# *Entry* **Synthetic Media: Deepfakes, AI-Generated Content, and Authenticity in the Digital Society**

**Dan Valeriu Voinea**

Department of Arts and Media, University of Craiova, 200585 Craiova, Romania; dan.voinea@edu.ucv.ro or dan.voinea@gmail.com

#### **Definition**

Synthetic media are digital artifacts (image, video, audio, text, or multimodal content) that are wholly or partly generated or materially manipulated by artificial intelligence (AI), particularly by deep-learning models. In consequence, their form, source, identity signal, or evidentiary relation to recorded events becomes partly or wholly artificial. A deepfake is the best-known subclass: AI-generated or AI-manipulated image, audio, or video content (including audio-only voice clones and still images) that realistically depicts an existing or fictitious person, object, place, or event and could falsely appear to be authentic. Contemporary scholarly and legal usage defines deepfakes by their technological origin and their capacity to deceive rather than by the creator's intent, so disclosed and beneficial applications (accessibility, dubbing, entertainment, and research) are synthetic media as well. They are distinguished from cheapfakes (or shallowfakes), which achieve deception through conventional, non-AI editing such as selective cropping, slowing, or recontextualization. The social significance of synthetic media is not intrinsic but depends on consent, context, disclosure, and distribution, and on the institutional conditions under which audiences judge authenticity across the expanding volume of AI-generated content (AIGC) in the digital society.

**Keywords:** synthetic media; deepfakes; AI-generated content; generative AI; media authenticity; content provenance; deepfake detection; disinformation

# **1. Introduction**

Synthetic media denote the class of images, sounds, videos, and texts whose production or alteration relies substantially on artificial intelligence (AI). The domain crystallized around the neologism "deepfake," a portmanteau of "deep learning" and "fake" that emerged in late 2017 when a Reddit user, and the community named "deepfakes," began posting AI face-swapped pornographic videos and released consumer face-swapping tools [\[1,](#page-12-0)[2\]](#page-12-1). The term rapidly generalized from that origin to name any hyper-realistic, AImanipulated depiction of a person, and then, as generative models matured, to sit within the broader umbrella of synthetic media alongside AI-generated content (AIGC) [\[1,](#page-12-0)[3\]](#page-12-2). Deepfakes are best understood as the most salient subclass of synthetic media, spanning image, audio, and video, not as a separate technology [\[2\]](#page-12-1).

The technical genealogy of the field is usually traced to the introduction of generative adversarial networks (GANs) in 2014, which established the modern paradigm of learning to generate realistic data [\[4\]](#page-12-3). The subsequent decade moved from research-grade face reenactment to consumer face-swap applications, and then, from about 2021, to diffusionbased text-to-image, text-to-audio, and text-to-video systems that placed high-fidelity

Academic Editors: Stylianos Mystakidis and Sandro Nuno Ferreira Serpa

Received: 5 July 2026 Revised: 13 August 2026 Accepted: 14 August 2026 Published: 18 August 2026

**Copyright:** © 2026 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the [Creative Commons](https://creativecommons.org/licenses/by/4.0/) [Attribution \(CC BY\) license.](https://creativecommons.org/licenses/by/4.0/)

generation in the hands of non-specialists. This diffusion of capability transformed synthetic media from a niche curiosity into an infrastructural feature of the contemporary information environment, described in legal scholarship as the arrival of a "synthetic society" in which any recorded appearance can, in principle, be fabricated or manipulated [\[5\]](#page-12-4).

The topic matters to the digital society in two ways. Synthetic media strain the evidentiary value that audiences have historically assigned to recordings, raising the prospect that authentic material can be dismissed as fake, an effect labeled the "liar's dividend" [\[6\]](#page-12-5). What synthetic media place under pressure is therefore not this or that recording but the relationship between authenticity, evidence, and public trust on which mediated communication rests; that relationship is the organizing concern of this entry. The same technologies also underpin legitimate and beneficial applications in accessibility, entertainment, education, and privacy protection, so synthetic media are better analyzed as a dual-use capability than as an intrinsic harm.

This entry approaches synthetic media from the perspective of the digital society: its defining contribution is a social-science, media-theory, and comparative-governance synthesis of what synthetic media do to authenticity, evidence, and public trust, and of how jurisdictions worldwide are responding, rather than a technical survey of generation and detection systems. On that footing, it synthesizes established scholarly and regulatory knowledge as of mid-2026, surveying competing definitions, a working typology, landmark technical milestones, detection and its documented limits, provenance and authentication, regulation and governance, societal and epistemic impacts, the media-theoretical framing of authenticity, and, closing the dual-use account, beneficial applications, before setting out open challenges. The account is deliberately bounded: it asserts only established findings, documented events, and enacted or formally proposed law, and presents genuinely contested questions as open debates. The focus accordingly falls on definitions, authenticity and evidence, human perception and trust, comparative regulation, and the media theory of the authentic; the intervening technical sections (Sections 4–6) are kept compact and reliability-oriented, and readers are pointed to related published work on the metaverse [\[7\]](#page-12-6) and on large language models [\[8\]](#page-12-7) rather than to re-explanations of generative-model mechanics or immersive-environment specifics.

A brief note on sources and method: as an encyclopedia entry, this text is a narrative synthesis rather than a systematic review. It rests on peer-reviewed scholarship (surveys, meta-analyses, and primary studies), on the official texts of the legal and regulatory instruments discussed, and, for developments not yet absorbed by the scholarly literature, on primary policy documents and documented incident reports, with coverage of the literature and of legal status through July 2026. Peer-reviewed and official primary sources are preferred wherever both exist; preprints, vendor announcements, and single-study or workshop-stage results are cited only with their provisional standing flagged. The same convention is applied throughout so that evidentiary weight stays visible: findings are identified in the text as experimental results, meta-analytic estimates, documented incidents, vendor-reported deployments, or early-stage demonstrations, and where authoritative sources genuinely conflict, as with the definition of the deepfake, the entry reports the disagreement rather than resolving it silently [\[3\]](#page-12-2).

# **2. Terminology and Competing Definitions**

No full scholarly consensus exists on the definition of "deepfake." A meta-review mapping the divergence across fifteen survey papers found that authors disagree on the term's boundaries even as its use has become ubiquitous [\[3\]](#page-12-2). Three recurring fault lines organize the disagreement: whether a deepfake must be produced with deep learning specifically, whether it must carry malicious or deceptive intent, and whether it must depict a real, identifiable person [\[3\]](#page-12-2). The most-cited definitional sources illustrate the spread. An influential early review characterizes deepfakes as hyper-realistic videos that apply AI to depict people saying and doing things that never happened [\[1\]](#page-12-0). A widely used working definition emphasizes AI-generated or AI-manipulated content with a high potential to deceive, while treating deception as typical but not strictly definitional [\[2\]](#page-12-1). A foundational law-review treatment restricts the paradigm case to audio and video of real people manipulated through machine learning [\[6\]](#page-12-5).

The best-supported resolution across the literature is that malicious intent is typical but not definitional: disclosed, benign, and beneficial synthetic media remain synthetic media [\[2\]](#page-12-1). This functional stance is reflected in binding law. The European Union's AI Act defines a deepfake as AI-generated or manipulated image, audio, or video content that resembles existing persons, objects, places, entities, or events and would falsely appear to a person to be authentic or truthful, a definition that turns on realism and capacity to appear authentic rather than on the creator's intent and that explicitly extends beyond real persons to objects, places, events, and realistic depictions of fictitious people [\[9\]](#page-12-8).

Several neighboring terms complete the vocabulary. AI-generated content (AIGC) is a modality-neutral umbrella covering text, images, audio, video, and virtual scenes generated or synthesized by AI. It is the dominant framing in Chinese regulation and platform policy [\[10\]](#page-12-9). "Generative media" refers more broadly to the output of generative models such as GANs and later diffusion and autoregressive systems, the latter generating output sequentially, element by element [\[4\]](#page-12-3).

Read together, these sources support a single operational framework, applied throughout what follows. First, AI involvement in generation or material manipulation is the boundary criterion: it separates synthetic media (AIGC, the umbrella preferred in regulatory usage, overlaps with but is not synonymous with this category: AIGC foregrounds content generated by AI, whereas synthetic media as defined here also include authentic material substantially manipulated by AI) from cheapfakes (or shallowfakes), which achieve deception through conventional, non-AI editing such as slowing, splicing, mislabeling, or recontextualizing footage. The criterion is the production method, not the degree of realism [\[11\]](#page-12-10). Second, within synthetic media, modality distinguishes synthetic still images, synthetic audio (including voice clones), synthetic video, synthetic text, and multimodal combinations. Third, the deepfake is the depiction-centered subset: content that realistically depicts persons, objects, places, or events and could falsely appear authentic. On the AI Act definition adopted here, that subset spans still images, audio-only voice clones, and video alike, although some scholarly definitions confine the term to moving audiovisual content, an openly acknowledged divergence [\[3,](#page-12-2)[6](#page-12-5)[,9\]](#page-12-8). Fourth, hybrid, partially edited content (authentic material substantively altered with AI tools) falls within synthetic media under the "materially manipulated" clause, whereas conventionally edited hybrids remain cheapfakes [\[2,](#page-12-1)[11\]](#page-12-10). Text-only output of large language models is therefore AIGC and synthetic media in the broad sense, but a deepfake only on the minority of definitions that extend the term beyond depiction [\[3\]](#page-12-2). The boundary matters for governance: because cheapfakes remain common, effective, and inexpensive to produce, a framework focused only on AI-generated content risks leaving conventional manipulation unaddressed [\[11\]](#page-12-10). The umbrella-and-subclass structure now standard in the literature follows accordingly: synthetic media is the superordinate category, and the deepfake its most recognizable member [\[1](#page-12-0)[,2\]](#page-12-1).

