Voice-Cloning as Crisis Threat: AI Detection Limits, Defender Latency and Escalation Risk in the 2024 Marcos Audio Deepfake Crisis in the Philippines
Loading...
Files
Size: 5.99 MB, File format: Adobe PDF
Date
Authors
Zagrusvon, Panthea
Supervisor
Barker, Sarah
Batistich-Vogels, Christina
Item type
Degree name
Journal Title
Journal ISSN
Volume Title
Publisher
Auckland University of Technology
Abstract
Deepfakes are increasingly used in reputational attacks, fraud and political disinformation, creating risks for public trust, crisis communication and institutional decision-making. This dissertation examines the April 2024 Philippines incident in which a fabricated audio recording, overlaid on South China Sea imagery, appeared to present President Ferdinand “Bongbong” Marcos Jr. directing the Armed Forces of the Philippines and related task groups to act against China. Philippine authorities treated the recording as a national-security concern. Because the original YouTube upload and associated DAPAT BALITA channel were later removed, the incident provides a bounded empirical case for examining how visible limits of AI-based anomaly detection and crisis response can be reconstructed when primary platform artefacts are no longer public. Using a single-case open-source intelligence design, the study integrates qualitative crisis reconstruction, digital trace analysis and bounded timing measures. It draws on lawfully obtainable online traces, including official and media reports, surviving platform traces and third-party YouTube analytics. These materials are used to reconstruct the earliest plausible public-availability window of the fabricated recording (T₀), measure defender latency (Δ) to first visible authoritative responses, and analyse the pre-crisis and trigger conditions that shaped escalation risk. The findings indicate that, under OSINT-only constraints, no AI-based anomaly-detection labels, warning banners, manipulated-media indicators or propagation-level analytical signals were publicly visible at the user-interface level. The Presidential Communications Office issued the first authoritative public advisory approximately three to four days after the reconstructed T₀ window. Major domestic news organisations amplified the denial and warning within hours, whereas platforms, cyber agencies and political or legislative actors became visible later through reported investigations, account deactivations, cross-platform takedowns and terrorism-related framing. The analysis identifies two layers of advance warning: a broader strategic environment shaped by alliance signalling, South China Sea legal and jurisdictional contestation, foreign-influence narratives and domestic tension; and a compressed 11-22 April period marked by high-level diplomacy, Balikatan-related announcements and intensified military exercise activity. In this escalation-prone context, a fabricated presidential voice apparently directing military action in disputed waters was both plausible and potentially consequential. The dissertation contributes a public-trace timing protocol for analysing deepfake crises where original content and platform records are unavailable, combining OSINT-based T₀ reconstruction with comparative defender-latency analysis. It advances interdisciplinary scholarship by integrating AI anomaly-detection research, audio-spoofing studies, crisis and issue management, platform governance, digital trace analysis and legal-policy perspectives. It treats timing, visibility and contextual escalation risk as observable indicators of how AI-supported detection systems and crisis-communication practices perform during a real-world national-security crisis. The study also offers practical implications for defenders and policymakers by showing why synthetic-voice incidents require earlier detection visibility, rapid authoritative intervention, evidence preservation and accountable governance of AI-based anomaly-detection outputs.