Dresden Research Insights · Generative AI and the New Disinformation Lifecycle · 22 September 2026
Can a dialogue written more than two thousand years ago help us recognize manipulation in contemporary social media? The DeMasKI researchers presented a compelling case for combining classical rhetoric with explainable artificial intelligence.
The project starts from a distinction often lost in automated content moderation: a statement can be manipulative without containing an easily identifiable factual falsehood. Personal attacks, emotional redirection, polarization and attacks on the conditions of rational debate can all change how an argument is received.
The researchers use passages from Plato’s Gorgias to develop annotations at multiple levels. At the broader level they identify narrative and argumentative shifts; at the finer level they identify rhetorical devices such as personal attacks. The ambition is to make these patterns computationally recognizable without discarding their interpretability.
The technical approach combines neural language processing with inductive logic programming. Neural models provide linguistic indicators, while symbolic learning can produce rules that humans can inspect. Preliminary results were promising but did not yet outperform a transformer baseline; the presenters explicitly acknowledged overfitting and the need for richer indicators.
This is an important scientific choice. A black-box classifier might flag a text, but an explainable system could show which rhetorical move caused concern. That makes the tool potentially useful for researchers, educators and journalists rather than merely generating another opaque score.
Research question: Should automated systems detect false claims, manipulative argumentation, or both—and how should they explain the distinction?
Watch the talk: YouTube recording. Explore the complete workshop archive and Dresden programme.
This editorial research insight summarizes a workshop presentation. It is not a verbatim transcript or a peer-reviewed assessment of the presented claims.
Leave a Reply