LE QUOTIDIEN DU SOIR · 12 MARCH 2026 · FICTIONAL EXAMPLE
One extra hour of sleep makes adults 4% more attentive
Sleeping one hour more increases attention by 4% in adults, according to an open-access study.
Loading
SYNESIATools and trainingagainst disinformation
SYNESIA is a free European project that does two things: tools to find, compare and analyse disinformation stories, even when they are written in another language, and training to learn to recognise them.
SYNESIA — Systemic Network for European Scrutiny & Information Architecture
In short
What you can do on SYNESIA
You bring the material: a link, a PDF, pasted text, or the page you are reading through the browser extension. Each tool shows its result and says what it does not know. A free Mythodologie account is all you need.
Find
Paste a text: SYNESIA finds, among the documents members have added, those that tell the same story, including in another language.
Check
Look into a source, pull out the claims of an article, compare two versions of a story, spot near-identical posts published in bursts, put events back in order.
Learn
Training modules, educational games and a weekly challenge to learn, in class or on your own, how manipulation techniques work.
FreeMultilingual6 partner countriesResults explained
The Challenge
False news spreads 6 times faster than the truth
With every share, the message distorts. Nuances disappear, claims become extreme, and original context is lost. Europe faces an unprecedented challenge to its information ecosystem.
0%
more likely to be shared than the truth
0%
of shared news links are never clicked
0%
actually check who is behind what they read
StudyPressSocial MediaViral
Scripted demo
The analysis in four steps
A scripted animation on invented examples: nothing here is analyzed live. Its four steps show how you add material (link, PDF, pasted text, browser extension), how the grouping engine links the same story across languages (in the pages members add, or failing that a built-in demo corpus), then, from what you supply, how to chart a claim's spread and flag likely coordinated posts (near-duplicates, timing bursts).
Cross-border propagation
How a single narrative crosses borders
Disinformation rarely stays in one country or language. This animated globe illustrates the idea: each arc stands for a narrative moving from one country to another. It shows no data: the arcs are drawn at random between a fixed set of capitals, and no Synesia tool feeds it.
Europe · illustration, not data
Capital (example)
Narrative spread
Citation traceability
What the cited source actually says
The cited document exists, it is open access, it can be found. The jump between what it says and what the article makes it say is written nowhere — so someone opens it, sets both sentences in the same measure, and records what changed. Grouping these echoes across languages and putting them in date order is not built in Synesia yet.
Worked example invented for this demonstration: no real outlet, journal, study or person. The identifier uses 10.5555, the DOI Foundation's reserved test prefix, and is deliberately not a link.
Invented example
Citation traceability, invented example. A fictional daily writes that sleeping one hour more increases attention by 4% in adults. A verifier opens the cited document, open access, under a fictional identifier: the document says one extra hour of sleep is associated with +4% attention among 214 pupils followed for six weeks, and adds that it cannot rule out other factors. The number is the same; the object, the scope and the verb have changed, and the caveat is absent. The verifier records the relation: the source partially supports. On a second claim from the same article, the same procedure returns: the source supports. Synesia groups 24 articles citing this document across four languages and orders them over eleven days: the verb moves from "is associated with" to "proves", the caveat drops at the institutional press release and never returns, and the walk back links the last echo to the document in seven hops. On the chain, two of four links resolve; the article cites another article, with no identifier. Over the 24, the distribution is: supports 7, partially supports 9, does not support 4, contradicts 1, cannot be determined 3. None of these values says whether the claim is true. In this section the source-claim relation is recorded by a person; a model in the product can compare a claim against a supplied passage, but it does not distinguish partial support. What has to be built: a persistent identifier on every citation, open access to the cited document, a public retraction register, a press release that adds nothing to the study. Necessary, but not sufficient.
COMPUTED BY SYNESIAcross-lingual grouping of the echoes · 384 dimensions · functionalordering the echoes in time · functional
RECORDED BY A PERSONarticle ↔ source relation · recorded by a persontraceability record · WP2 · roadmap
Here, the source ↔ claim relation is recorded by a person. A model in the product can compare a claim against a passage you supply it — supports, contradicts, cannot be determined — but it does not fetch the source for you, and it does not see partial support, which is the most common case.