# **3. A Typology of Synthetic Media**

Classification schemes for synthetic media combine a modality axis with a purpose axis. On the technical side, the canonical face-manipulation taxonomy distinguishes four operations: entire-face synthesis, identity or face swap, attribute manipulation, and expression swap or reenactment [\[12\]](#page-13-0). Modern surveys extend this visual scheme to a multimodal one that adds lip-sync and reenactment, full-body and scene synthesis, and audio manipulation through voice conversion, voice cloning, and text-to-speech impersonation [\[13,](#page-13-1)[14\]](#page-13-2). A modality axis spanning text, image, audio, video, and immersive scene has also been adopted by regulators; China's Deep Synthesis Provisions, for example, enumerate these modalities in defining the regulated category [\[15\]](#page-13-3).

The second axis concerns purpose and governance context. The literature repeatedly organizes uses along a spectrum from entertainment, satire, accessibility, education, film production, and privacy-preserving anonymization, to fraud, political disinformation, harassment, and non-consensual intimate imagery (NCII) [\[2\]](#page-12-1). A recurring observation is that the same technical form can be legitimate or harmful depending on consent, disclosure, and distribution, which is why the most useful organizing frame is two-dimensional: modality crossed with governance context [\[5\]](#page-12-4). Under this frame, an identity swap rendered for a consenting actor in a dubbed film occupies a different governance cell from a visually identical swap distributed without consent as intimate imagery, even though the two are technically indistinguishable [\[2\]](#page-12-1). Hybrid, partially AI-edited content occupies the same cells as its wholly synthetic counterparts, since the typology tracks production method and governance context, not the proportion of material generated [\[11\]](#page-12-10). NCII is a salient reference point on the harmful end of this axis: a 2019 census of deepfake videos online reported that the overwhelming majority were non-consensual pornography, almost exclusively targeting women, a finding often described as a "canary in the coalmine" for syntheticmedia governance, though the specific proportion is now dated [\[16\]](#page-13-4). Table [1](#page-3-0) presents a consolidated typology on these two axes.

<span id="page-3-0"></span>**Table 1.** A two-axis typology of synthetic media: modality (rows) × purpose and governance context (the two right-hand columns contrast benign and harmful uses of the same forms). Glosses: entire-face synthesis, a wholly new face; attribute editing, altering features such as age; voice conversion, one voice rendered as another; previsualization, pre-shoot scene planning; astroturfing, fake "grassroots" campaigns. Abbreviations: TTS, text-to-speech; NCII, non-consensual intimate imagery; VFX, visual effects; LLM, large language model.

| Modality                    | Representative Forms                                           | Illustrative Legitimate or                                | Illustrative Harmful or                                              |
|-----------------------------|----------------------------------------------------------------|-----------------------------------------------------------|----------------------------------------------------------------------|
|                             | and Techniques                                                 | Benign Uses                                               | Deceptive Uses                                                       |
| Still image                 | Entire-face synthesis;                                         | Stock and concept imagery;                                | Fabricated "photographic"                                            |
|                             | attribute editing;                                             | dataset anonymization; art                                | evidence; synthetic personas;                                        |
|                             | text-to-image generation                                       | and design                                                | visual disinformation                                                |
| Video (face and identity)   | Face swap; identity and<br>expression reenactment;<br>lip-sync | Film and television VFX;<br>dubbing; satire and parody    | NCII; political impersonation;<br>identity-based harassment          |
| Video (full body and scene) | Full-body and scene synthesis;                                 | Previsualization; education                               | Fabricated events; synthetic                                         |
|                             | text-to-video                                                  | and simulation; accessibility                             | "eyewitness" footage                                                 |
| Audio (voice)               | Voice conversion; voice<br>cloning; TTS impersonation          | Voice banking<br>and accessibility;<br>dubbing; narration | Impersonation and "CEO"<br>fraud; deceptive robocalls                |
| Text                        | LLM generation                                                 | Drafting; translation;<br>summarization                   | Astroturfing; spam and<br>low-grade "slop";<br>scaled disinformation |
| Multimodal and immersive    | Audio-visual avatars; virtual                                  | Gaming; training;                                         | Real-time impersonation;                                             |
|                             | scenes and telepresence                                        | remote collaboration                                      | deceptive avatars                                                    |

# **4. Technical Evolution: Landmark Milestones**

A scope note frames this and the two following sections: the milestones, detection methods, and provenance mechanisms reviewed here concern perceptual synthetic media, that is, still images, audio, and video, where the deepfake problem defined above is concentrated. The generation and detection of synthetic text form a distinct research area with its own methods and a parallel reliability debate, surveyed elsewhere [\[17\]](#page-13-5); text reenters this entry where provenance tooling and governance address it (Sections [6](#page-5-0) and [7\)](#page-6-0). The treatment of generation and detection here is deliberately compact and reliabilityoriented; for architectural depth, from generative-model families to detector taxonomies, benchmark datasets, and evaluation metrics, readers are referred to the complementary technical entry published in this journal by Gazis et al. [\[18\]](#page-13-6).

The generation lineage of synthetic media is marked by a sequence of documented milestones. Generative adversarial networks, introduced in 2014, are conventionally treated as the start of the modern generative lineage, not as a deepfake technology in their own right [\[4\]](#page-12-3). Real-time facial reenactment was demonstrated in 2016, showing that a source performance could drive a target face in a video stream [\[19\]](#page-13-7). In 2017, audio-driven lipsync of a public figure mapped speech onto realistic mouth movements; the same year, autoencoder-based face-swap tools spread through online communities and gave the field its name [\[1,](#page-12-0)[20\]](#page-13-8). Consumer tools then put face-swapping and voice cloning within non-expert reach, a democratization that, more than any single result, turned deepfakes into a public-policy concern [\[1\]](#page-12-0). Few-shot voice cloning (adaptation from a few samples) was demonstrated in 2018, when a short recording sufficed to seed a synthetic voice, foreshadowing later fraud capabilities [\[21\]](#page-13-9). High-realism, controllable face synthesis advanced markedly in 2019 with style-based generator architectures [\[22\]](#page-13-10).

A second inflection, text-prompted generation, arrived along two architectural routes. Denoising diffusion probabilistic models, described in 2020, underpin the diffusion lineage: models that learn to reverse a gradual noising process, refining random noise into a coherent output [\[23\]](#page-13-11). Zero-shot text-to-image generation (depicting prompted concepts a model was never explicitly trained on) followed in 2021, achieved not by diffusion but by an autoregressive transformer generating images as sequences of discrete tokens [\[24\]](#page-13-12). Latent diffusion in 2022 then enabled the high-resolution, openly available diffusionbased generators that brought text-to-image synthesis to a mass public [\[25\]](#page-13-13). Text-to-video generation matured next, with systems across 2023–2025 extending synthesis to short clips, exemplified by a model previewed in February 2024 and publicly released in December 2024 [\[26\]](#page-13-14) and a competing family announced in May 2024 [\[27\]](#page-13-15). In parallel, real-time and few-second voice cloning became a scalable consumer and fraud capability during the early 2020s [\[13\]](#page-13-1). The cumulative effect is a general lowering of the skill, time, and cost of producing convincing synthetic media, reframing every subsequent question of detection, provenance, and governance [\[14\]](#page-13-2).

# **5. Detection and Its Documented Limits**

Automated detection of synthetic media is usually grouped into four families: artifact or forensic methods seeking manipulation traces; biological-signal methods tracking physiological cues such as eye-blinking or pulse-related color changes; data-driven deep detectors trained to classify real versus fake; and multimodal-consistency methods testing whether audio and video agree [\[12](#page-13-0)[,14](#page-13-2)[,28\]](#page-13-16). No single family is robust in isolation, and detectors are typically trained and evaluated on curated datasets that imperfectly represent the compression, editing, and unseen generators met once content circulates online [\[14\]](#page-13-2). Progress is measured against public benchmarks, notably FaceForensics++ (one thousand real and four thousand manipulated sequences) [\[29\]](#page-13-17) and the DeepFake

Detection Challenge (DFDC) dataset, a prominent and widely used public benchmark of roughly 128,000 clips [\[30\]](#page-13-18). Gazis et al. [\[18\]](#page-13-6) survey these detector families, benchmark datasets, and evaluation metrics in architectural depth; the present section concentrates instead on what the record establishes about their reliability.

The most stable finding in the detection literature is generalization failure. Detectors that achieve high accuracy on their training distribution (high within-distribution performance is frequently observed, with accuracy and area-under-curve figures reported on curated benchmarks often falling in a 95–99% range, not at a single settled value) degrade sharply on unseen generators and under the compression and re-encoding typical of realworld sharing [\[14](#page-13-2)[,31\]](#page-13-19). The 2020 DFDC results illustrate the gap directly. Final standings were decided by log-loss score on a held-out, black-box test set of 10,000 videos that entrants could not inspect; on that set the winning entry, submitted by Selim Seferbekov, reached 65.18% average precision, a ranking-quality metric that is not interchangeable with classification accuracy. By contrast, the best model on the public leaderboard reached 82.56% average precision, a leaderboard on which the eventual winner had placed only fourth [\[30](#page-13-18)[,32\]](#page-13-20). Alongside this arms-race dynamic, reliability-focused surveys identify two persistent open problems: a transparency–accuracy dilemma (disclosing how a detector works aids evasion, while opacity undermines the scrutiny evidentiary use requires) and the absence of a generally recognized forensic standard tailored to deepfake detection [\[31](#page-13-19)[,33\]](#page-13-21).

The human baseline is weak. A meta-analysis of fifty-six papers, aggregating 137 effects and more than 86,000 participants, placed overall unaided human accuracy at about 55.5%, near chance (a pooled estimate across heterogeneous stimuli, tasks, and participant populations), with performance around 57% for video, 53% for image, and 52% for text, though training, AI assistance, and caricaturization can raise it [\[34\]](#page-13-22). A complementary finding helps explain this fragility: AI-synthesized faces can be indistinguishable from real ones and are, on average, rated slightly more trustworthy [\[35\]](#page-13-23). Humans and machine detectors also tend to err in different cases: in a large comparative experiment, crowds of ordinary observers reached accuracy comparable to the leading detection model, and observers who saw the model's prediction outperformed either alone, although the model's errors could also mislead them [\[36\]](#page-13-24). These results caution against human vigilance as a primary defense and motivate the provenance and governance approaches below.

