This study adopted a systematic review approach to synthesize current research on deepfakes, generative Artificial Intelligence (AI), and misinformation. The review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to ensure methodological transparency, replicability, and rigor. The process involved six key stages: (i) defining the research scope and questions, (ii) developing search strategies and selecting databases, (iii) applying inclusion and exclusion criteria, (iv) screening and extracting relevant data, (v) assessing methodological quality, and (vi) synthesizing and interpreting the findings.

The scope of this review covered the period 2014 to 2025, reflecting the evolution of deepfake research from its early technical demonstrations to contemporary multimodal and regulatory developments. Both qualitative and quantitative evidence were considered to capture the interdisciplinary nature of deepfake research across technical, social, and policy domains.

A comprehensive multi-database search strategy was designed to capture the full breadth of deepfake-related research, including studies on generation, detection, misinformation, and regulation. The Boolean search string used (Scopus format) was:

(“deepfake*” OR “fake news” OR “synthetic media” OR “AI-generated content” OR “misinformation” OR “disinformation”) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural network*” OR “natural language processing” OR “computer vision”)

From an initial pool of 120 database records, complemented by backward and forward citation chaining, this query was executed and adapted for each database, including IEEE Xplore, ACM Digital Library, Scopus, SpringerLink, and Web of Science. To ensure inclusion of the most recent advances, arXiv preprints were also screened, recognizing that peer review often lags rapid developments in generative AI.

Preprints were included only when they were later cited in peer-reviewed venues or associated with publicly recognized datasets to mitigate quality concerns.

For regulatory and governance perspectives, official EU repositories and government sites were reviewed to obtain key policy texts such as the EU Digital Services Act (European Parliament and Council of the European Union (EU DSA), 2022) and the EU AI Act (European Parliament and Council of the European Union (EU AI Act), 2024). Reference lists of selected articles were also examined through backward and forward citation chasing to identify additional relevant studies not captured by the initial search.

arXiv preprints were included selectively due to the rapid pace of advances in generative AI and were cross-validated based on subsequent peer-reviewed adoption, dataset impact, or citation prominence.

Inclusion criteria:

Studies were included if they met the following criteria:

Exclusion criteria:

These criteria ensured that the final selection consisted of studies with direct relevance, methodological rigor, and conceptual clarity.

The search strategy combined structured database keyword searches with both backward citations chasing (examining the references of included studies) and forward citation chasing (tracking newer works that cited them). Initial queries employed broad terms such as deepfake detection, generative adversarial networks (GANs), synthetic media, misinformation, and fake news. Throughout the review, the strategy was refined repeatedly to incorporate new methods and ideas. Other keywords comprised transformers, Vision Transformer (ViT), CLIP, voice generation, multimodal recognition, Xplainable AI (XAI), retrieval-augmented production (RAG), bias, equity, governance, and regulation. Boolean operators (e.g., “deepfake AND identification”, “synthetic media AND false information”, “transformer OR CLIP AND deepfake”) were employed to enhance recall and precision. This iterative approach ensured that both foundational studies and the most recent developments in technical, social, and regulatory domains were systematically captured.

The screening process followed a three-stage procedure to ensure systematic inclusion of high-quality and relevant studies:

At the full-text screening stage, studies were excluded for the following predefined reasons:

Reason 1: Primary focus on AI applications unrelated to misinformation or deceptive synthetic media (n = 6).

Reason 2: Lack of empirical evaluation, methodological transparency, or reproducible analysis (n = 5).

Reason 3: Conceptual or opinion-based papers without sufficient analytical depth or evidence synthesis (n = 7).

The systematic review included a total of 34 studies published between 2014 and 2025, capturing both the historical development and the latest advances in deepfake research. About 18 of them examined deepfake production, detection methods, and model architectures with an emphasis on computer vision and technical approaches. About eight studies from the social sciences looked at how deepfakes affect user behaviour, media trust, and disinformation in society. The other 8 studies focused on policy and regulatory frameworks, emphasizing ethical issues, governance strategies, and legal actions. Research was chosen for its capability to offer a broad viewpoint across various media formats text, images, video, and audio and for its role in enhancing knowledge of the technical, social, and policy aspects of deepfakes. This selection ensures a holistic view of the field, combining insights into both innovation and the societal challenges posed by synthetic media.

Although the ultimate set of 34 studies may seem narrow compared to the vastness of generative AI research, this demonstrates the careful use of stringent inclusion standards that prioritize interdisciplinary significance, methodological soundness, and clear involvement with detection, social consequences, or governance. This trade-off prioritizes analytical depth and coherence over exhaustive coverage and is acknowledged as a limitation of the review.

The methodological quality and potential risk of bias for each study were evaluated using an adapted version of the Critical Appraisal Skills Program (CASP) checklist. Technical studies were assessed based on criteria such as dataset transparency, model reproducibility, and evaluation robustness, while social science and policy studies were reviewed for methodological clarity, validity of interpretation, and evidence linkage. Each study received a quality rating (high, moderate, or low) based on these criteria, which informed the weight given during synthesis of the 34 studies evaluated, 21 were rated high quality, 9 moderate, and 4 low according to the adapted CASP checklist.

The studies included in this review were selected for several key reasons. They first offer a historical basis for comprehending the evolution of generative AI and the rise of misinformation by illustrating the progression of deepfakes from early technical demos to highly advanced synthetic media. Second, they showcase state-of-the-art technological detection techniques, such as multimodal analysis, explainable AI, and machine learning model advancements. Third, the research investigates sociological, ethical, and legal aspects, analysing how deepfakes influence media credibility, public opinion, privacy, and regulatory systems. Finally, the selection deliberately encompasses a range of media modalities text, images, video, and audio to ensure a comprehensive understanding of both the technical challenges and societal implications of deepfakes (see Figure 1).

