AI-Enabled Deepfake Psychological Operations#

1. Executive Summary#

The rapid proliferation of generative artificial intelligence (GenAI) has fundamentally altered the landscape of information warfare, enabling the deployment of AI-enabled deepfake psychological operations at unprecedented scale and velocity. This interdisciplinary report defines a deepfake psychological operation as the intentional use of synthetic or manipulated audio, video, imagery, or multimodal media to alter perceptions, create false evidence, impersonate a trusted figure, provoke action, discredit authentic evidence, or undermine trust for a political, military, ideological, criminal, or coercive objective. The analysis indicates that the threat environment is characterized by a dual vulnerability. First, malicious actors leverage highly realistic synthetic media to deceive human targets, bypass biometric security infrastructure, and manipulate democratic and financial systems. Second, the mere existence of these generative technologies engenders a pervasive epistemic skepticism that empowers wrongdoers to dismiss genuine, incriminating evidence as AI-generated—a phenomenon widely known as the "liar’s dividend"1. Through an exhaustive examination encompassing digital forensics, constitutional and international law, crisis communication, and cognitive psychology, this report demonstrates that technical realism is often less predictive of a deepfake’s success than its alignment with the target audience's pre-existing confirmation biases, the emotional intensity of the content, and the strategic timing of its release. The vulnerabilities of current provenance standards, such as the Coalition for Content Provenance and Authenticity (C2PA) and Google’s SynthID, are critically evaluated alongside emerging cryptographic solutions like Zero-Knowledge Proofs (ZKPs)4. Furthermore, the report outlines rigorous crisis-response protocols, analyzes the intersection of synthetic media with international humanitarian law and domestic free speech protections, and provides strategic recommendations for cultivating institutional and public resilience against cognitive manipulation.

2. Definitions and Synthetic-Media Taxonomy#

Understanding the mechanics of deepfake psychological operations requires a precise taxonomy of the media forms utilized to manipulate cognitive and emotional processing. These synthetic artifacts do not operate monolithically; they are engineered for specific strategic utilities depending on the target audience and the desired psychological effect. The following table delineates the primary forms of synthetic media, their mechanisms, and their strategic applications.

3. Historical Precedents in Manipulated Media#

The weaponization of media for psychological operations is deeply rooted in historical statecraft. Analog photo manipulation—such as the systematic airbrushing of political dissidents from official Soviet photographs under Joseph Stalin—served the same fundamental cognitive purpose as modern deepfakes: the alteration of objective reality to consolidate power, shape historical memory, and eliminate contradictory narratives. In the analog era, altering reality required state-level resources, highly skilled technicians, and significant time, inherently limiting the scale of disinformation campaigns. With the advent of digital editing software in the late twentieth century, the barrier to entry for media manipulation decreased significantly. This birthed the era of "shallowfakes"—media altered through relatively low-tech methods such as deceptive cropping, slowing down video to make a speaker appear intoxicated or cognitively impaired, or miscaptioning older footage to represent a current crisis. While shallowfakes democratized deception, they still required manual human intervention and were often bounded by the availability of existing media that could be recontextualized. The contemporary paradigm shift is defined by the automation, infinite scalability, and hyper-realism afforded by machine learning. The transition from manual digital manipulation to generative AI represents a qualitative leap in information warfare. Generative AI decouples the creation of evidence from physical reality entirely. State and non-state actors, transnational criminal syndicates, and lone provocateurs can now generate high-fidelity, multimodal propaganda at near-zero marginal cost and unprecedented velocity10. This historical evolution demonstrates a trajectory where the capacity to define reality has shifted from centralized state apparatuses to decentralized, algorithmically empowered networks, necessitating a fundamental reevaluation of how societies authenticate information.

4. Current Generation Capabilities and Limitations#

Current generative AI systems operate primarily via Large Language Models (LLMs) for textual generation, diffusion models for imagery and video, and advanced neural networks for voice cloning. These systems are highly capable of generating photorealistic and audiorealistic outputs that routinely defeat human perceptual processing. Behavioral research indicates that human subjects are essentially operating at chance levels when attempting to distinguish high-quality deepfakes from authentic media, frequently falling victim to a "seeing-is-believing" heuristic18. Despite their sophistication, these technologies possess distinct technical limitations. High-fidelity video generation, particularly in diffusion models, continues to struggle with temporal consistency. Over extended durations, synthetic videos often display artifacts such as flickering, violations of basic physics, morphing background elements, or anatomical anomalies (such as inconsistent dentition or poorly rendered extremities)10. Voice cloning, while highly accurate in replicating vocal timbre, can sometimes fail to capture the appropriate emotional cadence required for a specific context, or it may lack accurate ambient acoustic matching—such as a voice lacking the appropriate environmental reverberation when allegedly recorded in a large reverberant space12. Furthermore, text watermarking in LLMs faces an "entropy gap," where the low information-carrying capacity of text makes it exceedingly difficult to embed robust statistical signals without altering the semantic meaning of the output5. Crucially, however, high technical realism is frequently unnecessary for a psychological operation to succeed. Research demonstrates that low-tech edits, misleading captions, and poor-quality synthetic audio are often more effective than flawless, advanced deepfakes if they are deployed strategically. When released rapidly during an information vacuum, or when targeted at an audience with strong pre-existing ideological biases, rudimentary manipulation is sufficient to bypass critical evaluation10. Low-quality media can even provide a layer of operational security for the attacker; visual glitches or auditory anomalies in a highly compressed file can be plausibly attributed to poor network connections or artifacting rather than synthetic generation, as seen in various wartime influence operations11.

