Research / Disinformation
Summary of a research line · 2017 – todayDisinformation: from hunting falsehood to measuring trust
I started out asking how to detect a fake news story. Almost ten years later, the question I find useful is a different one: how much trust a piece of information deserves, who else is reporting it and with what evidence.
This page summarises what I have published on the subject and what has been built and measured in FactCheck. The full method and evaluation are in the working paper Measuring corroboration, not truth (2026, not peer reviewed).
What is new is not the lie, but how it spreads
Deception has always existed, and it has always been fought by looking for better sources. What social networks changed is speed: anyone can publish, nobody has to check first, and whatever fits what the audience wants to hear multiplies on its own. A newspaper used to investigate before publishing because its reputation depended on it; today a story can go viral without anyone corroborating it.
That is why this line does not study only the text of a piece of information, but its journey: who said it first, who repeats it, whether those repeating it are independent of each other, and whether anyone denies it.
Five stages
Most of the articles are in Spanish.
-
2009 – 2014
The foundations: understanding tone
Before talking about disinformation, something more basic had to be solved: getting a machine to understand the tone in which something is said. At Brand Rain and Anpro21 we worked on active listening, sentiment analysis and irony detection to measure brand reputation. That led to the 2014 paper in the SEPLN journal and to an idea that comes back in everything that followed: what is said is not always what is meant.
-
2017 – 2018
The problem: disinformation and democracy
After the 2016 US election I wrote about how disinformation shapes political debate and why the problem is the spread, not the deception itself. My conclusion then still holds: two responses are needed at once, citizens with more critical thinking and information systems able to slow the spread. The academic work on automatic detection at the time (listed in the references) already pointed to what we know today: text alone is not enough; you have to look at the source, the network and the spread.
-
2023
Detection: headlines, classifiers and the impostor technique
With generative AI on the horizon, I published a series on detection. The first part dealt with semantic headline detection: comparing what the headline promises with what the text says. The second proposed a two-stage neural classification. The third combined Claire Wardle's seven types of misinformation and Eliot Higgins's «four Ps» with what I called the impostor technique: locating «zone 0» (where and when something was first reported), weighing the authority of whoever says it, and having one AI classify while another questions what the first concludes.
I tested it on a real case: Sam Altman's firing from OpenAI, in November 2023, with a verification prototype built on TrawlingWeb data. These were methodological proposals and prototypes, not a measured system; measuring came later.
Part I (English) · Part II · Part III · The Sam Altman case (English)
-
2024 – 2025
New fronts: synthetic video and newsrooms
Sora changed the rules: if a believable video takes minutes to make, the coherence between text, image and emotion becomes one more signal. I also wrote about how generative AI affects journalism and proposed a «topic safety» tool for writers, with three axes that are still part of the method: reach, credibility («who else is talking about it?») and the authority of the source.
Sora and disinformation · Generative AI and journalism · Topic safety
-
2026
The turn: we do not look for truth, we measure trust
Keyword-based detection is over: a generative model writes with whatever vocabulary it needs. What can give away an artificial flow is its patterns: speed against volume, how alike the accounts spreading it are, sources with no track record. The question stops being «is it false?» and becomes «how much trust does it deserve, and on what evidence?». We provide the map; the judgement belongs to the reader.
Anatomy of digital deception (English) · Disinformation as a risk · Verification in the age of generative models
Four starting ideas
They guide the design of FactCheck. They are working hypotheses, not proven conclusions: the first one, for instance, has not been measured yet.
The spread may give it away
The hypothesis: real information usually spreads in stages, while fabricated information appears all at once, in many places, with no precursor. FactCheck gives it a low weight until it can be measured.
A hundred copies are one source
A wire story reproduced by a hundred outlets counts once. Trust comes from independent sources, in different places.
A binary verdict fails
Something true but recent or local may have little corroboration; a widely copied hoax, plenty of repetition. Saying «true» or «false» misleads in both cases.
Evidence in plain sight
A number without evidence asks for faith. Every score must come with the pieces that support or contradict it, so the reader can judge.
FactCheck: a trust score with its evidence
FactCheck applies this research on the GeriAI analysis engine, TrawlingWeb's intelligence layer. You give it a link to an article or a headline and it returns a 0-100 score: how much trust it deserves based on how well independent sources in the digital press corroborate it. It never says «true» or «false».
- Nine signals: independent sources, support or rebuttal, countries, key terms in the headline, audience, source quality, similarity, spread over time and first appearance. The weights used to combine them are provisional: they are not yet calibrated on data, and they are not published.
- Hand-set caps: with fewer than withheld independent sources, or if none supports it, the score cannot exceed withheld; if more sources rebut than support it, or fewer than withheld of evaluated pieces support it, it cannot exceed withheld.
- One model classifies, another reviews: a local model decides whether each piece supports or contradicts; unclear ones and contradictions are reviewed by a language model that only classifies. It is a version of the 2023 idea. In a test on 251 pairs of real headlines, the language model scores 0.94 against 0.78 for the local model. It is a preliminary measure: some of the labels were set by an AI and have not yet been reviewed by a person.
- Tried on real cases: in about ten cases, a fact reported by many independent outlets scored high (around 80); an already debunked hoax stayed at withheld and a made-up story at withheld, because of the caps. These are examples, not a systematic evaluation.
The marked figures are withheld to prevent stories from being crafted to score high. Researchers who need them can request them through the contact page.
It also says what it does not do yet: if a fact is told in very different words it may miss pieces, the track-record rating of outlets is not ready, and it does not analyse images or video. The method, the signals and the evaluation are explained in the working paper.
FactCheck by GeriAI Free, no sign-up: paste a link or a headline and see how much trust it deserves. Try FactCheck →What comes next
- Zone 0, looking back: finding the real first appearance of a piece of information outside the analysed period, not only within it.
- Source authority: whether the outlet or author already covered the topic before, and how consistently.
- The effect on people: how what we believe and share changes depending on how we process information. There are studies on how we judge credibility, but no work of my own yet; it is the next question in this line.
Where it has been applied
Disinformation and fake news
Impostor technique, discriminative and adversarial AI, natural language processing, deep learning and speech recognition models.
Anti-fraud for editorial news
Big Data, natural language processing, entity relations and cloud data analysis.
References
- Claire Wardle and Hossein Derakhshan, Information Disorder: Toward an interdisciplinary framework for research and policy making, Council of Europe, 2017.
- Eliot Higgins and the «four Ps» of disinformation: passion, politics, propaganda and pay.
- Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang and Huan Liu, «Fake News Detection on Social Media: A Data Mining Perspective», ACM SIGKDD Explorations, 2017.
- Eugenio Tacchini et al., «Some Like it Hoax: Automated Fake News Detection in Social Networks», 2017.
- Hadeer Ahmed, «Detecting Opinion Spam and Fake News Using N-gram Analysis and Semantic Similarity», thesis, University of Victoria, 2017.
- Journalism's oldest rule: check against independent sources before publishing.