Meaningful Synthetic Data? On the Possibility of Using Generative AI to Produce Images for Socio-Semiotic Research

This article explores the potential and limitations of generative artificial intelligence (AI) for producing synthetic images that preserve socio-semiotic interpretability while ensuring privacy. It situates the analysis within the European project Camouflage, which aimed to anonymize visual content from social media without resorting to conventional obfuscation techniques such as pixelation or blurring, which compromise meaning. The paper outlines the dual challenge of generating data that is both technically robust—balancing privacy, fidelity, and utility—and semiotically meaningful, capable of sustaining interpretative practices relevant to socio-cultural research. From a machine-learning perspective, anonymization demands techniques that suppress personal identity while retaining features relevant for downstream analysis. After surveying existing approaches, including inpainting methods, latent space manipulation, and latent diffusion models, the article presents Camouflage-Base and Camouflage-Light, two novel systems developed within the project. Empirical evaluations conducted on benchmark datasets demonstrate their capacity to reduce re-identification risks while preserving significant information for tasks such as attribute recognition. Complementary evaluation metrics—including Train on Synthetic, Test on Real paradigms, Visual DNA, and the newly proposed FRESCO-Score—are discussed as ways of assessing fidelity and interpretability beyond conventional measures. From a semiotic standpoint, however, the issue remains complex. Images—and in particular those usually studied in socio-cultural research, such as those posted by people on social media or those published on websites and blogs—are not mere carriers of biometric identifiers but polysemous signs whose meaning depends on physiognomic, pathognomic, plastic, and enunciative traits. Altering such features risks undermining typification processes central to socio-semiotic inquiry. A Delphi-based survey of semiotics scholars highlights the extent to which anonymized outputs diverge from originals in ways that shift their interpretive potential. The preservation of meaning is not only a technical matter but one requiring prior knowledge of the interpretive hypotheses guiding research. The article concludes that while generative AI offers promising avenues for privacy-preserving image synthesis, its application to socio-semiotic research requires caution. Ultimately, anonymization cannot be conceived as a purely computational task but as a semiotic negotiation, where the preservation of meaning depends on the convergence of technical constraints and interpretative practices.