5. Psychological and Crisis Effects#

To understand under what conditions deepfakes are likely to persuade, confuse, delay verification, or provoke immediate action, it is necessary to examine the cognitive mechanics of human information processing. The efficacy of a deepfake is rarely determined solely by its pixel-perfect realism; rather, it is dictated by its interaction with human psychology. Deepfakes thrive on the heuristic of "Truth-Default Theory," a psychological framework positing that humans inherently accept incoming information as true unless presented with explicit, overwhelming reasons for suspicion3. This default to truth is heavily modulated by confirmation bias and partisan-motivated reasoning. Empirical studies, including extensive behavioral experiments regarding political deepfakes, indicate that individuals are highly susceptible to synthetic media that aligns with their pre-existing political or social identities. Furthermore, partisans are highly likely to actively doubt the credibility of an authentic scandal if it reflects poorly on their own political in-group, demonstrating that ideological allegiance overrides objective evidence evaluation21. Technical realism matters significantly less than timing, source credibility, and emotional intensity. When a deepfake is released during a period of high anxiety—such as the immediate aftermath of a terrorist attack or during a closely contested election—human cognitive processing shifts from analytical, deliberate reasoning (System 2\) to intuitive, reactive processing (System 1). In this heightened emotional state, if the synthetic media confirms a deeply held prejudice or fear, the audience will actively rationalize away visual or auditory artifacts that might otherwise reveal the deception22. Verification during a fast-moving crisis is often outpaced by algorithmic amplification; by the time forensic experts can definitively debunk a piece of media, the emotional damage has been inflicted, and the target audience has often moved on, rendering the correction ineffective. Certain crisis settings are exceptionally vulnerable to these operations. Elections are prime targets, particularly during pre-election media moratoriums where candidates and journalists are legally restricted from communicating, preventing the rapid debunking of synthetic smears released in the final hours of a campaign10. Armed conflict introduces the fog of war, allowing synthetic media to incite panic, prompt unwarranted surrenders, or falsely implicate adversaries in atrocities, thereby manipulating international diplomatic responses24. Financial markets are susceptible to algorithmically driven bank runs; a synthetic audio clip of a central bank governor or corporate CEO announcing catastrophic failure can trigger automated trading sell-offs before human verification protocols can intervene. Similarly, in regions prone to communal violence, deepfakes depicting religious desecration or ethnic violence can rapidly incite fatal mob retaliation, leveraging deep-seated historical grievances to bypass any critical scrutiny.

6. Case Studies#

The following documented cases illustrate the diverse applications, distribution vectors, and measurable consequences of deepfake psychological operations across varying sectors, highlighting the practical challenges of verification and response.

These cases demonstrate that measurable effects range from localized reputational destruction and massive financial fraud to the destabilization of national elections and military coups. They underscore that the success of a psychological operation relies heavily on exploiting structural vulnerabilities—such as media moratoriums, hierarchical corporate trust, or deep-seated political paranoia—rather than purely relying on the technical perfection of the generative model.

7. Detection and Forensic Verification#

The forensic detection of synthetic media operates on a spectrum from human perceptual analysis to advanced algorithmic detection, representing an ongoing arms race between generative architectures and defensive forensics. Traditional forensic techniques analyze localized inconsistencies within the media file. Visually, analysts look for asymmetrical lighting, unnatural blending at the edge of the face mask, mismatched pupil reflections, or a lack of natural physiological movements such as blinking or breathing. Audio forensics rely heavily on spectrographic analysis to identify unnatural frequency gaps, robotic cadence, or the absence of the acoustic reverberation expected in the depicted environment11. However, the primary limitation of post-hoc forensic detection is its inherent fragility and reactive nature. As generation models improve, they are specifically trained to synthesize the exact physiological and physical cues that current detectors look for, rapidly closing the perceptual gap. Furthermore, malicious actors frequently employ adversarial techniques to intentionally degrade synthetic media—adding severe compression artifacts, gaussian noise, or artificial blurring—to destroy the high-frequency statistical signals that AI-based detectors rely upon11. Consequently, reliance on human evaluation performs no better than chance, and algorithmic detectors suffer from unacceptable rates of false positives and false negatives, rendering them insufficient as standalone legal or journalistic evidence in high-stakes environments7. The consensus among forensic experts is that attempting to prove a negative—that an event never happened—using black-box algorithmic detectors is a failing strategy.