# <span id="page-5-0"></span>**6. Provenance and Authentication**

Because detection asks whether a classifier predicts that content is synthetic, it differs conceptually from provenance, which asks what is verifiably known about a piece of content's origin and edit history. The leading provenance standard, developed by the Coalition for Content Provenance and Authenticity (C2PA) and surfaced to users as Content Credentials, binds cryptographically signed manifests describing origin and editing, including AI-generation flags, to media assets [\[37\]](#page-13-25). Adoption has broadened across 2023–2026 to include camera manufacturers implementing capture-time signing and firmware support, creative-software vendors, and major AI providers, per industry- and vendor-reported deployments [\[38\]](#page-13-26). A parallel approach is invisible watermarking, exemplified by Google DeepMind's SynthID for AI-generated images, audio, text, and video; a peer-reviewed account in *Nature* specifically describes the text variant, SynthID-Text, for watermarking large-language-model outputs [\[39](#page-14-0)[,40\]](#page-14-1). Several significant industry initiatives are progressing toward layered combinations of signed provenance, watermarking, and verification tooling [\[41\]](#page-14-2).

Both approaches have documented limits. C2PA metadata can be stripped by screenshotting, re-encoding, or re-uploading content, and the absence of credentials is not proof of inauthenticity, since most existing and analogue content carries no manifest [\[41\]](#page-14-2). Wa-

termarks degrade or can be removed under recompression, editing, and adversarial or optimization attacks. A formal analysis has shown that a class of invisible image watermarks is provably removable under a regeneration attack [\[42\]](#page-14-3), one instance of the broader co-evolution of watermarking and removal techniques. A usability study found that provenance signals help but not cleanly: they can improve calibration and reduce trust in deceptive media, yet incomplete or invalid provenance can also cause audiences to distrust honest media [\[43\]](#page-14-4). A recently demonstrated attack, termed an "integrity clash" (so far a single, unreplicated workshop-stage result), illustrates in principle that provenance and watermarking are technically independent layers that can be maliciously desynchronized, producing an "authenticated fake" in which a valid provenance manifest asserts human authorship while the pixels carry an AI watermark [\[44\]](#page-14-5).

These findings frame the recurring detection-versus-provenance debate. No single approach supersedes the others: the position that recurs across technical surveys and industry guidance, and that Gazis et al. [\[18\]](#page-13-6) likewise endorse, is that detection, provenance, watermarking, platform labeling, and media literacy function as complementary layers, none sufficient alone [\[28,](#page-13-16)[41\]](#page-14-2). Platform labeling illustrates the point in practice: major platforms have moved to disclose or label realistic synthetic content, with policies shifting toward broader AI labeling and, in some cases, automatic labeling of undisclosed photorealistic AI, though these policies change frequently and vary by service [\[45](#page-14-6)[–48\]](#page-14-7). Experimentally, such warnings can misfire, raising blanket disbelief in authentic video without improving detection [\[49\]](#page-14-8). Vendor figures reporting that large volumes of content have been watermarked describe deployment reach, not robustness, and should be read accordingly. The practical implication of the credential asymmetry is that provenance functions best as positive evidence of a trustworthy production pipeline rather than as a test that unmarked content is fake, which is why media literacy and institutional verification remain indispensable parts of the layered defense [\[41\]](#page-14-2). Media literacy here means less artifact-spotting (fragile, given near-chance human performance [\[34\]](#page-13-22)) than source-, provenance-, and institution-oriented verification habits. Scalable interventions have modestly but measurably improved mainstream-versus-false news discernment [\[50\]](#page-14-9).

# <span id="page-6-0"></span>**7. Regulation and Governance**

Regulation of synthetic media is fragmented across jurisdictions and changes quickly. Table [2](#page-8-0) consolidates the principal instruments, their status, and their key provisions; the legal status reported in this section and in Table 2 was last verified on 24 July 2026. In the European Union (EU), the AI Act (Regulation (EU) 2024/1689) sets transparency duties in Article 50: providers must mark synthetic output in a machine-readable way (Article 50(2)), and deployers must disclose deepfakes and certain public-interest AI text (Article 50(4)); these obligations apply from 2 August 2026 [\[9\]](#page-12-8). The EU's Digital Services Act (DSA, Regulation (EU) 2022/2065) complements this through systemic-risk duties on very large online platforms and search engines under Articles 34–35, covering civic discourse, elections, and disinformation [\[51\]](#page-14-10). A voluntary Code of Practice on Transparency of AI-Generated Content, published on 10 June 2026, supports standardized disclosure mechanisms for these duties [\[52\]](#page-14-11); on 8 July 2026 the European Commission concluded in its adequacy opinion, with the AI Board's concurring assessment, that the Code adequately covers the obligations of Articles 50(2), (4), and (5), so that signatories may rely on it to demonstrate compliance, although the Code is not itself binding law [\[53\]](#page-14-12).

China has moved earliest on mandatory labeling. Its Deep Synthesis Provisions, in force since 10 January 2023, require conspicuous labeling where deep-synthesis content could cause confusion (Article 17) alongside implicit technical markers (Article 16) [\[15\]](#page-13-3). Interim Measures for Generative AI Services took effect on 15 August 2023 [\[54\]](#page-14-13), and Labelling

Measures, accompanied by the mandatory national standard GB 45438-2025, came into force on 1 September 2025, requiring both explicit (visible) and implicit (metadata) labels on all AIGC [\[10,](#page-12-9)[55\]](#page-14-14). In the United States, the federal TAKE IT DOWN Act (Public Law 119-12) was enacted in May 2025, criminalizing NCII including AI "digital forgeries" and requiring covered platforms to operate a 48-h notice-and-removal process from 19 May 2026 [\[56\]](#page-14-15). The Federal Trade Commission's (FTC) Government and Business Impersonation Rule took effect on 1 April 2024, with a proposed extension to the impersonation of individuals still in rulemaking [\[57](#page-14-16)[,58\]](#page-14-17). Below the federal level, most states have enacted some deepfake law since 2019, and roughly thirty have election- or political-deepfake statutes, several of which have been enjoined or struck down on First-Amendment or intermediary-liability grounds, most prominently in *Kohls v. Bonta*, where California's AB 2839 was held unconstitutional and AB 2655 preempted by Section 230, with an appeal pending [\[59\]](#page-14-18). Because counts change and litigation is active, current counts are best taken from a live, dated tracker, not a frozen figure [\[60\]](#page-14-19). South Korea supplies a fourth, criminal-law-centered model: a 2024 amendment to its Act on Special Cases Concerning the Punishment of Sexual Crimes (Act No. 20459, in force 16 October 2024) criminalized even possessing or viewing sexually explicit deepfakes, removed the former distribution-intent requirement for the creation offense, and raised creation and distribution penalties [\[61\]](#page-15-0). Other jurisdictions have legislated on the same non-consensual-imagery front: Australia's Criminal Code Amendment (Deepfake Sexual Material) Act 2024 strengthened federal offenses for transmitting sexual material depicting a person without consent, expressly covering AI-created or AI-altered depictions [\[62\]](#page-15-1), and the United Kingdom criminalized sharing intimate images that "appear to show" a person through the Online Safety Act 2023 and, from February 2026, creating or requesting such "purported intimate images" of adults without consent through the Data (Use and Access) Act 2025 [\[63,](#page-15-2)[64\]](#page-15-3).

Two prominent measures remain proposals. Denmark's proposed amendment to its Copyright Act (Sections 65 a and 73 a) would grant individuals rights over their body, face, and voice against deepfakes, with carve-outs for satire and news; it rests on a June 2025 cross-party political agreement and a public consultation completed in August 2025, but the bill had not been introduced in the Folketing before Denmark's March 2026 general election, at which pending measures lapsed, so as of mid-2026 it remains a pre-parliamentary draft awaiting re-consultation [\[65,](#page-15-4)[66\]](#page-15-5). Likewise, the United States NO FAKES Act, a proposed federal digital-replica and right-of-publicity framework, remains unenacted: the 2025 bill stalled after introduction, and a 2026 successor was ordered reported by the Senate Judiciary Committee in June 2026 but has passed neither chamber [\[67](#page-15-6)[,68\]](#page-15-7). Together, the instruments reflect distinct regulatory philosophies: prescriptive labeling in China, transparency plus systemic-risk governance in the EU, a fragmented federal-plus-state patchwork in the United States, and criminal-law protection against sexual deepfakes in South Korea, Australia, and the United Kingdom, with no single global model. The scope of this section is deliberately selective: it examines leading regulatory models and makes no claim to comprehensive comparative coverage. India, for example, has issued deepfake advisories under its Information Technology Rules since 2023 [\[69\]](#page-15-8), and many jurisdictions in the Global South are at earlier stages. The evidence, like the impact findings below, skews toward high-income jurisdictions.

