Semantic Calibration in Media Streams

ICLR 2026 Conference Submission18248 Authors

Keywords: moderation, semantic calibration

Abstract: Current generative models can produce synthetic media that is visually indistinguishable from real content. As a result, traditional detection methods rely mostly on subtle artifacts introduced during generation. However, we show that such methods could eventually become ineffective. Anticipating this, we suggest that the main risk lies not in whether a media sample is synthetic or real, but in whether its semantic content is deceiving, that is, whether it distorts the information distribution in a way that misrepresents reality. To capture this, we formally introduce the notion of deception in the context of online media streams. Complementing standard detection approaches, we introduce semantic calibration to mitigate deception directly by processing semantic content using captioning and large language models, rather than relying on artifacts introduced by generative models. Our method is explainable, transparent, and modality agnostic, providing a rigorous foundation for developing new tools to combat online misinformation. We offer both theoretical justification and empirical evidence for its effectiveness.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 18248