8. Provenance and Authentication Systems#

Given the severe limitations of post-hoc detection, the industry paradigm is shifting toward "digital provenance"—the concept of cryptographically proving the origin, authorship, and edit history of media at the point of capture, ensuring a secure chain of custody through to publication. The most prominent effort in this space is the Coalition for Content Provenance and Authenticity (C2PA), which established a standard for binding cryptographically signed metadata to media files, forming what are known as "Content Credentials." While theoretically robust, independent security analyses have revealed severe structural vulnerabilities within the C2PA specification. A comprehensive 2026 formal-methods study by Golaszewski et al. demonstrated that C2PA fails to achieve its core security goals4. The vulnerabilities include timestamp forgery, where optional trusted timestamps can be replaced or modified without invalidating the underlying signature4. Furthermore, many C2PA validators fail to properly check for revoked certificates, permitting attackers who have compromised a camera's signing key to forge seemingly authentic media indefinitely39. Perhaps most critically, C2PA relies on an "exclusion range" to allow for privacy redactions; attackers can exploit this range to secretly alter vital metadata, such as GPS coordinates, without breaking the cryptographic signature4. Finally, the practical utility of C2PA is crippled by metadata stripping, as social media platforms routinely strip container metadata during image compression, destroying the provenance chain for an estimated 95% of shared media41. Watermarking offers an alternative approach, embedding imperceptible signals directly into the media content rather than attaching metadata. Google's SynthID represents the state-of-the-art in this domain, embedding statistical patterns into the pixels of images, the audio spectrograms of sound files, or the generation probabilities (logits) of text5. However, watermarking is highly brittle, particularly in text. Research demonstrates that SynthID text watermarks are acutely vulnerable to meaning-preserving attacks. Simple adversarial actions, such as heavy paraphrasing, synonym substitution, or back-translation through another language, degrade the watermark's statistical signal to the point of undetectability, all while preserving the semantic payload of the disinformation5. To address the surveillance concerns and redaction limitations of C2PA, advanced cryptographic frameworks utilizing Zero-Knowledge Proofs (ZKPs) are being developed. A ZKP allows a "prover" to demonstrate to a "verifier" that a specific statement about a piece of data is mathematically true, without ever revealing the underlying data itself6. In the context of digital media, a journalist operating in a hostile environment could take a cryptographically signed photograph, use software to crop out a sensitive source's face or redact a license plate, and generate a ZKP. This proof mathematically guarantees to the public that the published image is derived from an authentic, camera-signed photograph and was only subjected to permissible edits, without ever revealing the original unredacted photo or exposing the camera's secure private key6. ZKPs provide forensic-grade authenticity and omission detection (ensuring evidence wasn't selectively deleted) without creating centralized registries of surveillance that authoritarian regimes could exploit, effectively solving the privacy-versus-authentication paradox42. The following table summarizes the comparative strengths and limitations of these dominant authentication architectures:

9. The Liar’s Dividend#

The proliferation of synthetic media generates a secondary psychological effect that is arguably more insidious than the deepfakes themselves: the "liar’s dividend." Coined by legal scholars Robert Chesney and Danielle Citron, this concept describes how the mere existence of highly realistic deepfakes introduces a persistent, ambient informational uncertainty into society1. Wrongdoers can actively leverage this uncertainty by falsely claiming that genuine, unmanipulated audio or video evidence of their misconduct is an AI-generated fabrication1. This tactical denial exploits "Impostor Bias"—a newly identified cognitive heuristic where individuals systematically question the authenticity of real multimedia content, assuming it to be AI-generated due to their heightened awareness of synthetic capabilities50. In environments saturated with disinformation, the public's default setting shifts from trust to profound skepticism. Empirical research regarding the efficacy of the liar's dividend yields complex, highly contextual results. Initial studies by Schiff et al. (2025) suggested that falsely crying "deepfake" works exceptionally well to evade accountability for text-based scandals, allowing politicians to maintain support by activating partisan oppositional rallying, but that audiovisual evidence remained harder to deny51. However, subsequent 2026 research by Grohmann et al. demonstrated that as public trust in media institutions continues to decay, false deepfake claims regarding audiovisual evidence successfully yield a substantial dividend. Specifically, politicians utilizing this strategy experienced higher perceived leadership abilities from their base, while their claims simultaneously drove down the public's general trust in the media infrastructure attempting to hold them accountable1. The Ali Bongo case in Gabon serves as a visceral real-world manifestation of the liar's dividend, demonstrating how the pervasive suspicion of synthetic media can provide the necessary pretext to destabilize a sovereign nation35.