<span id="page-8-0"></span>**Table 2.** Consolidated regulatory timeline for synthetic media, 2019–2026; legal status last verified on 24 July 2026. Rows are ordered chronologically; the final two rows are formal proposals that had not been enacted as of that date.

| 10 January 2023<br>15 August 2023<br>31 January 2024/6 February 2026<br>1 April 2024 | Deep Synthesis Provisions,<br>Arts. 16, 17, 23 [15]<br>Interim Measures for Generative<br>AI Services [54]<br>Online Safety Act 2023, ss.<br>187-188; Data (Use and Access)<br>Act 2025, s. 138 [63,64]<br>Government and Business | China; in force<br>China; in force<br>UK; enacted, in force<br>US (FTC); in force | Labelling of confusion-prone deep<br>synthesis; modality definition<br>Governance and filing of<br>generative-AI services<br>Sharing (2024) and creation (2026)<br>offences for intimate images that<br>"appear to show" a person<br>Impersonation remedy;<br>individual-impersonation |
|--------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| April 2024 to 2026<br>1 August 2024 (entry into force)                               | Impersonation Rule [57,58]<br>Platform AI-labelling<br>policies [45–48]<br>AI Act, Reg. (EU) 2024/1689 [9]                                                                                                                         | Meta/YouTube/TikTok;<br>platform policy<br>EU; enacted, Art. 50 phased            | extension proposed<br>Disclosure and labelling of<br>realistic synthetic content<br>Framework in force; transparency                                                                                                                                                                   |
| September 2024 to August 2025                                                        | AB 2839; AB 2655 [59]                                                                                                                                                                                                              | California, US;<br>enjoined/invalidated,<br>appeal pending                        | duties phased<br>Election-deepfake bans held<br>unconstitutional (AB 2839) or<br>Section 230-preempted (AB 2655)                                                                                                                                                                       |
| 2 September 2024 (assent)                                                            | Criminal Code Amendment<br>(Deepfake Sexual Material)<br>Act 2024 [62]                                                                                                                                                             | Australia; enacted, in force                                                      | Federal offences for transmitting<br>non-consensual sexual material<br>incl. AI-created or<br>AI-altered depictions                                                                                                                                                                    |
| 16 October 2024 (in force)                                                           | Act on Special Cases Concerning<br>the Punishment of Sexual Crimes,<br>Act No. 20459 [61]                                                                                                                                          | South Korea; enacted                                                              | Criminalizes possession or<br>viewing of sexually explicit<br>deepfakes; raised creation and<br>distribution penalties                                                                                                                                                                 |
| 19 May 2025                                                                          | TAKE IT DOWN Act, Pub.<br>L. 119-12 [56]                                                                                                                                                                                           | US federal; enacted                                                               | Criminalizes NCII incl. AI<br>forgeries; platform 48-h takedown<br>from 19 May 2026                                                                                                                                                                                                    |
| 1 September 2025                                                                     | Labelling Measures<br>+ GB 45438-2025 [10,55]                                                                                                                                                                                      | China; in force                                                                   | Explicit visible and implicit<br>metadata labels on all AIGC                                                                                                                                                                                                                           |
| 2 August 2026                                                                        | AI Act Article 50 transparency<br>duties [9,52,53]                                                                                                                                                                                 | EU; enacted, applicable from<br>2 August 2026                                     | Provider machine-readable<br>marking 50(2); deployer<br>disclosure 50(4); Code of Practice<br>assessed adequate 8 July 2026                                                                                                                                                            |
| 2025 onward                                                                          | NO FAKES Act (S. 1367;<br>successor S. 4591) [67,68]                                                                                                                                                                               | US federal; formal proposal,<br>not enacted                                       | Federal<br>digital-replica/right-of-publicity<br>framework; successor reported<br>from committee June 2026                                                                                                                                                                             |
| 2025–2026 (lapsed at March 2026<br>election; re-consultation pending)                | Copyright Act amendment,<br>Sections 65 a/73 a [65,66]                                                                                                                                                                             | Denmark; pre-parliamentary<br>draft, not introduced                               | Likeness and voice rights vs.<br>deepfakes; satire and<br>news carve-outs                                                                                                                                                                                                              |

# **8. Societal and Epistemic Impacts**

Empirically, the central question is what synthetic media actually do to trust and knowledge. A frequently cited UK survey experiment (N = 2005) found that deepfakes tend to leave audiences more uncertain than actively deceived, and that this uncertainty in turn reduces trust in news encountered on social media [\[73\]](#page-15-11). The best-supported effect is therefore trust corrosion rather than reliable mass persuasion, a distinction that counsels against "infocalypse" (information-apocalypse) rhetoric, although the experimental base remains narrow, largely single-country survey experiments [\[73\]](#page-15-11). The complementary concept of the "liar's dividend" holds that, as public awareness of fakery grows, bad actors can more easily dismiss authentic evidence as fabricated [\[6\]](#page-12-5). Direct experimental evidence now exists: in five survey experiments (>15,000 US participants), false claims that scandal reports were "misinformation" increased politician support for text-based reports but were largely ineffective against video evidence [\[74\]](#page-15-12). Interventions intended to raise public skepticism, including some labeling schemes, can thus be turned against the truth, deepfake warnings themselves breeding indiscriminate disbelief in authentic video [\[6](#page-12-5)[,49\]](#page-14-8).

Documented real-world harms fall into several categories; the cases that follow are incident reports, which establish that a harm class exists and what it can cost, not evidence of how widespread it is. In fraud, a widely reported 2024 case saw a Hong Kong finance employee deceived by a deepfake video conference impersonating senior executives into transferring about HK\$200 million (roughly US\$25.6 million) across fifteen transactions [\[75\]](#page-15-13). The multinational engineering firm Arup later confirmed it was the victim [\[76\]](#page-15-14). In electoral manipulation, an AI-voice robocall imitating a US president circulated ahead of the January 2024 New Hampshire primary; the Federal Communications Commission (FCC) fined the responsible operative US\$6 million in a forfeiture order issued in September 2024 [\[77\]](#page-15-15), although a New Hampshire jury acquitted him of all related state voter-suppression and candidate-impersonation charges in June 2025, illustrating the uneven legal response to such incidents [\[78\]](#page-15-16). In image-based abuse, beyond the dated 2019 census noted above [\[16\]](#page-13-4), a ten-country survey of more than 16,000 respondents provides a contemporary crossnational baseline, reporting about 2.2% self-reported victimization and 1.8% perpetration for non-consensual synthetic intimate imagery, figures that rest on self-reports from online panels and carry the attendant reporting and sampling bias [\[79\]](#page-15-17).

A further, still-emerging concern is the flooding of the information ecosystem with low-grade synthetic content, colloquially "AI slop." The phenomenon is documented qualitatively as a source of clutter that degrades search, attention, and moderation, and its cultural salience is marked by the selection of "slop" as a dictionary word of the year for 2025, but ecosystem-level empirical measurement remains thin [\[80\]](#page-15-18). At the level of theory, the "epistemic backstop" argument holds that recordings have historically backstopped human testimony, and that widespread deepfakes erode that backstop and reduce the information a video reliably carries [\[81](#page-15-19)[,82\]](#page-15-20). This claim is genuinely contested: while some philosophers argue that deepfakes pose a grave epistemic threat, others contend that the catastrophist framing is overstated because audiences cross-check testimony coherentistically (against a web of corroborating sources) instead of trusting any single recording in isolation [\[83\]](#page-15-21). Reviews that separate established findings from alarmism conclude that the strongest evidence supports uncertainty and trust erosion, while claims of decisive electoral persuasion or an imminent epistemic collapse remain unproven [\[84\]](#page-16-0).

# **9. Authenticity in Media Theory**

Media and communication theory frame authenticity as the deeper issue beneath the deepfake. The realist tradition of photographic theory, associated with Bazin [\[85\]](#page-16-1), Barthes [\[86\]](#page-16-2), and the semiotics of Peirce [\[87\]](#page-16-3), treats the photograph as an index: a causal trace of what was in front of the lens. Analytic accounts of "transparent" pictures develop the same intuition that photographs put us in a special evidentiary relation to their subjects [\[88\]](#page-16-4). On this realist account, synthetic media sever the indexical link while preserving iconic resemblance: the image still looks like its subject, but it no longer stands in a causal relation to a real event. The epistemic-backstop and epistemic-threat arguments formalize that severing [\[81](#page-15-19)[,82\]](#page-15-20). The realist premise is contested: a constructivist countertradition holds that photographs were always stageable and retouchable, their evidentiary

authority an institutional achievement rather than a natural property, a point pressed by deflationists in the epistemic-apocalypse debate [\[83\]](#page-15-21). Even so, the distinction between iconic resemblance and indexical trace clarifies why a perfect synthetic likeness can be at once convincing and evidentially empty, and why authentication increasingly rests on provenance signals external to the image, not on inspecting the image alone [\[81\]](#page-15-19).

Later media theory complicates the realist picture from three directions, each anticipating generative media. Bolter and Grusin's account of remediation holds that new media refashion older media under a twin logic of immediacy and hypermediacy, so that the "authentic" feel of any medium is itself a produced and continually renegotiated effect [\[89\]](#page-16-5); on this view, synthetic media do not so much break a naturally trustworthy medium as extend the process by which mediation is alternately effaced and foregrounded. Manovich reaches a complementary conclusion from the side of production: once images are numerical representations, programmable like any other data, digital cinema becomes a particular case of animation, closer to painting than to recording, and the generated image is the ordinary condition of computational media rather than an aberration [\[90\]](#page-16-6). Gunning, finally, cautions against resting theories of cinematic realism on indexicality at all, locating the "impression of reality" in movement and the spectator's participatory perception rather than in the photograph's causal trace [\[91\]](#page-16-7). Read against the realist tradition, these positions suggest that the deepfake crisis is less the loss of an intrinsically credible image than the breakdown of historically specific conventions for certifying images, which is precisely why contemporary responses turn to externally certified provenance and institutional verification rather than to renewed faith in inspection of the image itself [\[37,](#page-13-25)[81\]](#page-15-19).

A second theoretical anchor is the uncanny valley, the affective discomfort provoked by near-human but imperfect synthetic faces and voices [\[92\]](#page-16-8). The concept has been extended to synthetic avatars and doppelgängers, where subtle mismatches can depress trust and comfort even when audiences cannot articulate what is wrong [\[93\]](#page-16-9). A third strand treats authenticity not as an intrinsic property of a file but as a communicative achievement: "mediated authenticity" is constructed and negotiated through genre cues, production conventions, and audience expectations, which explains why disclosed or synthetic content can still "feel" authentic, and why labeling does not automatically neutralize persuasive effect [\[94\]](#page-16-10). Together, these traditions reframe the policy problem: the goal is less to guarantee that a given artifact is "real" than to preserve the institutional and perceptual conditions under which audiences can reasonably calibrate their trust, a task that also underlies the evidentiary standards used to authenticate video in legal settings [\[33\]](#page-13-21).

