Detecting Deception, Not Deepfakes: Why Media Forensics Needs Social Theories

Jessee Ho, Shweta Khushu, Shaina Raza · 2026

Read Detecting Deception, Not Deepfakes: Why Media Forensics Needs Social Theories by Jessee Ho, Shweta Khushu, Shaina Raza (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-09007.txt.

Introduction

Jessee Ho, Shweta Khushu, and Shaina Raza say deepfake guards aim at the wrong mark. The mark should be the trick, not the trace. Guards that call real or fake score well on stock sets. They fall hard on new growth tools. In live calls, the harm is the trick act, not a flaw in the sound or sight. The team shows five props of trace work all wear thin as tools grow. They call the false calm the Generalization Illusion.

From speech act thought, from Grice's rules of talk, and from Cialdini's sway rules, they build a mate frame. It checks the line said, the talk as a whole, and how the hearer acts. Guards gain a social layer.

Catalog record

CREATORJessee Ho, Shweta Khushu, Shaina Raza
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.09007

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Detecting Deception, Not Deepfakes: Why Media Forensics Needs Social Theories Jessee Ho, Shweta Khushu, Shaina Raza Abstract For nearly a decade, deepfake detection has been framed as a classification task: given an audio or video clip, decide whether it is real or synthetic. Top detectors often report high accuracy on standard benchmarks; however, performance drops sharply on content from newer or unseen generators. We argue that better classifiers of synthetic media alone will not solve this problem, especially for interactive deepfakes such as impersonation in video and voice calls, where the harm lies not in the artifact (manipulated media signal) but in the act of deception. Deepfake detection therefore requires a complementary analytical layer focused on communicative interaction, not just media realism. We identify five assumptions that artifact-based detection (the forensic analysis of low-level signal traces) relies on and show that all five are eroding as generative models improve, producing what we call the Generalization Illusion. To address this, we draw on three well-established frameworks from philosophy of language and social psychology, namely, Speech Act Theory, Grice’s Cooperative Principle, and Cialdini’s principles of influence, to examine forensic signals at three levels: the utterance, the conversation, and the listener response. The result is a unified framework that complements existing forensic methods. We close with open problems for future work. (project webpage) 1 Introduction Deepfake 1 1 1 The term “deepfake” is widely believed to have originated on Reddit in 2017, where a user (“deepfakes”) shared AI-generated face-swapped videos 4 detection occupies an unusual position among machine learning (ML) subfields, where benchmark performance and deployment performance are increasingly diverging. State-of-the-art detectors achieve strong results on curated benchmark datasets such as FaceForensics++ 53 , Celeb-DF 41 , and DeeperForensics 27 . However, recent cross-distribution evaluations under real-world conditions find significant accuracy drops on contemporary generators 6 ; 34 . In parallel, deepfake-enabled fraud incidents rose roughly 40-fold between 2022 and 2024 60 ; 61 , with projected losses from generative AI-facilitated fraud in the United States alone expected to reach $40 billion annually by 2027 17 . Figure 1 highlights the rapid global growth of deepfake-enabled fraud 61 . Documented cases, such as the 2019 UK Energy voice-cloned executive impersonation 59 , the 2024 Arup video-conference fraud resulting in a $25.5M wire transfer 8 , and the attempted Ferrari impersonation 36 , share a common pattern: attacks are either completed undetected or noticed through contextual anomalies, such as unusual request channels, out-of-band verification, or inconsistent operational details, rather than through media forensics. Here, media forensics refers to the detection of manipulation in digital media by analyzing pixels, signals, and artifacts 9 . This pattern points to a detection gap that is fundamentally interaction-level, not signal-level. Figure 1: Global growth in deepfake-driven identity fraud from 2023 to 2024 (in %). Circle size scales with country-level growth in identity fraud, while regional labels highlight growth in deepfake-specific fraud 61 . We therefore focus on interactive deepfakes , defined as AI-generated audiovisual impersonation in video calls, voice calls, or targeted messaging 48 , excluding standalone synthetic media without a direct recipient. We argue that this pattern reflects not a gap that