10. Crisis-Response Protocols#

During a fast-moving crisis triggered by a suspected deepfake, the speed of forensic verification cannot match the velocity of algorithmic amplification on social media. By the time a definitive technical debunking is published, the cognitive damage has been inflicted. Consequently, government officials, newsrooms, platforms, and emergency services must adopt structured crisis communication protocols focused on rapid harm reduction, evidence preservation, and strategic psychological debunking. The fundamental principle of responding to a synthetic media incident is to avoid inadvertently amplifying the false content. Repeating the false claim, even to debunk it, strengthens the cognitive association in the audience's memory. Instead, communicators must strictly adhere to the "Truth Sandwich" model. This communication framework requires the speaker to lead with the factual reality, briefly contextualize the existence of the manipulative media without hyperlinking to it or reproducing it, and end by forcefully reiterating the truth alongside verified evidence52. To operationalize this response, organizations should adopt the following response timeline for high-impact suspected deepfake incidents:

Before an incident even occurs, proactive communication is essential. Organizations must establish public tip lines for citizens to report suspicious media, train executives on deepfake threats, and actively "prebunk" anticipated narratives. Prebunking exposes the public to a weakened form of the manipulative tactic—explaining exactly how voice cloning works and warning that it will likely be used in an upcoming crisis—thereby building cognitive antibodies that reduce susceptibility when the actual attack occurs11.

11. Legal and Regulatory Analysis#

The legal architecture governing deepfake psychological operations is highly fractured, attempting to balance the urgent necessity of preventing deception with the preservation of fundamental human rights and free speech protections. Different bodies of law apply depending on the context, intent, and target of the synthetic media.

Domestic Law and the First Amendment (United States)#

In the United States, regulating political deepfakes faces severe First Amendment hurdles. In response to the proliferation of election disinformation, California enacted Assembly Bill 2839 and Assembly Bill 2655 in 2024. These laws sought to prohibit the distribution of materially deceptive election deepfakes and required large online platforms to block such content during critical pre-election periods55. However, a federal judge swiftly enjoined these laws in the case of Kohls v. Bonta. The court ruled that the statutes acted as a "hammer instead of a scalpel," unconstitutionally stifling protected political satire and parody. The ruling also cited Section 230 of the Communications Decency Act in protecting platforms from liability for third-party synthetic content55. Conversely, legislation targeting the non-consensual exploitation of private individuals faces significantly fewer constitutional challenges. The federal TAKE IT DOWN Act of 2025 empowers the Federal Trade Commission to enforce removal requirements against platforms hosting non-consensual synthetic intimate imagery, as such material generally falls under obscenity or defamation exemptions outside First Amendment protections60. In the civil realm, victims of localized deepfake campaigns, such as the Pikesville High School principal, can pursue traditional tort claims including defamation, false light, and intentional infliction of emotional distress against the creators, though pursuing anonymous creators remains practically difficult31.

European Union Regulations#

The European Union approaches synthetic media through the lens of mandatory transparency. The EU AI Act attempts to mitigate deepfake harms by legally requiring deployers of AI systems to clearly label deepfakes and AI-generated text published on matters of public interest, provided the content lacks human editorial control (Article 50\)61. The regulatory philosophy assumes that an informed public will discount labeled synthetic media. However, the psychological effectiveness of labeling is heavily debated, as labels are frequently ignored by users engaged in intuitive processing, intentionally stripped by malicious actors, or weaponized to further polarize audiences.

International Humanitarian Law (IHL)#

In the context of armed conflict, the legality of deepfake psychological operations hinges entirely on the established distinction between perfidy and ruses of war under Article 37 of Additional Protocol I to the Geneva Conventions62.

• Perfidy (Prohibited): A deepfake operation violates IHL if it invites the confidence of an adversary with the intent to betray that confidence while feigning protected status. For example, generating a deepfake of an enemy commander falsely announcing a surrender, or deploying a synthetic video of the International Committee of the Red Cross calling a false medical truce to draw opposing troops into a lethal ambush, constitutes perfidy and a war crime62.
• Ruses of War (Permitted): Conversely, deepfakes used for general psychological warfare and deception do not feign protected status. Generating fake satellite imagery of troop movements, fabricating generic operational orders, or deploying synthetic videos to demoralize the civilian populace are generally considered lawful ruses of war under current interpretations of IHL. This remains true despite their potential to inflict severe psychological distress on civilian populations or trigger dangerous escalations, revealing a significant gap in modern humanitarian protections8.

