 2026

Overview

Facthem EU monitors plenary and committee debates of the European Parliament and automatically fact-checks factual claims made by MEPs. The entire pipeline — from raw transcript to published verdict — runs without human editorial intervention. This page explains every step.

Step 1 — Session scraping and intervention parsing

Every session of the European Parliament is scraped directly from the European Parliament's official website. The raw transcript of each session is downloaded and split into individual interventions — one per speaking turn. Each intervention is matched to its MEP, along with available metadata such as political group, nationality, and session date. This structured representation is what feeds the rest of the pipeline.

Step 2 — Claim extraction, normalisation and enrichment

Each intervention is processed by a language model tasked with identifying discrete, falsifiable factual claims. Opinions, predictions, rhetorical questions, and normative statements ("we should…", "it would be better if…") are excluded — only claims that assert a specific, verifiable fact about the world are extracted.

For each extracted claim, the model then:

  • Normalises it into a canonical, context-independent sentence that can be evaluated without reference to the original speech.
  • Categorises it by topic (economy, defence, climate, health, etc.) and geographic scope.
  • Enriches it with structured metadata — thematic tags, entities mentioned, and any implicit numerical or temporal claims — that improve the quality of the verification step.

Step 3 — Automated verification via LLM + web search

Each normalised claim is passed to a language model with access to live web search. The model autonomously searches for evidence, reasons over what it finds, and produces a verdict. This is not keyword matching — the model reads sources, evaluates their relevance and reliability, and constructs an argument for or against the claim before reaching a conclusion.

A claim can only receive a definitive verdict if the evidence meets one of the following thresholds:

  • 1 primary official source — a public institution, official statistical body, or intergovernmental organisation that directly confirms or refutes the claim (Eurostat, ECB, OECD, EUR-Lex, World Bank, etc.), or
  • 2 independent secondary sources — separate media outlets, academic publications, or reputable reference bodies that corroborate the same finding without relying on each other.

If neither threshold is met, the claim is marked Unverifiable regardless of what the model believes to be likely true. The output of this step includes the verdict, a confidence score, the sources cited, identified omissions in the original claim, and any factual errors detected. All of this is published on the claim's page.

Verdict definitions

Each claim receives exactly one of the following verdicts:

Verdict Meaning
✅ Confirmed The claim is accurate and supported by at least one primary official source, or two independent secondary sources.
⚠️ Nuanced The claim is broadly accurate but omits important context, qualifications, or caveats that materially affect its interpretation.
🔸 Inaccurate The claim contains factual errors but is not wholly false — the underlying direction or intention may be correct while specific figures or details are wrong.
🟠 Overestimated The claim asserts a value or magnitude that is verifiably higher than what the evidence supports.
🟠 Underestimated The claim asserts a value or magnitude that is verifiably lower than what the evidence supports.
🟠 Out of context The claim is literally accurate but presented in a way that creates a misleading impression by omitting decisive context.
❌ False The claim is directly contradicted by primary source evidence.
❓ Unverifiable Sufficient public evidence does not exist to confirm or refute the claim at the time of processing.

Confidence score

Each verdict is accompanied by a confidence score from 0 to 100. This score reflects the model's internal certainty about the verdict given the available evidence — it is not an objective probability. A high score means the evidence was clear and consistent; a low score means the evidence was ambiguous, sparse, or conflicting.

As a rule of thumb: scores above 80 indicate strong evidentiary support; scores between 50 and 80 indicate moderate support; scores below 50 suggest the verdict should be treated as tentative. Unverifiable verdicts typically carry low confidence scores by definition.

Source hierarchy

Sources are weighted in the following order. When a claim can be settled by a primary source, that source is determinative and secondary sources are listed as supplementary only.

Primary sources — data and documents published directly by public institutions: the Official Journal of the European Union, Eurostat, the European Central Bank, and reports from the OECD, IMF, World Bank, and other intergovernmental bodies.

Academic sources — peer-reviewed publications in indexed journals. Weighted below primary institutional data but above general press.

Secondary sources — specialised reference press, independent analytical bodies, and reputable news organisations. Used when primary sources are insufficient or unavailable. Carry less weight in the final verdict.

A single primary official source is sufficient for a definitive verdict. Two independent secondary sources are required when no primary source is available. If neither threshold is met, the claim is marked Unverifiable.

Limitations

This pipeline is experimental. Language models can hallucinate sources, misinterpret numerical claims, fail to retrieve recent data, or misjudge the relevance of evidence. Every verdict published on this site should be treated as a starting point for investigation, not a definitive conclusion.

No verdict has been reviewed, corroborated, or validated by a journalist, legal expert, or any other human professional before publication. See the legal notice for the full disclaimer.

Structured data

Every claim page includes schema.org ClaimReview structured data, following Google's fact-check structured data guidelines. This allows search engines to surface the verdict directly in results.

Contact

To report an error in a verdict or suggest an improvement to the methodology, write to .

Legal notice  ·  Methodology  ·  About  ·  Blog  ·  All claims  ·  Contact:  ·  ♥ Support us: PayPal Ko-fi