The passage
What the article claims, against what the cited source actually says.
CITED DOCUMENT · OPEN ACCESSdoi:10.5555/rosa.2026.0114 (fictional)opened by a verifier · 2026-03-12 14:07
Sleep and classroom attention: a pilot cohort of 214 pupils — Revue ouverte du sommeil et de l'apprentissage (fictional journal)
"Among the 214 pupils followed for six weeks, in March and April, one extra hour of sleep is associated with +4% attention. We cannot rule out that other factors explain this gap."
DOCUMENT
Among the 214 pupils followed, one extra hour of sleep is associated with +4% attention.
We cannot rule out that other factors explain this gap. absent from the article
ARTICLE
One extra hour of sleep increases attention by 4% in adults.SCOPE CONDITION REMOVED
number unchanged
- object
- 214 pupils followed → adults
- scope
- six weeks, March-April → (removed)
- verb
- is associated with → increases
coral marks a changed word or a missing link, never a false one
RELATION BETWEEN THE TWO DOCUMENTSRECORDED BY A VERIFIER, PASSAGE ATTACHED
- =the source supports
- ≈the source partially supports
- ≠the source does not support
- ×the source contradicts
- ?cannot be determined
The document does not say "increases": it says "is associated with", among 214 pupils, over six weeks, and it adds a caveat the article does not carry over.
Same procedure, two outcomes.
Cascade ledger · 24 citing articles · 11 days · one tick = one article↑ any tick is an entry point
Register detectionA · ASSERTIVE REGISTER43 words · 272 chars B · TRACEABLE REGISTER66 words · 378 chars
The same fact, two wrappings
Two texts argue the same position about the same fact. In each one we count the entries of three published word lists, scaled to 1,000 words. What differs is not the position — it is what the text leaves you to check with.
Word lists, surface statistics and language identification: published code from Synesia's Influence module (stores/influence.ts). No trained model and no accuracy figure — this module is not a detector and carries no tool status badge.
frendeuk
POSITION BOTH TEXTS ARGUE
The 12 August vote on the low-emission zone was timed so as to limit turnout.
Invented texts. No truth score is computed.
Register detection, invented texts. Two texts argue the same position about a vote timed for August. Text A is written in an assertive register: capitals, exclamation marks, an emoji, and entries from the alarmist and suspicious lists. Text B is written in a traceable register: a resolution number, a time, a vote count, figures, a quotation, and entries from the hedging list. Counts are scaled to 1,000 words because the two texts differ in length. The count is of entries present in three published word lists, across four languages, with no syntax, no lemmatisation, no negation and no irony. Where this count fails: a genuine civil-safety alert saturates the alarmist list; a perfectly polished misrepresentation of a study triggers no alarm at all; legitimate anger has the register of a fake. A register count does not say whether a text is true, and it never will. No truth score is computed. Necessary, not sufficient: it also takes open fact-checking databases, lateral reading taught in class, and collectives that verify together.
URGENT 🚨 The low-emission zone vote went through on 12 AUGUST. Mid-holiday, when nobody is watching. Obviously deliberate. They lie to us about the real target: dodging the debate. Banned from the big outlets, clearly. Total scandal!! Share before it disappears. Wake up.
ALARMIST4→93.0/1000
SUSPICIOUS3→69.8/1000
HEDGING0→0.0/1000
SURFACECAPS 7.4 %·! 2·" " 0·numbers 1
Resolution no. 2026-214 was adopted on 12 August at 6.40 p.m., by 23 votes out of 61 seats; the minutes are available at the town hall. According to the county mobility observatory, the measure may cover about 18 % of the local fleet; the council, however, puts it at 12 %. Some councillors therefore contest the timing: "an August vote reduces turnout", the research note says.
ALARMIST0→0.0/1000
SUSPICIOUS0→0.0/1000
HEDGING4→60.6/1000
SURFACECAPS 1.3 %·! 0·" " 1·numbers 8
0 220 entries per 1,000 words
Both texts argue the same position. What changes is what they leave you to check with.
43 words / 66 words. Every count is scaled to 1,000 words.