Emerging legal scholarship argues that the unfettered use of synthetic media in conflict zones necessitates the recognition of a new human right: "cognitive liberty," defined as the right to mental self-determination and protection against algorithmic manipulation that bypasses rational cognitive processes8.

12. Institutional and Public Resilience#

Because technical detection is locked in a perpetual arms race, and legal regulation is constrained by constitutional and jurisdictional boundaries, the ultimate defense against deepfake psychological operations relies on cultivating profound institutional and public resilience. Institutions must abandon the assumption that seeing is believing. Authentication infrastructure must transition toward absolute zero-trust principles, moving away from the subjective evaluation of media realism toward the strict cryptographic verification of data provenance from the moment of capture18. Public resilience requires a fundamental shift in media literacy. Rather than teaching the public to spot technical glitches—which will inevitably disappear as models improve—education must focus on recognizing the emotional manipulation tactics and dissemination patterns utilized by adversaries. By consistently applying inoculation theory and prebunking methodologies, societies can build the cognitive antibodies necessary to evaluate information based on its verifiable cryptographic chain of custody rather than its visceral emotional appeal.

13. Future Scenarios and Warning Indicators#

The trajectory of synthetic media points unequivocally toward real-time, multimodal, and interactive generation capabilities. Warning indicators for the next evolution of psychological operations include:

• Real-Time Interactive Avatars: The tactical shift from pre-recorded deepfake videos to real-time interactive deepfakes deployed in live video conferencing (as foreshadowed in the Hong Kong CFO fraud) or via automated, conversational telephone calls. These systems will be capable of holding dynamic, persuasive negotiations or extracting sensitive information autonomously.
• Hyper-Personalized Micro-Targeting: The integration of LLMs with vast datasets of stolen personally identifiable information (PII) to generate highly individualized synthetic media. For example, generating a voice clone of a specific voter's family member urging them to go to the wrong polling station, deployed simultaneously to thousands of targets53.
• The Epistemic Apocalypse: A scenario where the sheer volume of algorithmically generated "AI slop" and synthetic noise collapses public reliance on digital media entirely. This creates a fractured environment where objective truth is discarded, and reality is decided solely by in-group tribal affiliation and confirmation bias, severely undermining the functioning of democratic institutions18.

14. Research Gaps#

Despite rapidly expanding academic literature and forensic capabilities, critical research gaps remain that hinder effective policy responses. The long-term psychological impacts of sustained exposure to hyper-realistic synthetic media on baseline societal anxiety and democratic participation are currently poorly understood. Furthermore, while Zero-Knowledge Proofs (ZKPs) offer a highly robust theoretical solution to the privacy and surveillance limitations of current provenance standards like C2PA, the computational overhead and latency of executing ZKPs natively on mobile camera hardware require extensive engineering research to achieve viability at scale47. Finally, rigorous empirical studies are urgently required to determine whether the mandatory labeling of deepfakes (as required by the EU AI Act) effectively mitigates psychological harm, or if it inadvertently creates a false sense of security for unlabeled, yet equally deceptive, manipulative media.

15. Conclusion#

AI-enabled deepfake psychological operations represent a fundamental and escalating challenge to the epistemic foundation of modern society. The asymmetric advantage inherently favors the attacker: generating hyper-realistic synthetic media requires minimal financial resources and technical skill, while forensic detection, cryptographic verification, and psychological debunking demand immense time, coordinated institutional effort, and highly specialized expertise. As demonstrated by global case studies ranging from localized extortion to international armed conflict, the threat is profoundly compounded by human cognitive vulnerabilities. Specifically, confirmation bias and the liar's dividend allow deepfakes to persuade populations effortlessly, while empowering wrongdoers to deny genuine evidence of misconduct. Defending against this complex threat matrix requires a multi-layered, interdisciplinary approach. It necessitates advancing privacy-preserving cryptographic provenance architectures like Zero-Knowledge Proofs, aggressively updating international humanitarian and domestic legal frameworks to address synthetic warfare, and fundamentally restructuring public communication strategies toward prebunking to build cognitive resilience against emotional deception.

