<![CDATA[The proliferation of generative artificial intelligence has transformed the cyber threat landscape, particularly in the domain of social engineering. Deepfake technology, encompassing synthetic audio and video generation, has emerged as a vector for advanced phishing campaigns that can bypass conventional controls, exposing limitations of traditional Security Awareness Training (SAT) when confronted with AI-driven deception. This paper presents a systematic analysis of empirical evidence on deepfake-enabled social engineering and its implications for existing security-awareness frameworks, drawing on 56 primary studies involving 86,155 participants, supplemented by 47 documented corporate incidents and a review of current technical detection methods. Pooled human detection accuracy for deepfake content was 55.54% (95% CI [52.3%, 58.8%]), only marginally above chance; the pooled estimate is, however, accompanied by substantial heterogeneity (I² = 78.4%) and should be interpreted as an average performance rather than a uniform inability to discriminate. Traditional SAT was associated with a non-significant +1.6% improvement in deepfake detection, whereas SAT incorporating synthetic-media examples produced significant gains of +8.1% to +15.5%. The study identifies three persistent limitations in current awareness training: the absence of deepfake-specific detection heuristics, inadequate calibration of trust in response to synthetic authority cues, and insufficient inoculation against cognitive-load manipulation; a 21.1-percentage-point laboratory-to-field performance gap was also observed. The paper proposes a resilience-oriented training framework that integrates technical literacy, psychological preparedness, and organisational verification mechanisms. Rather than declaring the “human firewall” obsolete, the analysis argues for its reconceptualisation as a complementary safeguard within layered, procedure-anchored defence.]]>
Copyrights © 2026
