Radar Aos Fatos is a real-time monitoring system built to fight disinformation. Its public facade consists of a monitor that automatically detects potentially misleading contents that are trending on social platforms such as Twitter, Facebook, Instagram, YouTube and WhatsApp. Everyday, publications are swept by an algorithm that maps language patterns and classifies them according to their informative quality. Thus, it is possible to observe the evolution of disinformation campaigns and how they are amplified. By doing this, it analyzes an average of 200.000 publications weekly applying a methodology that combines linguistics, social and data science.
In a country where the most prominent politicians use disinformation as a political tool, Radar Aos Fatos produces relevant journalism that makes those in power uncomfortable. Since its launch, in March 2020, Radar’s reporters have already published more than 30 stories showing that the most pernicious disinformation campaigns in Brazil come from the top: president Jair Bolsonaro and his allies are the most common misinformers on social media.
This work motivated multiple lawsuits against Aos Fatos from bolsonaristas, such as one brought by federal deputy Bia Kicis (PSL) and another one by the public prosecutor and right-wing influencer Ailton Benedito. Aos Fatos’ team has been doxxed and attacked in social media provoked by Bolsonaro’s allies.
Besides that, Radar’s editorial branch published high impact stories, shared by over 20 media outlets, including Folha de S.Paulo, O Globo and Veja. For instance, we found out that a representative who was being considered for Secretary of Health was the biggest disinformation spreader about the pandemic (https://www.aosfatos.org/noticias/cotado-para-saude-osmar-terra-e-congressista-que-mais-difundiu-desinformacao-sobre-coronavirus-no-twitter/). After this story, his name stopped being considered for office.
Radar Aos Fatos won the 2020 Gabriel García Márquez award on innovation and the Digital Media LATAM 2020 for best digital project.
To build Radar’s algorithm, we started by identifying patterns of misinformation on longform texts. It was needed to develop a parser to apply linguist rules such as alarmist text (alarmist words like “urgent”, “calamity”), exaggerated or generalistic text (“the best”, “100% cured”) and texts strings that frequently appears on misinformative texts (“vaccine from Israel”, “hidroxycloroquine with azithromycin”).
Then, this set of rules was applied and a total score was calculated. To create a more approachable score, two statisticians translated it to a scale from 1 to 10, where 1 means low quality and 10 is high quality content. For the purpose of Radar, we only display texts that scored 5 or less.
The same process of recognizing linguistic patterns was applied on posts on Twitter, YouTube, WhatsApp, Instagram and Facebook. We used their public APIs to collect text with a scrapper written in Python, using a Django web framework and a Celery task manager. Data is being saved to a PostgreSQL database, which also helps us filter content we need.
Even though the data scraping is pretty much the same, each platform has its set of rules to recognize potentially misinformative patterns.
During this process, we created a dataset of websites (over 400 websites that published at least one misinformative piece), Twitter handles (over 7,000, containing politicians, official governmental profiles), Facebook’s public pages (over 600 that have ever shared a content debunked by Aos Fatos), WhatsApp public groups and YouTube channels (over 650). This is an ongoing process and we rely on Aos Fatos daily debunking and monitoring.
Aos Fatos also created an editorial branch for Radar. Previously planned for the second semester, Covid-19 demanded a change of plans, and a team of journalists and data scientists was hired in March 2020.
What was the hardest part of this project?
A project this complex lacks references. It’s unique from backend development to frontend render. This specificity becomes extra challenging during a pandemic, because it forced our whole team to work from home, hindering creative and collective construction processes.
Also, Radar’s monitor aims to garner attention from audiences from different backgrounds: regular Aos Fatos’ readers, data scientists, scholars, journalists, political analysts. That’s why it is designed to provide an unique but also granular dataset about mis/disinformation in Brazil. The tool itself provides general numbers of our dataset, as well as a broad overview on over 200.000 publications in small charts. Radar invites new users to dig deeper in each social network and explore, by themselves, the different modulations of potentially misleading, low quality content.
This simple yet unique dataviz demands a deep understanding of our own dataset and our abilities to collect, analyze and display a large amount of data that is updated constantly. Plus, since it’s an ongoing project, both layouts and database needed to be planned for reasonable growth, which means more storage and processing power, a flexible layout and programmable charts. Yet, the tool need to fully load within 5 seconds, otherwise we will lose our readers.
What can others learn from this project?
By having access to sophisticated analysis about the landscape of broad political mis/disinformation strategies, Aos Fatos’ editorial team and other journalists now have a tool to act faster against fraudulent content campaigns. By looking at most common searches and cross-monitoring viral content in many platforms simultaneously, it is easier to detect misinformation patterns and act against it. As a broader result, the accuracy of disinformation investigations are higher, and so is their impact.