16. Annotated Bibliography#

• Boneh, D. (2024). Using ZK Proofs to Fight Disinformation. A foundational technical breakdown detailing how Zero-Knowledge Proofs can be utilized to verify the provenance of digital media. The author explains how ZKPs allow for legitimate media editing (such as cropping or redaction for source protection) without ever revealing the original signed photo, offering a secure, privacy-preserving alternative to the centralized trust models currently dominating the industry6.
• **Chesney, R., & Citron, D. (2019). Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security. California Law Review.** This seminal legal text introduces the concept of the "liar’s dividend," arguing presciently that the proliferation of deepfakes will allow malicious actors to escape accountability by falsely labeling genuine, incriminating audiovisual evidence as AI-generated, fundamentally altering the nature of legal and political evidence2.
• **Golaszewski, E., et al. (2026). Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short. arXiv:2604.24890.** A highly critical, independent security analysis of the C2PA standard. The authors utilize formal-methods analysis to demonstrate that C2PA fails to achieve its core security goals, documenting specific structural vulnerabilities including timestamp forgery, the widespread acceptance of revoked credentials, and the devastating impact of undetectable metadata stripping4.
• **Grohmann, L., Halle, F. A., & Appel, M. (2026). Deepfake! A Liar's Dividend for Audiovisual Material. Psychology of Popular Media.** An essential empirical study confirming that politicians who falsely claim authentic video evidence is a deepfake do indeed reap a significant liar's dividend. The study proves that in low-trust media environments, deceptive politicians experience higher perceived leadership abilities while concurrently driving down the public's general trust in the press1.
• **Lucas, C., et al. (2024). Political deepfake videos are no more deceptive than other fake news. Journal of Politics.** This rigorous behavioral experiment reveals that while deepfake videos successfully deceive a large portion of the public (approximately 42%), they are not statistically more persuasive than traditional text or audio disinformation. The study underscores that partisan-motivated reasoning—rather than technical realism—is the primary cognitive driver of deception in political contexts21.
• **Schiff, K. J., et al. (2025). The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability? American Political Science Review.** A comprehensive survey experiment analyzing how politicians utilize false claims of "fake news" and "deepfakes" to survive scandals. The study found that while invoking informational uncertainty effectively protects politicians facing text-based scandals, it is less effective against raw video evidence, highlighting the nuanced application of the liar's dividend prior to the total collapse of media trust51.

