Skip to content
FactCheck by GeriAI

Trial version · free, no sign-up

How much trust does this headline deserve?

Works best with facts ("X announces Y") rather than opinions, and without "according to X".

On your phone: Android iPhone and iPad

Time window
Fine-tune the analysis

With a link, the window is centred on the article's date; with a headline, it counts back from today. 14 days at most.

The first paragraph helps find the pieces that report the same thing.

Languages

No translation: in English and Portuguese it only searches for proper names.

See the full website
  1. Paste a link to an article or type the headline.
  2. GeriAI looks for everyone else reporting the same fact in the digital press.
  3. You get a trust score from 0 to 100 and every piece that supports or rebuts it.

It measures trust, not truth.

You give it a link to an article or a headline, with its standfirst if you like, and a time window. FactCheck looks through the digital press that TrawlingWeb processes for pieces that report that same fact, and works out how much trust it deserves from who reports it and who rebuts it. The article itself does not count as a source.

The scale runs from 0, when nobody else reports it, to 100, when many independent sources from several countries do.

0–25 Very low trust Not enough independent sources, or more sources rebut it than back it.
25–50 Low trust A few sources carry it, but few are independent of each other.
50–75 Moderate trust Several independent sources back it.
75–100 High trust Many independent sources from several countries, and they say it in their headline.

FactCheck never says "true" or "false".

A headline can earn little trust today and still be right, because it is recent or local. And a hoax copied a hundred times is still one source.

That is why the score counts independent sources, not repetitions, and leaves the verdict to the reader.

Nine readings. None decides alone.

As in a land survey, each reading is taken on its own and then they are combined. There are limits: without enough corroboration, or with more rebuttals than support, the score cannot be high.

How it works inside: the steps and the principles

No.ReadingWhat it looks at
1Independent sourcesHow many separate media groups report it. A hundred copies of the same wire story count as one.
2Support or rebuttalWhether each piece backs the fact, hedges it or contradicts it, explicit rebuttals included.
3CountriesHow many different countries report it. Fabricated facts rarely cross borders.
4Key terms in the headlineWhether the names and figures of the fact appear in each piece's headline, not just buried in the body.
5AudienceThe reach of each outlet reporting it.
6Source qualityA ranking of outlets by track record. Still in preparation: today every outlet weighs the same.
7SimilarityHow closely each piece matches the fact under analysis.
8Spread over timeWhether it appears gradually, the way a real fact travels, or all at once.
9First appearanceWhere and when it was first reported.

Build your own triangulation.

Tap each station to change what it says: supports, rebuts or stays silent. The band changes instantly.

Trust score

Very low trust

Independent sources
0
Countries
0
Support
0
Rebut
0
Limit
None

A1 and A2 belong to the same media group: together they count as one.

An illustration with made-up outlets: it shows the idea of triangulation, not the engine’s calculation, which is why it gives no number. It does not analyse anything real.

Every number comes with its evidence.

The result is not a bare grade. It lists every matching piece with its outlet, country and date, and says whether it supports or rebuts the headline. You can trace where every point comes from.

Models that compare, not models that make things up.

It is computed by GeriAI's own models, which compare texts, recognise names, places and figures, and judge whether a piece supports or contradicts. When a piece is unclear, a language model rereads it and only says whether it supports, contradicts or is neutral. None of them writes sources: each one is a real piece, with its link.

What it does not do yet.

  1. If a fact is told in very different words, it may miss some of the pieces.
  2. The ranking of outlets by track record is not ready yet.
  3. The headline has to state the fact, without "according to X".
  4. Tested in Spanish, English and Portuguese.
  5. Calibrated on real cases: a widely reported fact scores high; hoaxes already rebutted and a made-up story stay at the low end of the scale.
  6. It looks at 14 days around the headline at most.
  7. No translation: in English and Portuguese it only searches for proper names.

The research behind FactCheck

Óscar Trabazos Founder of TrawlingWeb and GeriAI

Since 2009 I have worked on how a machine can read what gets published and understand more than words: first sentiment, then irony, then emotions that do not match what is being said. In 2017 I began writing about hoaxes and, soon after, about how to detect them with data. Over the years I learned that the useful question is not whether something is true, but how much trust it deserves and who else is reporting it. FactCheck is that question turned into a tool.

What inspires it. Claire Wardle's seven types of misinformation, Eliot Higgins's «four Ps», and journalism's oldest rule: check against independent sources.

Analysis engine: GeriAI, by TrawlingWeb, led by Óscar Trabazos. Web design and performance: the CronosAI team.

For anyone who reads a headline.

For people who pause before sharing, for people who write and want to cross-check, for people who teach how to read the press.

It is a trial version, free and without sign-up. Analyses run one at a time and take 15 to 55 seconds; if there is a queue, we tell you your place.

Analyse a headline