Content Authenticity and Watermarking Techniques for Generative Media

Abstract

The rapid democratisation of high-quality generative AI -- text, image, audio, and video synthesis at human-indistinguishable quality -- has created a critical societal challenge: the ability to distinguish authentic human-created content from AI-generated synthetic media is becoming increasingly difficult, with significant implications for information integrity, electoral processes, legal systems, and public trust. Watermarking and content authenticity mechanisms -- techniques for embedding verifiable provenance signals in AI-generated content -- have emerged as primary technical responses to this challenge, but existing approaches are fragmented across modalities, vulnerable to removal attacks, and lack standardised evaluation methodologies. This paper proposes the Generative Media Authenticity (GMA) framework, a unified approach to content watermarking and authenticity verification across text, image, audio, and video modalities, comprising four watermarking techniques: semantic watermarking for text (SWT) that embeds verifiable signals in token selection patterns; spectral watermarking for images (SpWI) that embeds imperceptible frequency- domain signals; psychoacoustic watermarking for audio (PWA) that uses auditory masking thresholds; and temporal-spatial watermarking for video (TSWV) that coordinates frame-level and motion-level signals. GMA is evaluated on watermark robustness (resistance to removal attacks), imperceptibility (signal quality degradation), and detection accuracy across four generative AI systems. GMA achieves mean watermark detection accuracy of 97.4% (SD = 1.8%) after 12 attack types including compression, paraphrasing, and adversarial removal, while maintaining 98.6% of baseline generation quality (imperceptibility). The false positive rate (human-created content incorrectly flagged as AI-generated) is 0.8% -- meeting the threshold required for deployment in legal and forensic contexts. The study contributes the GMA specification, a Content Authenticity Score (CAS) metric, and the first cross-modal watermarking robustness evaluation benchmark.