# **10. Beneficial and Legitimate Applications**

The technology that enables the harms above also supports substantial legitimate uses. In accessibility, personalized text-to-speech (TTS) that blends a person's residual voice with donor recordings supports augmentative and alternative communication (perceptual evaluation of the resulting voices reports promising intelligibility), and AI voice generation has been used to preserve or restore the voices of people with degenerative conditions [\[95\]](#page-16-11). In localization and media production, AI dubbing enables lip-synced multilingual delivery, and film and television increasingly rely on synthesis for visual effects and post-production, uses that are disclosed and consensual by design [\[2\]](#page-12-1). A widely cited 2019 public-health campaign used AI visual dubbing to deliver one spokesperson's anti-malaria appeal in nine languages, impersonation-grade lip-sync scaled to legitimate messaging [\[2\]](#page-12-1).

The duality recurs across domains. In privacy protection, GAN-based face replacement can anonymize individuals in datasets and imagery while preserving analytic utility (the DeepPrivacy generator retains about 99.3% of a face detector's original average precision on anonymized images), offering a route to data sharing that is compatible with data-protection requirements [\[96\]](#page-16-12). In education, direct evidence has begun to accumulate: a randomized comparison with 83 adult learners found that micro-lessons delivered by a realistic synthetic AI instructor produced pre-to-post learning gains statistically indistinguishable from those of a traditionally produced instructor video, with no difference in how learners perceived the two formats [\[97\]](#page-16-13). In healthcare communication, synthetic video and voice have been explored for patient education and support in qualitative work with nursing students, alongside an explicit accounting of the associated risks [\[98\]](#page-16-14). Media-forensics scholarship stresses that these constructive applications are real, even as it notes that the benefits literature is less quantified and less integrated into governance debates than the literatures on harm, detection, and regulation [\[28,](#page-13-16)[84\]](#page-16-0).

# **11. Open Challenges**

Several questions remain unsettled. The definitional boundary of "deepfake" is one: technical, legal, and popular usage still diverge on the roles of deep learning, deception, and depiction of real persons [\[3\]](#page-12-2). The balance between deception and corrosion is another. The strongest evidence supports increased uncertainty and reduced trust, while direct evidence of real-world electoral persuasion remains thin [\[73\]](#page-15-11). Behind both stands the open epistemic-apocalypse debate, which pits grave-threat accounts against arguments that coherentist cross-checking survives [\[81](#page-15-19)[–83\]](#page-15-21).

On the technical side, detection faces a persistent ceiling. Benchmark accuracy does not equal real-world reliability, generalization failure and adversarial drift continue, and no generally recognized forensic standard tailored to deepfakes exists [\[30,](#page-13-18)[31\]](#page-13-19). Provenance and watermarking remain strippable and removable (cross-layer desynchronization is an emerging, single-study demonstration), so they function as accountability infrastructure, not as proof that unmarked content is authentic [\[42](#page-14-3)[,44\]](#page-14-5).

Governance and measurement lag behind both. Regulatory approaches are incoherent across jurisdictions, and it would be misleading to imply a single global model [\[9,](#page-12-8)[10\]](#page-12-9). Prevalence data are dated or thin, with contemporary cross-modality measurement only beginning to accumulate [\[16](#page-13-4)[,79\]](#page-15-17), and the ecosystem-level effects of "AI slop" are documented qualitatively but not yet quantified [\[80\]](#page-15-18). Beneficial uses are under-integrated into governance debates [\[84\]](#page-16-0). Human-in-the-loop interventions such as warning labels show mixed evidence, and some labels appear to trigger indiscriminate skepticism that can feed the liar's dividend [\[6,](#page-12-5)[43,](#page-14-4)[49\]](#page-14-8).

# **12. Conclusions**

Synthetic media are AI-generated or AI-manipulated artifacts whose defining feature is a partly or wholly artificial relation between appearance and reality, with the deepfake as their most recognizable depiction-centered subclass, spanning image, audio, and video. The established knowledge base supports several central conclusions: on a widely held scholarly and legal view, malicious intent is typical but not definitional; unaided human detection sits near chance; automated detection is powerful in-distribution but fragile in the wild; provenance and watermarking are useful but strippable, removable, and incomplete; and the dominant societal effect documented to date is the erosion of trust and the growth of uncertainty, not reliable mass persuasion. At the same time, synthetic media are genuinely dual-use, with substantial and legitimate applications in accessibility, localization, privacy protection, education, and healthcare. Governance has advanced unevenly, from China's mandatory labeling and the EU's phased transparency duties to criminal-law responses in South Korea, Australia, and the United Kingdom and a fragmented US patchwork, while high-profile measures in Denmark and the proposed US NO FAKES Act remain unenacted. For the digital society at large, the practical implication is that authenticity is best treated

not as a property to be certified once, but as a socio-technical condition sustained through complementary layers of detection, provenance, disclosure, regulation, and media literacy. That layered strategy is a recurrent recommendation of the technical and policy literatures rather than a documented consensus, and its composite effectiveness remains empirically unmeasured [\[28](#page-13-16)[,41](#page-14-2)[,84\]](#page-16-0). The same holds for human-centric smart cities, where citizens' confidence in official communications presupposes this authenticity infrastructure; for the immersive dimension, see published work on the metaverse [\[7\]](#page-12-6). Journalism-specific work on AI-generated imagery reaches the same practical conclusion: maintaining trust requires rigorous verification, clear labeling, disclosure, and institutional policies for synthetic visual content [\[99\]](#page-16-15). The most durable open questions are less about any single generative system than about the institutions through which audiences will continue to calibrate trust in what they see and hear.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** No new data were created or analyzed in this study. Data sharing is not applicable to this article.

**Acknowledgments:** During the preparation of this manuscript, the author used Anthropic Claude Opus 4.8 and OpenAI GPT-5.5 xhigh for limited assistance with idea development, literature discovery, manuscript organization, and language editing. The author independently evaluated and revised all AI-assisted material and verified every factual claim, legal and scholarly source, citation, and interpretation against the original sources. No AI system was used to make final scholarly judgments, conduct independent analysis, generate or validate data, or assume authorship. The author takes full responsibility for the accuracy, integrity, and originality of the final manuscript.

**Conflicts of Interest:** The author declares no conflicts of interest.

# **References**

- <span id="page-12-0"></span>1. Westerlund, M. The Emergence of Deepfake Technology: A Review. *Technol. Innov. Manag. Rev.* **2019**, *9*, 39–52. [\[CrossRef\]](http://dx.doi.org/10.22215/timreview/1282) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/42589149)
- <span id="page-12-1"></span>2. Kietzmann, J.; Lee, L.W.; McCarthy, I.P.; Kietzmann, T.C. Deepfakes: Trick or Treat? *Bus. Horiz.* **2020**, *63*, 135–146. [\[CrossRef\]](http://dx.doi.org/10.1016/j.bushor.2019.11.006)
- <span id="page-12-2"></span>3. Altuncu, E.; Franqueira, V.N.L.; Li, S. Deepfake: Definitions, Performance Metrics and Standards, Datasets, and a Meta-Review. *Front. Big Data* **2024**, *7*, 1400024. [\[CrossRef\]](http://dx.doi.org/10.3389/fdata.2024.1400024) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/39296632)
- <span id="page-12-3"></span>4. Goodfellow, I.J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative Adversarial Nets. In *Advances in Neural Information Processing Systems 27 (NeurIPS 2014)*; Curran Associates: Red Hook, NY, USA, 2014.
- <span id="page-12-4"></span>5. van der Sloot, B.; Wagensveld, Y. Deepfakes: Regulatory Challenges for the Synthetic Society. *Comput. Law Secur. Rev.* **2022**, *46*, 105716. [\[CrossRef\]](http://dx.doi.org/10.1016/j.clsr.2022.105716)
- <span id="page-12-5"></span>6. Chesney, R.; Citron, D.K. Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security. *Calif. Law Rev.* **2019**, *107*, 1753–1820. [\[CrossRef\]](http://dx.doi.org/10.15779/Z38RV0D15J)
- <span id="page-12-6"></span>7. Mourtzis, D. The Metaverse in Industry 5.0: A Human-Centric Approach towards Personalized Value Creation. *Encyclopedia* **2023**, *3*, 1105–1120. [\[CrossRef\]](http://dx.doi.org/10.3390/encyclopedia3030080)
- <span id="page-12-7"></span>8. Friedman, R. Large Language Models and Logical Reasoning. *Encyclopedia* **2023**, *3*, 687–697. [\[CrossRef\]](http://dx.doi.org/10.3390/encyclopedia3020049)
- <span id="page-12-8"></span>9. Regulation (EU) 2024/1689 of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act), Arts. 3(60) and 50. *Off. J. Eur. Union* **2024**, *L*, 2024/1689. Available online: <https://eur-lex.europa.eu/eli/reg/2024/1689/oj> (accessed on 24 July 2026).
- <span id="page-12-9"></span>10. Cyberspace Administration of China. Measures for Labelling of Artificial Intelligence-Generated Synthetic Content (in force 1 September 2025). 2025. Available online: [https://www.cac.gov.cn/2025-03/14/c\\_1743654684782215.htm](https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm) (accessed on 5 July 2026).
- <span id="page-12-10"></span>11. Paris, B.; Donovan, J. *Deepfakes and Cheap Fakes: The Manipulation of Audio and Visual Evidence*; Data & Society Research Institute: New York, NY, USA, 2019. Available online: [https://datasociety.net/research-library/deepfakes-and-cheap-fakes-the](https://datasociety.net/research-library/deepfakes-and-cheap-fakes-the-manipulation-of-audio-and-visual-evidence/)[manipulation-of-audio-and-visual-evidence/](https://datasociety.net/research-library/deepfakes-and-cheap-fakes-the-manipulation-of-audio-and-visual-evidence/) (accessed on 4 July 2026).

