Researchers have introduced FakeSpotter, a system designed to estimate whether text carries the structural signals of viral misinformation without directly declaring its claims true or false. The approach is intended for emerging narratives, where fact-check labels or historical examples may not yet exist.
FakeSpotter measures features across language, narrative structure, logic and critical-thinking dimensions. It combines repeated assessments by language models with separate logistic-regression classifiers for short and long texts. Instead of a single opaque label, its interpretation layer reports feature scores, agreement among warning signals and a caution index that can help analysts decide what deserves closer inspection.
On a labeled collection of 764 social-media and FakeNewsNet texts, the researchers report macro F1 scores of 0.788 for short material and 0.793 for long material on a held-out test set. Macro F1 balances performance across classes rather than letting a common category dominate the score. The result suggests structural clues may help prioritize content for early review, but the sample is limited and the tool does not verify factual accuracy. Human analysts would still need to investigate the underlying claims, context and sources before reaching a conclusion. The paper is a new arXiv preprint and has not been presented as a deployed moderation product.