Detecting the Fake News at Its Source, Media Literacy, and Regulatory Compliance

Definitions and scope

Misinformation vs. disinformation vs. propaganda:

• Misinformation — false information spread unintentionally, without intent to harm
• Disinformation — deliberately false information, intentionally created and spread to cause harm
• Propaganda — expressions deliberately designed to influence opinion and action, with intent to persuade (but not necessarily false; can be true statements deployed manipulatively)

The talk emphasizes that while falsity matters, the more important criterion is harm: whether information is designed to manipulate, mislead, or damage society, a person, or a community. Example: cherry-picked facts (all true) presented to distort perception are as dangerous as outright falsehoods.

Why fake news is a problem

Speed and scale: False information on Twitter spreads six times faster than truth. More critically, 50% of the viral spread of a false claim happens within the first ten minutes—before fact-checkers or platforms can intervene. This means reactive fact-checking cannot keep pace. The problem became acute with social media's ability to micro-target false claims at specific demographics (by location, age, gender, inferred interests) using psychological profiling and advertising infrastructure.

Real-world harms:

• Political: Fake news weaponized in elections (2016 U.S., others), with candidates/campaigns targeting different messages to different groups.
• Medical: Conspiracy theories and anti-vaccine misinformation drove a 300% spike in measles outbreaks globally; measles re-emerged in developed countries (Greece, Czech Republic, UK) after eradication.
• Violent: In India, false claims on WhatsApp and Instagram led to mob violence and deaths. A UN report blamed misinformation for the Rohingya genocide.
• Financial: In 2013, a hacked AP Twitter account posting false claims about explosions at the White House caused a $136 billion market drop.

Fact-checking: necessary but insufficient

Fact-checking organizations (Snopes, FactCheck.org, PolitiFact, etc.) have checked thousands of claims. However:

• Scale problem: A fact-checking task can take weeks. Checking ~12,000–33,000 claims per organization over years is a drop in an ocean of online claims.
• Speed problem: Given the 10-minute viral threshold, even fast fact-checkers cannot catch false information before it spreads.

Solution: Shift from checking claims to checking sources — profile news outlets themselves by factuality and bias.

Source-level detection: media profiling

The key insight: you can predict whether a claim from a particular outlet is trustworthy by profiling the outlet's entire track record.

What Taqaddum analyzes about news outlets:

1. Published content — what the outlet writes (language, framing, emotional appeals)
2. Social-media presence — Twitter account description, verification status, audience composition (left/right bias)
3. Audience demographics — inferred from Facebook/Twitter follower profiles
4. Wikipedia coverage — how the outlet is described externally
5. URL structure — suspicious domains (abcnews.com.co, angry-patriot-movement.com) are red flags
6. YouTube channels — speech, emotional tone, inflammatory language

Outputs:

• Political ideology — predicted on a 3-point (left/center/right) or 7-point scale, achieved 81% F1 by combining article content, Twitter audience bias, and Wikipedia context
• Factuality/reporting quality — 3-point scale (high/mixed/low)
• Source bias — general bias and topic-specific bias (how does this outlet cover gun control, immigration, climate change, COVID-19?)

Key finding: Extreme outlets (far left or far right) tend to have both high bias and low factuality. Outlets closer to center tend to be more trustworthy. This allows modeling factuality and bias jointly.

Media literacy and propaganda techniques

Media literacy strategy: Make readers aware that they are being manipulated. A quote from Mahatma Gandhi's tomb inspired the approach: "Don't listen to rumors, but if you do, don't believe it." The idea is that awareness breaks propaganda's spell — propaganda becomes ineffective once you know it's being deployed.

Fine-grained propaganda technique detection:

Nakov's team annotated 350,000 words of text with ~18 propaganda techniques and built systems to detect them automatically. Techniques include:

Emotional appeals:

• Appeal to fear
• Appeal to emotion / loaded language
• Flag waving (emotional nationalism)
• Appeal to authority
• Glittering generalities

Logical fallacies:

• Bandwagon ("most people believe X, so join us")
• False dilemma
• Name-calling
• Exaggeration / minimization
• Rumor-mongering
• Oversimplification

Example detections: A statement like "We will pursue any aggressor" with destruction appeals and world language (coded meanings) can be tagged with technique spans, not just whole-document labels.