Works cited#

1. Deepfake! A Liar's Dividend for Audiovisual Material - Ovid, https://www.ovid.com/journals/popmed/pdf/10.1037/ppm0000665\~deepfake-a-liars-dividend-for-audiovisual-material 2. Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security, https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security 3. A LIAR'S DIVIDEND 1 Deepfake! A Liar's Dividend for Audiovisual Material Lara Grohmann, Franziska A. Halle, and Markus Appel, https://www.mcm.uni-wuerzburg.de/fileadmin/06110000/2026/Grohmann\_\_Halle\_\_\_Appel\_2026\_\_Preprint\_.pdf 4. Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short - arXiv, https://arxiv.org/html/2604.24890v1 5. Robustness Assessment and Enhancement of Text Watermarking for Google's SynthID, https://arxiv.org/html/2508.20228v1 6. Using ZK Proofs to Fight Disinformation | by Dan Boneh - Medium, https://medium.com/@boneh/using-zk-proofs-to-fight-disinformation-17e7d57fe52f 7. Challenges in Deepfake Authentication: Harmonizing International Rules for Criminal Evidence, https://ijlmh.com/paper/challenges-in-deepfake-authentication-harmonizing-international-rules-for-criminal-evidence/ 8. Emerging need to regulate deepfakes in international law: the Russo–Ukrainian war as an example - Oxford Academic, https://academic.oup.com/cybersecurity/article-pdf/11/1/tyaf008/63133008/tyaf008.pdf 9. School principal was framed using AI-generated racist rant, police say. A co-worker is now charged. - CBS News, https://www.cbsnews.com/baltimore/news/maryland-framed-principal-racist-ai-generated-voice/ 10. Commentary: Deepfakes are still new, but 2024 could be the year they have an impact on elections - CNA, https://www.channelnewsasia.com/commentary/deepfakes-impact-election-tech-regulation-4214396 11. How to Identify and Investigate AI Audio Deepfakes, a Major 2024 Election Threat, https://gijn.org/resource/tipsheet-investigating-ai-audio-deepfakes/ 12. AI tools usage for disinformation in the war in Ukraine - DFRLab, https://dfrlab.org/2024/07/09/ai-tools-usage-for-disinformation-in-the-war-in-ukraine/ 13. Incident 573: Deepfake Recordings Allegedly Influence Slovakian Election, https://incidentdatabase.ai/cite/573/ 14. Deepfake Satire and the Possibilities of Synthetic Media | Afterimage - UC Press Journals, https://online.ucpress.edu/afterimage/article/50/3/81/197199/Deepfake-Satire-and-the-Possibilities-of-Synthetic 15. A Deep Dive Into the Pikesville, MD, Racist/Antisemitic Deepfake Audio - EDRM, https://edrm.net/2026/07/a-deep-dive-into-the-pikesville-md-racist-antisemitic-deepfake-audio/ 16. Deepfake Fight: AI-Powered Disinformation and Perfidy Under the Geneva Conventions - NDLScholarship, https://scholarship.law.nd.edu/cgi/viewcontent.cgi?article=1035\&context=ndlsjet 17. Cognitive Warfare: Generative AI, False Realities, and International Humanitarian Law, https://digitalcommons.wcl.american.edu/cgi/viewcontent.cgi?article=1156\&context=research 18. F is for Fake: How AI Changes What We Believe and what that means for Cyber Security, https://www.compromisingpositions.co.uk/podcast/episode-62-fake 19. Fooled twice: People cannot detect deepfakes but think they can - PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC8602050/ 20. SynthID: A Technical Deep Dive into Google's AI Watermarking Technology - Medium, https://medium.com/@karanbhutani477/synthid-a-technical-deep-dive-into-googles-ai-watermarking-technology-0b73bd384ff6 21. Political deepfake videos no more deceptive than other fake news, research finds, https://source.washu.edu/2024/08/political-deepfake-videos-no-more-deceptive-than-other-fake-news-research-finds/ 22. The detection of political deepfakes - Oxford Academic, https://academic.oup.com/jcmc/article/27/4/zmac008/6650406 23. 5 Shocking Cases of AI-Generated Deepfakes Interfering in Global Politics | Incode, https://www.incode.com/blog/5-shocking-cases-of-ai-generated-deepfakes-interfering-in-global-politics 24. Deepfakes and the Geneva Conventions: Does Deceptive AI-Generated Misinformation Directed at an Enemy During Armed Conflict Violate International Humanitarian Law? A Critical Discussion - MDPI, https://www.mdpi.com/2075-471X/14/6/83 25. Deepfake Fight: AI-Powered Disinformation and Perfidy Under the Geneva Conventions, https://www.researchgate.net/publication/356383164\_Deepfake\_Fight\_AI-Powered\_Disinformation\_and\_Perfidy\_Under\_the\_Geneva\_Conventions 26. Undermining Ukraine: How the Kremlin Employs Information Operations to Erode Global Confidence in Ukraine - Atlantic Council, https://www.atlanticcouncil.org/wp-content/uploads/2023/02/Undermining-Ukraine-Final.pdf 27. Zelenskyy deepfake reflects new front in Ukraine conflict - Information Age | ACS, https://ia.acs.org.au/article/2022/zelenskyy-deepfake-reflects-new-front-in-ukraine-conflict.html 28. PARLIAMENTARY HANDBOOK ON DISINFORMATION, AI AND SYNTHETIC MEDIA, https://www.cpahq.org/media/sphl0rft/handbook-on-disinformation-ai-and-synthetic-media.pdf 29. How Generative AI Is Transforming Journalism: Development, Application and Ethics - MDPI, https://www.mdpi.com/2673-5172/5/2/39 30. Athletic Director Charged in Pikesville High School AI Case | Baltimore County Government, https://www.baltimorecountymd.gov/departments/police/news/athletic-director-charged-pikesville-high-school-ai-case 31. Former Pikesville High School principal sues Baltimore County Schools over racist AI case, https://www.cbsnews.com/baltimore/news/pikesville-high-school-principal-sues-baltimore-county-schools-racist-ai-recording/ 32. Former principal sues Baltimore County schools over alleged racist AI deepfake | K-12 