Overview
Preslav Nakov presents the TANBIR project from Qatar Computing Research Institute, arguing that the key to winning the war on fake news lies in education and source credibility assessment. Rather than attempting to fact-check every individual claim in real-time (impossible given the scale and speed of spread), the approach profiles media sources in advance to estimate their factuality and bias, enabling "fact-checking before the news is even written." The talk covers three main innovations: a news aggregator detecting propaganda, media profiling tools, and fine-grained propaganda analysis techniques, with collaborations across academic institutions (Northwestern, MIT) and media outlets (Al Jazeera, Associated Press, RT).
Key Topics
The Education Answer
Finland's success in combating fake news stems not from technology but from media literacy education embedded in schools, teaching students critical thinking, image verification, and claim-checking. A recent European Union study found that 75% of young people don't recognize fake news, indicating the scope of the literacy gap. As Nazi propaganda minister Joseph Goebbels observed, "propaganda becomes ineffective the moment we are aware of it" — awareness raising is the fundamental defense. Simply having the discourse around fake news in mainstream conversation represents progress toward collective awareness.
The Speed Problem
A seminal study showed that fake news spreads six times faster than real news on Twitter and reaches far more users. More critically, 50% of the eventual spread of false information happens in the first 10 minutes. This timeline precludes traditional fact-checking approaches: journalists need time to investigate, and mainstream media must publish before their analysis reaches the public. Automatic fact-checking also faces barriers — many claims require observing user reactions on social media or waiting for mainstream coverage to assess claims. Time is the core constraint.
Profiling the Source: The TANBIR Approach
Nakov's core insight: fake news creators reuse the same set of websites rather than creating new ones for each piece of disinformation. By identifying and profiling these source websites in advance, the system can estimate credibility instantaneously when a new article appears, bypassing the need for content-level fact-checking. The TANBIR ("awareness/warning" in Arabic) project develops this strategy on two information layers:
Media Content Analysis:
- Measuring subjectivity and appeal to emotions
- Detecting moral categories (war, religion, identity)
- Analyzing linguistic style and framing
Social Media Profiling:
- Analyzing how media sources self-describe on social platforms
- Tracking social media accounts (Facebook, Twitter, YouTube)
- Assessing audience bias based on user engagement patterns
- Integrating Wikipedia information about the source
- Using traffic data (Alexa rank) and other signals
Classification Performance
The system classifies websites on two dimensions:
- Factuality of Reporting — accuracy and reliability based on content and reputation signals (baseline majority-class accuracy ~46%, achieves 80% accuracy)
- Political Ideology / Bias — left, center-left, center, center-right, right along a seven-point scale (baseline ~37%, achieves 82% accuracy)
Fine-Grained Propaganda Detection
Beyond source-level credibility, TANBIR detects specific propaganda techniques in text:
- Emotional appeals (appealing to prejudices and emotions)
- Logical fallacies (e.g., "with us or against us" false dilemmas)
These techniques are presented to users with explainability, teaching them to recognize manipulation in text themselves. The project ran successful hackathons on this and collaborates with media organizations to apply detection at the article and sentence level.
Media Profiles in Practice
Examples from the system:
- Sputnik: Low factuality, high propagandistic content
- RT: High bias and propaganda; notably polarizing by promoting both left and right causes to maximize societal division
- BBC: Central political position, low hyper-partisanship; analysis of 50,000+ articles shows 95.8% are very unlikely to contain propaganda
The system also profiles specific events and topics (Brexit, vaccines, climate change, Christchurch shooting) showing:
- % of content that is propagandistic
- Geographic and organizational distribution of propagandistic sources
- How topics are framed (e.g., Brexit 80% clean, vaccines 60% clean, Christchurch 30% clean)
- Whether coverage is driven by political, cultural identity, safety, or other frames
Public Tools and Integration
The TANBIR news aggregator is publicly available. As users read news articles, the system flags content likely to contain propaganda. An API is available for third-party integration (with performance agreements for heavy users). The system enables integration into news platforms, browser extensions, or media literacy tools to support critical thinking development.
Notes
A practically-oriented talk bridging NLP/computational methods and media literacy. Nakov's core contribution — shifting from claim-level fact-checking to source-level credibility assessment — addresses the speed-of-spread problem inherent in reactive fact-checking systems. The tripartite approach (news aggregator, media profiling, fine-grained propaganda analysis) reflects a coherent system design: first flag suspicious sources, then provide fine-grained evidence of manipulation techniques, both to support immediate filtering and to educate users in critical consumption.
The talk usefully distinguishes between user-facing tools (increasing awareness) and research/API tools (supporting integration), noting that both serve the education goal. The emphasis on education and media literacy aligns with evidence from Finland that systematic literacy instruction outperforms technological solutions alone. Notably prescient on platform vulnerabilities and the necessity of transparency in content curation algorithms.