Difference from hate speech: Propaganda analysis is distinct from toxicity or hate-speech detection — it focuses on persuasion techniques and influence tactics, not merely offensive language. A statement can be true, non-offensive, and still propaganda if it employs manipulative techniques to shape opinion.

Detecting media bias

Approach: Analyze news outlets' framing of topics. For COVID-19, the same situation (high death rates, economic disruption) can be framed through political, economic, health, or moral lenses. Different outlets emphasize different frames, revealing bias.

Example: Two headlines about COVID-19 may both be factual but convey opposite impressions:

• "COVID-19 is much more lethal than the flu" (emphasizing severity)
• "We're vastly overestimating flu deaths vs. COVID-19" (downplaying severity)

Both can be simultaneously true early in a pandemic, but the framing—and the outlet's consistent choice—reveals bias.

Regulatory compliance and automated tools

Industry need: Media outlets themselves need tools to self-monitor their content and catch potential problems before publication or after. The Taqaddum project includes:

1. Claim detection and fact-check linking — upload an article/video; the system identifies claims and links them to existing fact-checks, alerting editors to known falsehoods.
2. Bias reporting — shows a media outlet its own political bias and how it differs across topics.
3. Graphic-content detection — flags sensitive visuals.
4. Propaganda analysis — fine-grained propaganda technique detection on submitted articles.

These tools are exposed via APIs and browser extensions, allowing newsrooms and researchers to integrate them into workflows.

COVID-19 infodemic

Nakov's team annotated tweets about COVID-19 across multiple languages (English, Arabic) with a multi-perspective annotation schema. Annotation questions:

• Does the tweet contain a verifiable factual claim?
• Is the claim false?
• Is it of interest to the general public?
• Is it harmful? (To public health, to society, to a person or group?)
• Is it worth reporting?

Examples of COVID-19 misinformation identified:

• Drugs (false cures like garlic water or bleach)
• Rumor-mongering (military helicopters spraying pesticides; Bill Gates microchip vaccines)
• Medical disinformation (conspiracy theories about vaccine components)
• Advice and recommendations (legitimate but unverified; needs expert vetting)
• Panic-spreading (apocalyptic framing: "Americans are hoarding guns")
• Blame assignment (criticizing authorities or specific groups)
• Action calls (e.g., "reopen barbershops and salons")
• International comparisons (e.g., "Germany does 1 million tests per week")
• Discussion of possible cures (e.g., speculation about heat killing the virus)
• Factual questions (e.g., "Will COVID-19 disappear in summer?")

Notes

This talk is notable for its source-centric framing of misinformation detection: rather than asking "Is this claim true?" (reactive and slow), it asks "Is this source trustworthy?" (proactive and scalable). The empirical finding that false information spreads six times faster than truth, and that half the viral spread happens in ten minutes, is a damning critique of reactive fact-checking—the human and organizational capacity simply cannot scale to the speed of social media.

Nakov's distinction between misinformation (false, unintentional) and disinformation (false, intentional) is standard, but his emphasis on propaganda as a separate category—true or false, but designed to manipulate—is important. It broadens the scope beyond falsity to include legitimate information weaponized through selective presentation, emotional appeals, and framing.

The fine-grained propaganda technique detection is a key contribution: it moves beyond binary "propaganda" / "not propaganda" labels to identify specific persuasion techniques (appeal to fear, bandwagon, loaded language, etc.), making the manipulation visible and, per Nakov's media literacy thesis, less effective.

The multidimensional media profiling (combining outlet content, audience, Wikipedia, URL structure, YouTube channel) is sophisticated—no single signal suffices, but combined signals are predictive. The observation that extreme outlets tend to be both biased and low-factuality is important for understanding media ecology and may guide interventions (e.g., prioritizing fact-checking from high-bias, low-factuality sources or promoting media literacy strategies targeted at different audiences).

The work on COVID-19 misinformation and the multi-perspective annotation schema (verifiability, falsity, interest, harm, actionability) shows the complexity of categorizing pandemic-related claims—many are neither simple true/false nor harmless/harmful, requiring nuanced judgment. This work informs policy makers, public-health officials, and social-media platforms on what to prioritize.