better classifiers will close, but a conflation of two distinct problems that were observed together in the early deepfake literature (2018–2020) and came apart as generative models matured after 2022. The first is media classification : determining whether a given artifact, such as a video, image, or audio clip is synthetic. The second is communicative deception : determining whether a given interaction is being used to mislead its recipient. Current detectors ask "was this generated by a machine?” We argue the right question is "is this being used to deceive someone?” Both are classification problems, but the second requires inputs that current detectors typically ignore: speech acts, conversational coherence, and influence patterns, rather than pixels and frequencies. We take the position that this is not an engineering shortfall but a category error: media synthesis detection has been mistaken for the defining question, when it should be treated as one signal within the larger problem of deception detection. Figure 2 illustrates this position: detection must shift from analyzing media artifacts to analyzing the interaction itself. Position Statement Current deepfake detectors focus on identifying traces left by specific generative tools, rather than detecting synthesis itself. As these tools evolve, their traces fade and detection performance declines. We argue that the field needs a new analytical target, shifting from the media artifact to the communicative interaction, and from synthetic realism to operational deception, while operating alongside existing media classification rather than replacing it. Contributions. This paper makes two contributions, structured as a diagnosis and a prescription. (1) Diagnosis (§ 2 ) We identify five premises underlying current detection: spatial artifacts, frequency signatures, temporal inconsistencies, biological signals, and signal survival under compression (P1–P5). We show that these premises are eroding simultaneously as generative models improve and that real-world base-rate effects further compound this problem. Together, these factors produce what we call the Generalization Illusion : the systematic overestimation of real-world capability based on static benchmarks. (2) Prescription (§ 3 –§ 5 ). We propose a three-level framework for interaction-grounded detection, drawing on Speech Act Theory 56 at the utterance level, Grice’s Cooperative Principle 23 at the conversation level, and Cialdini’s principles of influence 11 at the recipient-response level. We ground each level in computational work that supports feasibility and close with a four-problem research agenda and deployment considerations. A glossary of abbreviations and notations is given in Appendix A.2 . Figure 2: The interrogation analogy. Traditional detection focuses on surface-level media artifacts. An interrogation-based approach evaluates speech-act validity, interaction coherence, and manipulative intent, shifting detection from perceptual analysis to communicative structure and behaviour. 2 The Generalization Illusion State-of-the-art benchmark performance in deepfake detection does not necessarily translate into real-world capability. Many reported results are based on evaluation settings that lag behind recent advances in generative AI, widening the gap between benchmark performance and deployment effectiveness. We call this the Generalization Illusion . 2.1 Deepfake Forensic Premises By abstracting across major classes of prior work, including artifact-based, temporal consistency, physiological signal, and model-specific forensic approaches, we derive a five premises. Although not exhaustive, this set provides a minimal but representative basis for analyzing the shared assumptions of existing detectors. We list these premises (P1–P5) below and illustrate them in Figure 3 . P1: Spatial artifacts (2017–2022). This premise assumes that synthetic faces leave visible traces when composited onto real backgrounds. Early face-swap pipelines created such traces at blending boundaries and in local textures. Detectors such as XceptionNet ( 53 ) , Face X-ray ( 39 ) , and LAA-Net ( 46 ) were built to detect these traces. End-to-end diffusion models now synthesize entire frames, leaving no blending step and no boundary to detect ( 52 ; 70 ) . P2: Frequency signatures (2019–2022). This premise assumes that generative models leave a recognizable fingerprint in the frequency content of an image, distinct from the patterns produced by real cameras. GAN-based generators left characteristic spectral fingerprints, such as checkerboard patterns and anomalous frequency distributions, which methods such as F3-Net ( 50 ) , FreqNet ( 65 ) , and FE-CLIP (

Read the complete text