- <span id="page-13-1"></span><span id="page-13-0"></span>13. Masood, M.; Nawaz, M.; Malik, K.M.; Javed, A.; Irtaza, A.; Malik, H. Deepfakes Generation and Detection: State-of-the-Art, Open Challenges, Countermeasures, and Way Forward. *Appl. Intell.* **2023**, *53*, 3974–4026. [\[CrossRef\]](http://dx.doi.org/10.1007/s10489-022-03766-z)
- <span id="page-13-2"></span>14. Mirsky, Y.; Lee, W. The Creation and Detection of Deepfakes: A Survey. *ACM Comput. Surv.* **2021**, *54*, 1–41. [\[CrossRef\]](http://dx.doi.org/10.1145/3425780)
- <span id="page-13-3"></span>15. Cyberspace Administration of China. Provisions on the Administration of Deep Synthesis of Internet Information Services, Arts. 16, 17 and 23 (in force 10 January 2023). 2022. Available online: [https://www.cac.gov.cn/2022-12/11/c\\_1672221949354811.htm](https://www.cac.gov.cn/2022-12/11/c_1672221949354811.htm) (accessed on 5 July 2026).
- <span id="page-13-4"></span>16. Ajder, H.; Patrini, G.; Cavalli, F.; Cullen, L. *The State of Deepfakes: Landscape, Threats, and Impact*; Deeptrace: Amsterdam, The Netherlands, 2019. Available online: [https://regmedia.co.uk/2019/10/08/deepfake\\_report.pdf](https://regmedia.co.uk/2019/10/08/deepfake_report.pdf) (accessed on 4 July 2026).
- <span id="page-13-5"></span>17. Wu, J.; Yang, S.; Zhan, R.; Yuan, Y.; Chao, L.S.; Wong, D.F. A Survey on LLM-Generated Text Detection: Necessity, Methods, and Future Directions. *Comput. Linguist.* **2025**, *51*, 275–338. [\[CrossRef\]](http://dx.doi.org/10.1162/coli_a_00549)
- <span id="page-13-6"></span>18. Gazis, A.; Karypidis, E.; Santamouri, K.; Vavouras, T.; Mastorakis, N.E.; Pappas, S. Deepfakes and Synthetic Media: Generation, Detection, and Governance. *Encyclopedia* **2026**, *6*, 165. [\[CrossRef\]](http://dx.doi.org/10.3390/encyclopedia6080165)
- <span id="page-13-7"></span>19. Thies, J.; Zollhöfer, M.; Stamminger, M.; Theobalt, C.; Nießner, M. Face2Face: Real-Time Face Capture and Reenactment of RGB Videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016. [\[CrossRef\]](http://dx.doi.org/10.1109/CVPR.2016.262)
- <span id="page-13-8"></span>20. Suwajanakorn, S.; Seitz, S.M.; Kemelmacher-Shlizerman, I. Synthesizing Obama: Learning Lip Sync from Audio. *ACM Trans. Graph.* **2017**, *36*, 95. [\[CrossRef\]](http://dx.doi.org/10.1145/3072959.3073640)
- <span id="page-13-9"></span>21. Jia, Y.; Zhang, Y.; Weiss, R.J.; Wang, Q.; Shen, J.; Ren, F.; Chen, Z.; Nguyen, P.; Pang, R.; Lopez-Moreno, I.; et al. Transfer Learning from Speaker Verification to Multispeaker Text-to-Speech Synthesis. *arXiv* **2018**, arXiv:1806.04558.
- <span id="page-13-10"></span>22. Karras, T.; Laine, S.; Aila, T. A Style-Based Generator Architecture for Generative Adversarial Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 15–20 June 2019. [\[CrossRef\]](http://dx.doi.org/10.1109/CVPR.2019.00453)
- <span id="page-13-11"></span>23. Ho, J.; Jain, A.; Abbeel, P. Denoising Diffusion Probabilistic Models. In *Advances in Neural Information Processing Systems 33 (NeurIPS 2020)*; Curran Associates: Red Hook, NY, USA, 2020.
- <span id="page-13-12"></span>24. Ramesh, A.; Pavlov, M.; Goh, G.; Gray, S.; Voss, C.; Radford, A.; Chen, M.; Sutskever, I. Zero-Shot Text-to-Image Generation. *arXiv* **2021**, arXiv:2102.12092.
- <span id="page-13-13"></span>25. Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; Ommer, B. High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, 18–24 June 2022. [\[CrossRef\]](http://dx.doi.org/10.1109/CVPR52688.2022.01042)
- <span id="page-13-14"></span>26. OpenAI. Sora (Text-to-Video Model). 2024. Available online: <https://openai.com/index/sora-is-here/> (accessed on 4 July 2026).
- <span id="page-13-15"></span>27. Google DeepMind. Veo (Text-to-Video Model). 2024. Available online: <https://deepmind.google/technologies/veo/> (accessed on 4 July 2026).
- <span id="page-13-16"></span>28. Verdoliva, L. Media Forensics and DeepFakes: An Overview. *IEEE J. Sel. Top. Signal Process.* **2020**, *14*, 910–932. [\[CrossRef\]](http://dx.doi.org/10.1109/JSTSP.2020.3002101)
- <span id="page-13-17"></span>29. Rössler, A.; Cozzolino, D.; Verdoliva, L.; Riess, C.; Thies, J.; Nießner, M. FaceForensics++: Learning to Detect Manipulated Facial Images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Republic of Korea, 27 October–2 November 2019. [\[CrossRef\]](http://dx.doi.org/10.1109/ICCV.2019.00009)
- <span id="page-13-18"></span>30. Dolhansky, B.; Bitton, J.; Pflaum, B.; Lu, J.; Howes, R.; Wang, M.; Ferrer, C.C. The DeepFake Detection Challenge (DFDC) Dataset. *arXiv* **2020**, arXiv:2006.07397.
- <span id="page-13-19"></span>31. Wang, T.; Liao, X.; Chow, K.P.; Lin, X.; Wang, Y. Deepfake Detection: A Comprehensive Survey from the Reliability Perspective. *ACMComput. Surv.* **2024**, *57*, 1–35. [\[CrossRef\]](http://dx.doi.org/10.1145/3699710)
- <span id="page-13-20"></span>32. Meta AI. Deepfake Detection Challenge Results: An Open Initiative to Advance AI. 12 June 2020. Available online: <https://ai.meta.com/blog/deepfake-detection-challenge-results-an-open-initiative-to-advance-ai/> (accessed on 4 July 2026).
- <span id="page-13-21"></span>33. Maras, M.-H.; Alexandrou, A. Determining Authenticity of Video Evidence in the Age of Artificial Intelligence and in the Wake of Deepfake Videos. *Int. J. Evid. Proof* **2019**, *23*, 255–262. [\[CrossRef\]](http://dx.doi.org/10.1177/1365712718807226)
- <span id="page-13-22"></span>34. Diel, A.; Lalgi, T.; Schröter, I.C.; MacDorman, K.F.; Teufel, M.; Bäuerle, A. Human Performance in Detecting Deepfakes: A Systematic Review and Meta-Analysis of 56 Papers. *Comput. Hum. Behav. Rep.* **2024**, *16*, 100538. [\[CrossRef\]](http://dx.doi.org/10.1016/j.chbr.2024.100538)
- <span id="page-13-23"></span>35. Nightingale, S.J.; Farid, H. AI-Synthesized Faces Are Indistinguishable from Real Faces and More Trustworthy. *Proc. Natl. Acad. Sci. USA* **2022**, *119*, e2120481119. [\[CrossRef\]](http://dx.doi.org/10.1073/pnas.2120481119) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/35165187)
- <span id="page-13-24"></span>36. Groh, M.; Epstein, Z.; Firestone, C.; Picard, R. Deepfake Detection by Human Crowds, Machines, and Machine-Informed Crowds. *Proc. Natl. Acad. Sci. USA* **2022**, *119*, e2110013119. [\[CrossRef\]](http://dx.doi.org/10.1073/pnas.2110013119) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/34969837)
- <span id="page-13-25"></span>37. Coalition for Content Provenance and Authenticity (C2PA). C2PA Technical Specification. 2024. Available online: <https://c2pa.org/specifications/> (accessed on 4 July 2026).
- <span id="page-13-26"></span>38. Content Authenticity Initiative. Content Credentials: Camera-Maker Adoption (Leica M11-P; Nikon Z6III C2PA Firmware Announced 14 October 2024). 2024. Available online: <https://contentcredentials.org/> (accessed on 4 July 2026).