A thematic review approach was used to organize and interpret the selected studies according to recurring patterns, concepts, and research priorities. Finding commonly discussed subjects, assembling related concepts into clusters, and honing these clusters into cohesive themes were the steps in the analysis process. This approach facilitated the emergence of both chronological and conceptual links, underscoring the progression of research on deepfakes and synthetic media from initial technical trials to wider societal and policy-focused conversations.

From this analysis, four overarching themes were identified:

The trajectory of deepfake development mirrors broader advances in generative AI. Early methods, notably Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) established the adversarial paradigm that underpins synthetic image and video generation. Subsequent variants such as StyleGAN2/3 (Lehtinen & Aila NVIDIA) dramatically improved realism and controllability. Meanwhile, transformer architectures (Vaswani et al., 2017) and large-scale language models (Brown et al., 2020) expanded synthetic generation to text and audio, creating a multimodal threat landscape.

These architectural advances not only enabled high-fidelity media synthesis but also reshaped detection research. As deepfakes grew more realistic, early CNN-based detectors emphasizing pixel-level inconsistencies gave way to transformer and CLIP-based frameworks capable of modelling contextual and semantic relationships. This shift reflects an evolution from surface-level artifact detection toward cross-modal, meaning-aware analysis, mirroring the progression of misinformation itself from isolated falsifications to integrated multimodal narratives.

This architectural transition reflects more than incremental performance improvement. Detectors based on CNNs chiefly leverage low-level visual artifacts, which modern generative models increasingly reduce. Transformer and CLIP-based approaches instead model long-range spatial, temporal, and semantic dependencies, enabling improved robustness to compression and post-processing. Nonetheless, these benefits come with trade-offs in computational expense, data reliance, and interpretability. The literature rarely addresses how such models can be deployed in real-time or resource-constrained environments, revealing a gap between benchmark success and operational feasibility (see Table 1).

To synthesize and contrast the findings across the selected studies, a comparative analysis was performed focusing on methodological design, evaluation performance, and disciplinary orientation. To find similarities, contrasts, and latest trends, this analysis combines data from computer vision, multimodal AI, social science, and policy research.

A comparative analysis identified three primary methodological clusters CNN-based, Transformer-based, and CLIP-based/multimodal frameworks each presenting unique advantages and disadvantages. CNN architectures (e.g., Section 3: Thematic Review/Detection Models, EfficientNet) exhibit high accuracy on established benchmarks such as FaceForensics++ and DFDC but demonstrate limited generalization to unseen datasets or manipulation techniques. Although they require more data and have greater computing costs, Vision Transformers (ViT, Swin-T) and hybrid models provide better contextual modeling and modest improvements in cross-dataset generalization. CLIP-based and multimodal methods utilize vision-language pretraining to enable zero-shot and few-shot detection, indicating scalability potential across modalities, yet they continue to be vulnerable to adversarial perturbations and domain shifts.

Recent research has delved deeper into multimodal fusion and extensive pretraining techniques to enhance generalization and resilience in deepfake detection, especially in cross-dataset and real-world scenarios (Chen et al., 2024a; Chen et al., 2024b).

Foundational deep learning architectures underpinning contemporary detection systems include Xception networks (Chollet, 2017), transformer-based language models such as BERT (Devlin et al., 2019), and generative adversarial networks including StyleGAN (Karras et al., 2021). These architectures have directly impacted the generation pipelines for deepfakes and the detection methods assessed in benchmark datasets like FaceForensics++ (Dolhansky et al., 2019).

From a disciplinary standpoint, methodological research use diverges:

The framework integrates technical, social, and governance dimensions through feedback loops that connect model transparency, user trust, and regulatory compliance to build accountable and resilient AI detection systems (see Figure 2).

The framework is organized in a three-tier system:

These components are interconnected through bidirectional feedback loops. Technical detection systems generate transparent outputs that inform social understanding and policy decisions, while governance structures provide ethical and compliance feedback to guide model development and user interaction.

This framework extends prior conceptual models (Mirsky and Lee, 2021; Zhou et al., 2019) by explicitly incorporating explainability and feedback dynamics that connect machine learning performance with human trust and policy accountability. Unlike earlier models focusing solely on detection or social impacts, this integrated design emphasizes mutual reinforcement between technology, transparency, and governance to create a resilient AI ecosystem for mitigating deepfake threats.

Despite significant advances in deepfake detection, several key gaps and contentious issues remain. Generalization is a major challenge, as many detectors perform well on benchmark datasets but struggle to maintain accuracy across new or unseen datasets, limiting their real-world applicability. Multimodal deepfakes, which combine video, audio, and text, are understudied compared to single-modality cases, leaving detection methods less prepared for increasingly sophisticated manipulations. The balance between explainability and security also presents a dilemma: while explainable AI tools improve transparency, they can potentially be exploited by adversaries to bypass detection. Additionally, ethical and legal protections remain insufficient, and harmful such as non-consensual deepfake pornography and political disinformation highlight gaps in current laws and enforcement. Finally, the lack of standardized, adversarial resilient evaluation benchmarks complicates the comparison of detection methods across studies, making it difficult to assess progress and deploy robust solutions in practice.

Future detection models should reflect the complexity of multimodal manipulations and evolving threat landscapes: