Disinformation Detection¶
Disinformation detection encompasses methods and systems for identifying deliberately false, misleading, or deceptive information designed to manipulate public opinion or erode trust. It differs from general misinformation detection by focusing on intentional falsehood rather than unintentional error.
Core problem¶
Distinguishing intentional disinformation from: - Misinformation: False information spread without intention to deceive (e.g., shared misunderstanding) - Rumors: Unverified claims that may or may not be false (intent to deceive unclear) - Satire and parody: Intentionally false but labeled as humorous or fictional
Detection approaches¶
Claim-level approaches: Verify the factuality of individual claims using fact-checking databases, knowledge bases, or evidence retrieval systems.
Source credibility: Assess credibility of the source publishing the claim using publication history, institutional affiliation, and editorial standards.
Linguistic analysis: Detect linguistic markers associated with deceptive content (excessive emotional language, vague claims, logical fallacies).
Network and propagation analysis: Analyze how content spreads; disinformation often exhibits distinctive cascade patterns (rapid initial spread, concentration in echo chambers).
Multimodal approaches: For social media, examine images, videos, and metadata alongside text to identify synthetic media or coordinated inauthentic behavior.
Machine-generated content detection: Identify whether content was created by LLMs or other automated systems, which may increase suspicion of coordinated campaigns.
Key challenges¶
Scale: Fact-checkers can verify only a tiny fraction of claims; computational approaches must handle volume.
Subjectivity: Determining intent to deceive requires understanding publisher motivation, which is often opaque or ambiguous.
Context-dependence: Factuality depends on temporal and geographical context; "true" in one time period may be "false" in another.
Multilingual content: False narratives cross language boundaries; most detection research focuses on English.
Adversarial evolution: As detection methods improve, disinformation tactics adapt (paraphrasing, image manipulation, bot amplification).
Key papers¶
- Modzelewski et al. (2026) — MALicious INTent Dataset and Inoculating LLMs for Enhanced Disinformation Detection: First English corpus annotating both disinformation and malicious intents (5 categories); proposes intent-augmented reasoning (IBI) that improves zero-shot LLM detection by ~9% on MALINT and ~20% on other datasets across six languages.
- Macko et al. (2026) — Beyond Speculation: Measuring the Growing Presence of LLM-Generated Texts in Multilingual Disinformation: Measures prevalence of LLM-generated content in fact-checked disinformation datasets across languages and platforms; documents temporal increase post-ChatGPT and identifies vulnerable contexts (specific languages, platforms)
- A Survey on Computational Propaganda Detection — Comprehensive survey of computational propaganda detection; distinguishes propaganda (persuasion) from disinformation (falsity) and covers network and NLP approaches
Related topics¶
- Fake news detection (broader concept encompassing misinformation and disinformation)
- Misinformation (unintentional false information)
- Fact-checking and corrections (verification of claims)
- Multilingual Disinformation (language-specific aspects)
- Machine-generated text detection (detection of synthetic content)