- <span id="page-14-1"></span><span id="page-14-0"></span>40. Dathathri, S.; See, A.; Ghaisas, S.; Huang, P.-S.; McAdam, R.; Welbl, J.; Bachani, V.; Kaskasoli, A.; Stanforth, R.; Matejovicova, T.; et al. Scalable Watermarking for Identifying Large Language Model Outputs. *Nature* **2024**, *634*, 818–823. [\[CrossRef\]](http://dx.doi.org/10.1038/s41586-024-08025-4) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/39443777)
- <span id="page-14-2"></span>41. OpenAI. Advancing Content Provenance: C2PA, SynthID, and Verification Tooling. 2026. Available online: [https://openai.com/](https://openai.com/index/advancing-content-provenance/) [index/advancing-content-provenance/](https://openai.com/index/advancing-content-provenance/) (accessed on 4 July 2026).
- <span id="page-14-3"></span>42. Zhao, X.; Zhang, K.; Su, Z.; Vasan, S.; Grishchenko, I.; Kruegel, C.; Vigna, G.; Wang, Y.-X.; Li, L. Invisible Image Watermarks Are Provably Removable Using Generative AI. In *Advances in Neural Information Processing Systems 37 (NeurIPS 2024)*; Curran Associates: Red Hook, NY, USA, 2024.
- <span id="page-14-4"></span>43. Feng, K.J.K.; Ritchie, N.; Blumenthal, P.; Parsons, A.; Zhang, A.X. Examining the Impact of Provenance-Enabled Media on Trust and Accuracy Perceptions. *Proc. ACM Hum.-Comput. Interact. (CSCW)* **2023**, *7*, 270. [\[CrossRef\]](http://dx.doi.org/10.1145/3610061)
- <span id="page-14-5"></span>44. Nemecek, A.; He, H.; Cheng, G.; Ayday, E. Authenticated Contradictions from Desynchronized Provenance and Watermarking. *arXiv* **2026**, arXiv:2603.02378.
- <span id="page-14-6"></span>45. Meta. Our Approach to Labeling AI-Generated Content and Manipulated Media. 5 April 2024. Available online: [https://about.](https://about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/) [fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/](https://about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/) (accessed on 5 July 2026).
- 46. YouTube Help. Disclosing Use of GenAI Content. 2024–2026. Available online: [https://support.google.com/youtube/answer/](https://support.google.com/youtube/answer/14328491) [14328491](https://support.google.com/youtube/answer/14328491) (accessed on 5 July 2026).
- 47. TikTok. Community Guidelines: Integrity and Authenticity—Edited Media and AI-Generated Content (AIGC). Available online: <https://www.tiktok.com/community-guidelines/en/integrity-authenticity> (accessed on 5 July 2026).
- <span id="page-14-7"></span>48. TikTok Newsroom. Partnering with Our Industry to Advance AI Transparency and Literacy. 9 May 2024. Available online: [https://](https://newsroom.tiktok.com/en-us/partnering-with-our-industry-to-advance-ai-transparency-and-literacy) [newsroom.tiktok.com/en-us/partnering-with-our-industry-to-advance-ai-transparency-and-literacy](https://newsroom.tiktok.com/en-us/partnering-with-our-industry-to-advance-ai-transparency-and-literacy) (accessed on 5 July 2026).
- <span id="page-14-8"></span>49. Ternovski, J.; Kalla, J.; Aronow, P.M. The Negative Consequences of Informing Voters about Deepfakes: Evidence from Two Survey Experiments. *J. Online Trust Saf.* **2022**, *1*, 1–16. [\[CrossRef\]](http://dx.doi.org/10.54501/jots.v1i2.28)
- <span id="page-14-9"></span>50. Guess, A.M.; Lerner, M.; Lyons, B.; Montgomery, J.M.; Nyhan, B.; Reifler, J.; Sircar, N. A Digital Media Literacy Intervention Increases Discernment between Mainstream and False News in the United States and India. *Proc. Natl. Acad. Sci. USA* **2020**, *117*, 15536–15545. [\[CrossRef\]](http://dx.doi.org/10.1073/pnas.1920498117) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/32571950)
- <span id="page-14-10"></span>51. Regulation (EU) 2022/2065 of the European Parliament and of the Council on a Single Market for Digital Services (Digital Services Act), Arts. 34–35. *Off. J. Eur. Union* **2022**, *L 277*, 1–102. Available online: <https://eur-lex.europa.eu/eli/reg/2022/2065/oj> (accessed on 4 July 2026).
- <span id="page-14-11"></span>52. European Commission. Code of Practice on Transparency of AI-Generated Content. 10 June 2026. Available online: <https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content> (accessed on 24 July 2026).
- <span id="page-14-12"></span>53. European Commission. Commission Opinion on the Assessment of the Code of Practice on Transparency of AI-Generated Content. 8 July 2026. Available online: [https://digital-strategy.ec.europa.eu/en/library/commission-opinion-assessment-code](https://digital-strategy.ec.europa.eu/en/library/commission-opinion-assessment-code-practice-transparency-ai-generated-content)[practice-transparency-ai-generated-content](https://digital-strategy.ec.europa.eu/en/library/commission-opinion-assessment-code-practice-transparency-ai-generated-content) (accessed on 24 July 2026).
- <span id="page-14-13"></span>54. Cyberspace Administration of China. Interim Measures for the Management of Generative Artificial Intelligence Services (in Force 15 August 2023). 2023. Available online: [https://www.cac.gov.cn/2023-07/13/c\\_1690898327029107.htm](https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm) (accessed on 5 July 2026).
- <span id="page-14-14"></span>55. *GB 45438-2025*; Cybersecurity Technology—Labelling Method for Content Generated by Artificial Intelligence. State Administration for Market Regulation: Beijing, China, 2025. Available online: [https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F3](https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F32EA2A561F1886CD8D606513512D547) [2EA2A561F1886CD8D606513512D547](https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F32EA2A561F1886CD8D606513512D547) (accessed on 5 July 2026).
- <span id="page-14-15"></span>56. TAKE IT DOWN Act, Public Law 119-12, United States (Enacted 19 May 2025; Covered-Platform Obligations from 19 May 2026). 2025. Available online: <https://www.congress.gov/119/plaws/publ12/PLAW-119publ12.pdf> (accessed on 5 July 2026).
- <span id="page-14-16"></span>57. Federal Trade Commission. Trade Regulation Rule on Impersonation of Government and Businesses, 16 CFR Part 461 (Effective 1 April 2024). *Fed. Regist.* **2024**, *89*, 15017–15031. Available online: [https://www.federalregister.gov/documents/2024/03/01/2](https://www.federalregister.gov/documents/2024/03/01/2024-04335/trade-regulation-rule-on-impersonation-of-government-and-businesses) [024-04335/trade-regulation-rule-on-impersonation-of-government-and-businesses](https://www.federalregister.gov/documents/2024/03/01/2024-04335/trade-regulation-rule-on-impersonation-of-government-and-businesses) (accessed on 5 July 2026).
- <span id="page-14-17"></span>58. Federal Trade Commission. Supplemental Notice of Proposed Rulemaking on Impersonation of Individuals (Pending, Not Finalized as of July 2026). *Fed. Regist.* **2024**, *89*, 15072–15083. Available online: [https://www.federalregister.gov/documents/20](https://www.federalregister.gov/documents/2024/03/01/2024-03793/trade-regulation-rule-on-impersonation-of-government-and-businesses) [24/03/01/2024-03793/trade-regulation-rule-on-impersonation-of-government-and-businesses](https://www.federalregister.gov/documents/2024/03/01/2024-03793/trade-regulation-rule-on-impersonation-of-government-and-businesses) (accessed on 5 July 2026).
- <span id="page-14-18"></span>59. Kohls v. Bonta, No. 2:24-cv-02527-JAM-CKD (E.D. Cal.): Order Granting Preliminary Injunction Against AB 2839, 2 October 2024; Final Judgment Holding AB 2655 Preempted by 47 U.S.C. § 230, 20 August 2025; Order Permanently Enjoining AB 2839, 29 August 2025; Appeal Docketed, No. 25-6138 (9th Cir., 19 September 2025). Available online: [https://www.courtlistener.com/](https://www.courtlistener.com/docket/69172215/kohls-v-bonta/) [docket/69172215/kohls-v-bonta/](https://www.courtlistener.com/docket/69172215/kohls-v-bonta/) (accessed on 5 July 2026).
- <span id="page-14-19"></span>60. Public Citizen; National Conference of State Legislatures (NCSL); Ballotpedia. Trackers of US State Deepfake and AI-in-Elections Legislation. 2026. Available online: <https://www.citizen.org/article/tracker-legislation-on-deepfakes-in-elections/> (accessed on 4 July 2026).