Dive, https://www.k12dive.com/news/baltimore-county-schools-lawsuit-principal-deepfake/737105/ 33. Tools for trustworthy AI - IBM, https://www.ibm.com/think/insights/ai-ethics-tools 34. AI Deepfake CFO Scam Nets HK$200 Million - OECD.AI, https://oecd.ai/en/incidents/2024-02-02-e719 35. Concern Grows as 'Deepfakes' Spread Misinformation - Africa Defense Forum, https://adf-magazine.com/2023/04/concern-grows-as-deepfakes-spread-misinformation/ 36. The Bizarre and Terrifying Case of the “Deepfake” Video that Helped Bring an African Nation to the Brink - Mother Jones, https://www.motherjones.com/politics/2019/03/deepfake-gabon-ali-bongo/ 37. Gabon's Bongo Opens Summit After Stroke - VOA, https://www.voanews.com/a/africa\_gabons-bongo-opens-summit-after-stroke/6181268.html 38. Gabon Coup Attempt May Have After-Effects - VOA, https://www.voanews.com/a/gabon-coup-attempt-may-have-after-effects/4733811.html 39. (PDF) Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short, https://www.researchgate.net/publication/404281688\_Verifying\_Provenance\_of\_Digital\_Media\_Why\_the\_C2PA\_Specifications\_Fall\_Short 40. Verifying Provenance of Digital Media: Why the C2PA Specifications Fall Short - arXiv, https://arxiv.org/abs/2604.24890 41. What Is C2PA? The Standard, Its Metadata and Real Limits - TrueScreen, https://truescreen.io/articles/c2pa-standard-history-limitations/ 42. C2PA vs CPP: Why Content Provenance Needs a New Approach for the Forensic Era, https://veritaschain.org/blog/posts/2026-01-25-c2pa-vs-cpp-comparison/ 43. Google's SynthID: A Guide With Examples - DataCamp, https://www.datacamp.com/tutorial/synthid 44. \[2508.20228\] Robustness Assessment and Enhancement of Text Watermarking for Google's SynthID - arXiv, https://arxiv.org/abs/2508.20228 45. ZKPROV: A Zero-Knowledge Approach to Dataset Provenance for Large Language Models, https://arxiv.org/html/2506.20915v1 46. How to Beat Deep Fakes (Part 1\) - Ingonyama, https://www.ingonyama.com/post/how-to-beat-deep-fakes-using-zero-knowledge-cryptography-for-audio-video-and-image-verification 47. \[2602.11887\] Verifiable Provenance of Software Artifacts with Zero-Knowledge Compilation, https://arxiv.org/abs/2602.11887 48. Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security - Scholarly Commons at Boston University School of Law, https://scholarship.law.bu.edu/cgi/viewcontent.cgi?article=1640\&context=faculty\_scholarship 49. RPTR FORADORI EDTR CRYSTAL NATIONAL SECURITY CHALLENGES OF ARTIFICIAL INTELLIGENCE, MANIPULATED MEDIA, AND "DEEPFAKES", https://docs.house.gov/meetings/IG/IG00/20190613/109620/HHRG-116-IG00-Transcript-20190613.pdf 50. GenAI Mirage: The Impostor Bias and the Deepfake Detection Challenge in the Era of Artificial Illusions - arXiv, https://arxiv.org/html/2312.16220v2 51. The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability? | American Political Science Review - Cambridge University Press & Assessment, https://www.cambridge.org/core/journals/american-political-science-review/article/liars-dividend-can-politicians-claim-misinformation-to-evade-accountability/687FEE54DBD7ED0C96D72B26606AA073 52. The prebunking playbook for communicators |, https://elmcommunications.com.au/2025/08/21/the-prebunking-playbook-for-communicators/ 53. The 'Truth Sandwich' and 11 Other Ways to Combat Election Misinformation with your Reporting - PEN America, https://pen.org/combat-election-misinformation-with-your-reporting/ 54. How to Spot Fake News: A Step-by-Step Detection Guide | MastersInCommunications.org, https://www.mastersincommunications.org/truth-about-fake-news/ 55. AI deepfake policy in California - Ballotpedia, https://ballotpedia.org/AI\_deepfake\_policy\_in\_California 56. Defending Democracy from Deepfake Deception Act of 2024 - Senate Judiciary Committee, https://sjud.senate.ca.gov/system/files/2024-06/ab-2655-berman-sjud-analysis.pdf 57. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 Brian R. Chavez-Ochoa (CA Bar No. 190289)\* brianr@cha, https://hlli.org/wp-content/uploads/2024/09/kohls.78.Ps-opp-to-DMSJ-AB-2839.pdf 58. Federal Judge Strikes Down California Deepfake Law - The Conference Board, https://www.conference-board.org/research/ceo-center-newsletters-alerts/federal-judge-strikes-down-california-deepfake-law 59. Deepfakes pose an obvious peril in politics, but California's bans amount to censorship, https://calmatters.org/commentary/2025/08/deepfake-politics-california-law-censorship/ 60. Take It Down Act enforcement starts now: What to know about the FTC and TIDA, https://www.ftc.gov/business-guidance/blog/2026/05/take-it-down-act-enforcement-starts-now-what-know-about-ftc-tida 61. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act#:\~:text=Under%20the%20AI%20Act%2C%20deployers,review%20or%20editorial%20control%20(Article 62. research brief - digital disinformation operations in armed conflict - The Geneva Academy of International Humanitarian Law and Human Rights, https://archives.geneva-academy.ch/joomlatools-files/docman-files/Digital%20disinformation%20operations%20in%20Armed%20Conflict%20(1).pdf.pdf) 63. The new era of disinformation wars - Völkerrechtsblog, https://voelkerrechtsblog.org/de/the-new-era-of-disinformation-wars/ 64. Toward Clarifying the Gap Between Prohibited Perfidy and Ruses, https://doshisha.repo.nii.ac.jp/record/29547/files/028004340003.pdf 65. Safeguarding human values: rethinking US law for generative AI's societal impacts - PMC, https://pmc.ncbi.nlm.nih.gov/articles/PMC12058884/ 66. From AI Rights to Neuro Privacy, Cybersecurity Law Struggles to, https://me.pcmag.com/en/security/29652/from-ai-rights-to-neuro-privacy-cybersecurity-law-struggles-to-keep-up