Traditional automated fact-checking systems treat truth as a binary switch: a claim is either entirely true or entirely false. However, real-world misinformation rarely operates in absolute fabrications. Instead, it thrives in the "gray areas"—incorporating accurate data but distorting its implications, cherry-picking facts, or presenting true information within a deceptive context. As noted by Zhijiang Guo et al. in A Survey on Automated Fact-Checking, binary frameworks are fundamentally insufficient for assessing real-world claims that contain a complex mixture of accurate and misleading information.
When models force these nuanced statements into a binary classification, they fail to provide practical utility for professional journalists and undermine user trust. For instance, Preslav Nakov et al. argue in Automated Fact-Checking for Assisting Human Fact-Checkers that this mismatch between model outputs and the complex reality of journalism limits the adoption of AI tools. Furthermore, a simple "true or false" tag lacks the persuasive power needed to correct misconceptions; Saadia Gabriel et al. demonstrate in MisinfoEval: Generative AI in the Era of "Alternative Facts" that without detailed explanations and context, users are unlikely to internalize corrections or improve their media literacy.
Decomposing the Spectrum of Misinformation
To capture this complexity, recent research shifts toward fine-grained veracity classification. Instead of a single label, models must learn to place claims along a multi-class scale or evaluate them across multiple dimensions.
Ashim Gupta introduced the X-Fact: A New Benchmark Dataset for Multilingual Fact Checking benchmark to evaluate claims across a multi-class scale (e.g., 'mostly true' or 'partly true'), allowing models to learn complex distinctions between varying degrees of factual accuracy. Similarly, Kai Nakamura et al. developed the Fakeddit: A New Multimodal Benchmark Dataset for Fine-grained Fake News Detection dataset, which categorizes misinformation into specific types such as satire, misleading content, imposter content, and manipulated media.
Rather than just measuring factual accuracy, some frameworks evaluate content across entirely separate axes. For example, Ipek Baris Schlicht et al. propose a three-dimensional taxonomy in Co-Creation of Misinformation Management Policies that evaluates content based on truthfulness, intent to harm, and verifiability. This allows a system to distinguish between a harmless piece of satire and a malicious conspiracy theory, enabling more precise and appropriate content moderation.
Granular Attribution and Localized Inconsistencies
Why do models struggle with fine-grained classification? In a binary setup, a model can often rely on coarse-grained features or event-specific keywords. But as Jing Ma et al. point out in A Coarse-to-fine Cascaded Evidence-Distillation Neural Network for Explainable Fake News Detection, distinguishing between subtle degrees of truthfulness—such as differentiating between 'barely true' and 'false'—requires capturing highly nuanced linguistic and contextual cues.
To address this, models must move from bulk-labeling entire documents or events to pinpointing specific localized errors. Soumya Suvra Ghosal et al. highlight in ReSCo-CC: Unsupervised Identification of Key Disinformation Sentences that disinformation is frequently embedded within otherwise factual or neutral text. Their work focuses on identifying the exact sentences containing core disinformation, preventing over-generalized warnings that destroy user trust.
Similarly, in natural language inference (NLI) and hallucination detection, standard models often reduce conflicts to simple "entailment" or "contradiction" labels. Tathagata Raha et al. address this in Neural models for Factual Inconsistency Classification with Explanations by categorizing factual inconsistencies into specific linguistic types and localizing the exact text spans where the error occurs. By verbalizing the specific mechanism of deception—whether it is a minor detail omission or a complete fabrication—systems can explain how and where a claim goes wrong, which Hao Guo et al. show in Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection is vital for building explainable and trustworthy detection pipelines.
Navigating Ambiguity and Undecidability
A major bottleneck in binary fact-checking is handling claims that cannot be cleanly verified due to linguistic ambiguity or a lack of evidence. When forced to choose between true or false, models misclassify these edge cases, leading to poor alignment with human judgment.
Farima Fatahi Bayat et al. demonstrate in FactBench: A Dynamic Benchmark for In-the-Wild Language Model Factuality Evaluation that incorporating an "undecidable" category into factuality evaluation significantly improves model reliability by explicitly accounting for claims that are subjective, ambiguous, or lack sufficient external evidence. This mirrors earlier findings by Madian Khabsa et al. in Towards Few-shot Fact-Checking via Perplexity, who noted that distinguishing between explicitly refuted claims and those with insufficient evidence remains a core challenge for automated systems.
To resolve this ambiguity entirely, Ieva Staliunaite et al. proposed the task of claim disambiguation in Dis2Dis: Explaining Ambiguity in Fact-Checking. Instead of simply assigning a discrete label to an underspecified claim, their framework rewrites the claim into multiple, evidence-backed versions, clarifying the specific contexts or interpretations under which the statement holds true or false.