- <span id="page-15-0"></span>61. Act on Special Cases Concerning the Punishment of Sexual Crimes (Republic of Korea), as Amended by Act No. 20459 of 16 October 2024. Korean Law Information Center. Available online: <https://www.law.go.kr/lsInfoP.do?lsId=011187> (accessed on 5 July 2026).
- <span id="page-15-1"></span>62. Criminal Code Amendment (Deepfake Sexual Material) Act 2024, No. 78 of 2024 (Commonwealth of Australia; Assent 2 September 2024). Federal Register of Legislation. Available online: <https://www.legislation.gov.au/C2024A00078/latest/text> (accessed on 5 July 2026).
- <span id="page-15-2"></span>63. Online Safety Act 2023, c. 50, Sections 187–188 (United Kingdom; inserting Sections 66A–66D into the Sexual Offences Act 2003; Section 188 in force 31 January 2024). Available online: <https://www.legislation.gov.uk/ukpga/2023/50/section/188> (accessed on 5 July 2026).
- <span id="page-15-3"></span>64. Data (Use and Access) Act 2025, c. 18, Section 138 (United Kingdom; Inserting Sections 66E–66H into the Sexual Offences Act 2003; in Force 6 February 2026). Available online: <https://www.legislation.gov.uk/ukpga/2025/18/section/138> (accessed on 5 July 2026).
- <span id="page-15-4"></span>65. Kulturministeriet (Danish Ministry of Culture). Bred Aftale om Deepfakes Giver Alle Ret til Egen Krop og Egen Stemme (Political Agreement, 26 June 2025). 2025. Available online: [https://kum.dk/aktuelt/nyheder/bred-aftale-om-deepfakes-giver-alle-ret-til](https://kum.dk/aktuelt/nyheder/bred-aftale-om-deepfakes-giver-alle-ret-til-egen-krop-og-egen-stemme)[egen-krop-og-egen-stemme](https://kum.dk/aktuelt/nyheder/bred-aftale-om-deepfakes-giver-alle-ret-til-egen-krop-og-egen-stemme) (accessed on 5 July 2026).
- <span id="page-15-5"></span>66. Høring over Forslag til Lov om Ændring af Lov om Ophavsret (Sections 65 a and 73 a), Høringssag 70269 (Consultation Closed 21 August 2025; Pre-Parliamentary Draft, Not Introduced). 2025. Available online: [https://hoeringsportalen.dk/Hearing/](https://hoeringsportalen.dk/Hearing/Details/70269) [Details/70269](https://hoeringsportalen.dk/Hearing/Details/70269) (accessed on 5 July 2026).
- <span id="page-15-6"></span>67. NO FAKES Act of 2025, S. 1367, 119th Congress, United States—Formal Proposal, Not Enacted. 2025. Available online: <https://www.congress.gov/bill/119th-congress/senate-bill/1367> (accessed on 5 July 2026).
- <span id="page-15-7"></span>68. NO FAKES Act of 2026, S. 4591, 119th Congress, United States—Formal Proposal, Not Enacted (Ordered Reported by the Senate Judiciary Committee on 18 June 2026). 2026. Available online: <https://www.congress.gov/bill/119th-congress/senate-bill/4591> (accessed on 5 July 2026).
- <span id="page-15-8"></span>69. Ministry of Electronics and Information Technology (MeitY), Government of India. MeitY Issues Advisory to All Intermediaries to Comply with Existing IT Rules (Deepfake Advisory). Press Information Bureau, 26 December 2023. Available online: <https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=1990542> (accessed on 4 July 2026).
- <span id="page-15-9"></span>70. Assembly Bill 730, 2019–2020 Reg. Sess., Chapter 493, Statutes of 2019 (California, US); Sunset Extended to 1 January 2027 by Assembly Bill 972, Chapter 745, Statutes of 2022. Available online: [https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?](https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=201920200AB730) [bill\\_id=201920200AB730](https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=201920200AB730) (accessed on 5 July 2026).
- 71. Assembly Bill 602, 2019–2020 Reg. Sess., Chapter 491, Statutes of 2019 (California, US). Available online: [https://leginfo.](https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=201920200AB602) [legislature.ca.gov/faces/billTextClient.xhtml?bill\\_id=201920200AB602](https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=201920200AB602) (accessed on 5 July 2026).
- <span id="page-15-10"></span>72. Senate Bill 751, 86th Legislature, Reg. Sess. (Texas, US, 2019); Codified at Tex. Elec. Code § 255.004, Effective 1 September 2019. Available online: <https://capitol.texas.gov/BillLookup/History.aspx?LegSess=86R&Bill=SB751> (accessed on 5 July 2026).
- <span id="page-15-11"></span>73. Vaccari, C.; Chadwick, A. Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty, and Trust in News. *Soc. Media Soc.* **2020**, *6*, 2056305120903408. [\[CrossRef\]](http://dx.doi.org/10.1177/2056305120903408)
- <span id="page-15-12"></span>74. Schiff, K.J.; Schiff, D.S.; Bueno, N.S. The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability? *Am. Polit. Sci. Rev.* **2025**, *119*, 71–90. [\[CrossRef\]](http://dx.doi.org/10.1017/S0003055423001454)
- <span id="page-15-13"></span>75. Chen, H.; Magramo, K. Finance Worker Pays Out \$25 Million After Video Call with Deepfake "Chief Financial Officer". CNN, 4 February 2024. Available online: <https://www.cnn.com/2024/02/04/asia/deepfake-cfo-scam-hong-kong-intl-hnk> (accessed on 4 July 2026).
- <span id="page-15-14"></span>76. CNN Business. Arup Revealed as Victim of \$25 Million Deepfake Scam Involving Hong Kong Employee. 16 May 2024. Available online: <https://www.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk> (accessed on 4 July 2026).
- <span id="page-15-15"></span>77. Federal Communications Commission. FCC Issues \$6M Fine for NH Robocalls (Forfeiture Order). September 2024. Available online: <https://www.fcc.gov/document/fcc-issues-6m-fine-nh-robocalls> (accessed on 4 July 2026).
- <span id="page-15-16"></span>78. New Hampshire Public Radio. Political Operative Who Admitted to Creating Fake Biden Robocalls Found Not Guilty. 13 June 2025. Available online: [https://www.nhpr.org/nh-news/2025-06-13/political-operative-fake-biden-robocalls-nh](https://www.nhpr.org/nh-news/2025-06-13/political-operative-fake-biden-robocalls-nh-primary-found-not-guilty)[primary-found-not-guilty](https://www.nhpr.org/nh-news/2025-06-13/political-operative-fake-biden-robocalls-nh-primary-found-not-guilty) (accessed on 5 July 2026).
- <span id="page-15-17"></span>79. Umbach, R.; Henry, N.; Beard, G.; Berryessa, C. Non-Consensual Synthetic Intimate Imagery: Prevalence, Attitudes, and Knowledge in 10 Countries. In Proceedings of the CHI Conference on Human Factors in Computing Systems, Honolulu, HI, USA, 11–16 May 2024. [\[CrossRef\]](http://dx.doi.org/10.1145/3613904.3642382)
- <span id="page-15-18"></span>80. Merriam-Webster. 2025 Word of the Year: "Slop". 2025. Available online: [https://www.merriam-webster.com/wordplay/word](https://www.merriam-webster.com/wordplay/word-of-the-year)[of-the-year](https://www.merriam-webster.com/wordplay/word-of-the-year) (accessed on 4 July 2026).
- <span id="page-15-19"></span>81. Rini, R. Deepfakes and the Epistemic Backstop. *Philos. Impr.* **2020**, *20*, 1–16.
- <span id="page-15-20"></span>82. Fallis, D. The Epistemic Threat of Deepfakes. *Philos. Technol.* **2021**, *34*, 623–643. [\[CrossRef\]](http://dx.doi.org/10.1007/s13347-020-00419-2) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/32837868)
- <span id="page-15-21"></span>83. Habgood-Coote, J. Deepfakes and the Epistemic Apocalypse. *Synthese* **2023**, *201*, 103. [\[CrossRef\]](http://dx.doi.org/10.1007/s11229-023-04097-3)

- <span id="page-16-1"></span><span id="page-16-0"></span>85. Bazin, A. The Ontology of the Photographic Image. *Film Q.* **1960**, *13*, 4–9. Gray, H. (Translator). [\[CrossRef\]](http://dx.doi.org/10.2307/1210183) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/17213806)
- <span id="page-16-2"></span>86. Barthes, R. *Camera Lucida: Reflections on Photography*; Howard, R., Translator; Hill and Wang: New York, NY, USA, 1981.
- <span id="page-16-3"></span>87. Peirce, C.S. *Collected Papers of Charles Sanders Peirce, Volume II: Elements of Logic*; Hartshorne, C., Weiss, P., Eds.; Harvard University Press: Cambridge, MA, USA, 1932.
- <span id="page-16-4"></span>88. Walton, K.L. Transparent Pictures: On the Nature of Photographic Realism. In *Photography and Philosophy: Essays on the Pencil of Nature*; Walden, S., Ed.; Blackwell: Oxford, UK, 2007; pp. 14–49. [\[CrossRef\]](http://dx.doi.org/10.1002/9780470696651.ch1)
- <span id="page-16-5"></span>89. Bolter, J.D.; Grusin, R. *Remediation: Understanding New Media*; MIT Press: Cambridge, MA, USA, 1999.
- <span id="page-16-6"></span>90. Manovich, L. *The Language of New Media*; MIT Press: Cambridge, MA, USA, 2001.
- <span id="page-16-7"></span>91. Gunning, T. Moving Away from the Index: Cinema and the Impression of Reality. *Differences* **2007**, *18*, 29–52. [\[CrossRef\]](http://dx.doi.org/10.1215/10407391-2006-022)
- <span id="page-16-8"></span>92. Mori, M.; MacDorman, K.F.; Kageki, N. The Uncanny Valley [From the Field]. *IEEE Robot. Autom. Mag.* **2012**, *19*, 98–100. [\[CrossRef\]](http://dx.doi.org/10.1109/MRA.2012.2192811)
- <span id="page-16-9"></span>93. Weisman, W.D.; Peña, J.F. Face the Uncanny: The Effects of Doppelganger Talking Head Avatars on Affect-Based Trust Toward Artificial Intelligence Technology Are Mediated by Uncanny Valley Perceptions. *Cyberpsychol. Behav. Soc. Netw.* **2021**, *24*, 182–187. [\[CrossRef\]](http://dx.doi.org/10.1089/cyber.2020.0175) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/33646048)
- <span id="page-16-10"></span>94. Enli, G. *Mediated Authenticity: How the Media Constructs Reality*; Peter Lang: New York, NY, USA, 2015. [\[CrossRef\]](http://dx.doi.org/10.3726/978-1-4539-1458-8)
- <span id="page-16-11"></span>95. Mills, T.; Bunnell, H.T.; Patel, R. Towards Personalized Speech Synthesis for Augmentative and Alternative Communication. *Augment. Altern. Commun.* **2014**, *30*, 226–236. [\[CrossRef\]](http://dx.doi.org/10.3109/07434618.2014.924026) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/25025818)
- <span id="page-16-12"></span>96. Hukkelås, H.; Mester, R.; Lindseth, F. DeepPrivacy: A Generative Adversarial Network for Face Anonymization. In *Advances in Visual Computing (ISVC 2019)*; Springer: Cham, Switzerland, 2019. [\[CrossRef\]](http://dx.doi.org/10.1007/978-3-030-33720-9_44)
- <span id="page-16-13"></span>97. Leiker, D.; Gyllen, A.R.; Eldesouky, I.; Cukurova, M. Generative AI for Learning: Investigating the Potential of Learning Videos with Synthetic Virtual Instructors. In *Artificial Intelligence in Education (AIED 2023)*; Communications in Computer and Information Science, Volume 1831; Springer: Cham, Switzerland, 2023; pp. 523–529. [\[CrossRef\]](http://dx.doi.org/10.1007/978-3-031-36336-8_81)
- <span id="page-16-14"></span>98. Navarro Martínez, O.; Fernández-García, D.; Cuartero Monteagudo, N.; Forero-Rincón, O. Possible Health Benefits and Risks of DeepFake Videos: A Qualitative Study in Nursing Students. *Nurs. Rep.* **2024**, *14*, 2746–2757. [\[CrossRef\]](http://dx.doi.org/10.3390/nursrep14040203) [\[PubMed\]](http://www.ncbi.nlm.nih.gov/pubmed/39449440)
- <span id="page-16-15"></span>99. Voinea, D.V. The Ethics of AI-Generated Imagery in Journalism. *Soc. Sci. Educ. Res. Rev.* **2023**, *10*, 330–335. [\[CrossRef\]](http://dx.doi.org/10.5281/zenodo.15254301)

**Disclaimer/Publisher's Note:** The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.