Word lists, surface statistics and language identification: published code from the Influence module (stores/influence.ts). No trained model, no accuracy figure.
How it Works
From raw data to actionable intelligence
Nothing is collected on a schedule: you supply the material — a link, a PDF, a DOCX, pasted text, or the current page via our browser extension. A multilingual model then finds, across languages, the texts closest in meaning to yours, among those members have contributed.
01
Collect
Members add the material by hand: link, PDF, DOCX, pasted text, extension. Google News watch on demand. No scheduled crawling, no social connector.
Google News RSS (read-only)Browser extensionURL / PDF / DOCX
02
Analyze
Claim extraction by a self-hosted language model, with FR/EN keyword patterns as a fallback (the method is shown with each result). Similarity compares your text to contributed narratives — a semantic distance, not a threat level. No manipulation detector.
NLP / LLMsEmbeddingsSimilarity
03
Cluster
The grouping engine is real (multilingual embeddings, hierarchical clustering): it brings together the pages members add when they tell the same story, across languages, and runs on a built-in demo corpus while there are none. Nothing is tracked over time.
Cross-lingualHierarchical clusteringTemporal
04
Read
Each tool shows its result on screen, with what it knows and what it ignores. No alerts, no automated reports, no email or webhook delivery.
On-screen resultsKnows / ignoresNo auto-send
Use Cases
Built for the people on the front lines
Whether you're debunking a viral claim, investigating a coordinated campaign, or teaching media literacy in a classroom, Synesia gives you the tools.
Researchers
"I need to track how a climate denial narrative spreads from Russian-language Telegram to French Facebook groups."
- Search by meaning, from one language to another
- Temporal propagation analysis
Journalists & Fact-checkers
"A viral claim is spreading in 4 languages. I need to find the original source and trace how it was distorted."
- Citation tracing, shown on a demo example
- Claim extraction by a language model, with a keyword-pattern fallback
- Coordinated-content triage (near-duplicates, timing bursts)
Educators
"I want my students to understand how disinformation works, with interactive simulations they can try themselves."
- Interactive prebunking games
- Disinformation chain simulator
- Structured training curriculum
NGOs & Institutions
"We follow a narrative around an election and want a documented file: which page was published where, who added it and when, and what the cited sources actually say."
- Keyword watch on Google News, on demand
- Documents added to the shared corpus, with provenance
- Coordinated-content triage on posts you supply
Roadmap
Where we're going
Synesia is being built in the open. Here's our journey from prototype to a full-scale European platform.
Phase 1Completed
Research & Prototyping
Cross-lingual embeddings research, NLP pipeline architecture, narrative clustering algorithms, initial prototype with 4 analysis engines.
Phase 2In Progress
Platform & Consortium
Building the web platform, assembling the 6-country consortium, developing the educational module, and securing Creative Europe funding (selected 2026).
Phase 32027
Pilot Deployment
Live monitoring across 6 countries, training workshops with journalists and educators, first public threat intelligence reports, API access for partners.
Phase 42028
Scale & Open Access
Open-source release of core tools, expansion to 12+ languages, partnership with EU institutions, integration with existing fact-checking networks.
The Consortium
6 countries, 6 organizations, 1 mission
A European partnership uniting expertise in media literacy, fact-checking, investigative journalism, and AI research, backed by 13 associated partners that extend the network across 13 countries.
FR
Mythodologie
Coordinator — AI & Platform
LT
15MIN
Fact-checking & Media
RO
Atlatszo Erdely
Investigative Journalism
UA
SUNPU
Academic Research
ES
Lichen
Media Literacy Education
ME
Atlantic Council of Montenegro
Digital Forensics
Open CollaborationGet Involved
Be Part of the Solution
Disinformation doesn't stop at borders. Neither should we. We're building an open European coalition of researchers, journalists, educators, and technologists.
Researchers
Journalists & Fact-checkers
Educators
Developers
NGOs & Funders
or write to us at contact@mythodologie.fr
SourcesMIT, Vosoughi, Roy & Aral, Science (2018). Columbia & Inria, Gabielkov et al., "Social Clicks", SIGMETRICS (2016). Stanford History Education Group (2021).