Why Content-Authentication Countermeasures Fall Short Against AI-Enabled Propaganda: A Comparative Review and a Defense Architecture

Watermarking, content-provenance standards, and forensic detection are the technical countermeasures most often proposed against AI-generated propaganda. This review evaluates them, together with psychological prebunking, against a common set of criteria and distinguishes failures that follow from a measure's design from failures that better engineering could remove. Two failures are structural: watermarking cannot reach actors who run their own models, and capture provenance certifies a recording pipeline rather than the world it records while depending on near-universal adoption. Most other documented failures, including C2PA's specification flaws, are contingent but recurrent. Prebunking shows no structural failure, but its evidence base is weaker than often presented. All three technical measures authenticate individual artifacts, while the evidence locates AI's impact on propaganda elsewhere: in the cost, breadth, and coordination of broadcast campaigns and in conversational persuasion, which leaves no artifact to authenticate. Building on the influence-operation "kill chain" of Goldstein et al. (2023), the paper maps each defense layer against three tiers of adversary, identifies where layers interact or undermine one another, and sets out recommendations for AI developers, platforms, regulators, and researchers.