Plain answers, with citations.

Any image, audio, or video artifact produced or materially altered by a generative model — including face-swap, voice-clone, full-frame synthesis, and parameter-level edits indistinguishable from authentic capture. The term is broader than “deepfake” and includes both consensual uses (film VFX, accessibility dubbing, research) and adversarial ones (impersonation, fraud, image-based abuse). See the glossary.

Synthetic media is the umbrella term. Deepfake is the colloquial label for one subset — typically face-swap or voice-clone content created with deep neural networks. Synthetic media also includes full-frame text-to-image generation, voice cloning, lip-sync edits, and audio manipulation that is not face-swap.

It depends on how they are made and used. Generation alone is generally not illegal. Distribution can be civil or criminal under the TAKE IT DOWN Act (federal law since 19 May 2025), the FCC TCPA AI-voice rule, federal wire-fraud statutes, or roughly 30 state-level deepfake-NCII statutes plus broader state NCII / right-of-publicity laws. The DEFIANCE Act (S.1837, 119th Congress) has passed the Senate twice but not the House — it is not yet federal law.

U.S. enterprise scale: the FBI's 2024 IC3 Annual Report recorded $16.6 billion in total cybercrime losses (a 33% YoY increase), with $2.77 billion attributable to BEC alone across 21,442 incidents. Personal scale: the Cyber Civil Rights Initiative's 2017 nationwide study found approximately 1 in 8 U.S. social-media users had had a sexually explicit image shared without consent or had been threatened with such, with women approximately twice as likely as men to be targets.

A Generative Adversarial Network is a pair of neural networks — a generator that produces candidate samples and a discriminator that judges them — trained against each other in a minimax game. Goodfellow et al. introduced the framework in 2014 (NeurIPS, arXiv:1406.2661). See /research-lab.

A diffusion model that operates in the compressed latent space of a pretrained autoencoder rather than in pixel space. Introduced by Rombach et al. at CVPR 2022 (arXiv:2112.10752); the architecture behind Stable Diffusion and most modern open-weight image generators.

Detectors trained on outputs from one generator overfit to that generator's specific spectral and structural artifacts. Applied to images from an unseen architecture, accuracy can drop sharply. This cross-generator generalization gap is the central open problem in synthetic-media forensics. See Wang et al. (arXiv:1912.11035) and Frank et al. (arXiv:2003.08685).

Groh et al. (PNAS 2022, n=15,016) reported, on a curated holdout set: leading automated detector ~65% accuracy; recruited crowds (multiple observers averaged) ~74%; non-recruited crowds ~80%; individual untrained observers tracked the model. So crowds beat individuals; trained reviewers do better still; but a single untrained viewer should not assume better-than-model performance without forensic tooling.

A deepfake detection method by Ciftci, Demir, and Yin (IEEE TPAMI 2020) that exploits photoplethysmographic (PPG) signals — subtle skin-tone shifts from blood circulation — present in real portrait videos but not preserved in synthetic content. Reported accuracies in the paper: 96% on the original FaceForensics dataset (2018), 94.65% on FaceForensics++, 91.50% on Celeb-DF. Degrades on compressed video. (Intel later commercialised a real-time variant under the same name.)

The most commonly cited are FaceForensics++ (Rössler et al., ICCV 2019; arXiv:1901.08971); the Deepfake Detection Challenge dataset (Dolhansky et al., 2020; arXiv:2006.07397); and ASVspoof for audio (Yamagishi et al., 2021). See /research-lab.

An out-of-band callback to a number stored in your corporate directory — never the number that initiated the request. Combine with a 30-minute soft hold on payments above a defined threshold and a pre-shared challenge phrase rotated quarterly. The IC3 Recovery Asset Team reports a 66% success rate freezing funds when reported quickly. See the full six-step protocol on /war-room.

In late January / early February 2024, an Arup Hong Kong finance employee made fifteen wire transfers totaling roughly USD 25 million after a video conference in which every other participant — including the supposed CFO — was an AI-generated deepfake. Arup confirmed publicly in May 2024 (CNN, 16 May 2024).

A Ferrari executive received WhatsApp messages and a phone call using an AI-generated voice clone of CEO Benedetto Vigna. The executive asked the caller to name the title of a book Vigna had personally recommended a few days earlier. The synthetic system did not have the answer (Bloomberg, 26 July 2024). The technique generalizes — see “challenge phrase” in /war-room.

In the United States, yes. The FCC's Declaratory Ruling FCC 24-17 (8 February 2024) confirmed AI-generated voice calls fall within the TCPA's restrictions on “artificial or prerecorded voice” calls and require prior express consent. Effective immediately on adoption.

File with the FBI Internet Crime Complaint Center at ic3.gov within 24 hours, preserve audio and message logs, and notify your bank's fraud team to attempt a recall. The IC3 Recovery Asset Team can attempt to freeze fraudulent transfers if alerted promptly.

(1) Preserve evidence (archive URLs, screenshot with system clock, save originals). (2) Hash and submit to StopNCII.org (adults) or NCMEC Take It Down (imagery from when you were under 18). (3) Report on the platform under both NCII and copyright (Qiwei et al. 2024 found copyright reports cleared faster than NCII reports on a major platform). (4) For sexual or minor-involving imagery, file with FBI tips.fbi.gov. (5) Consult a lawyer about civil claims under existing state NCII / right-of-publicity / IIED law (and under the DEFIANCE Act if and when it becomes federal law). Walk-through on /safety-suite.

The DEFIANCE Act (Disrupt Explicit Forged Images and Non-Consensual Edits Act) is a pending federal bill, not enacted law. It would create a federal civil cause of action for adults depicted in non-consensual intimate digital forgeries — liquidated damages of $150,000 ($250,000 with aggravating factors), 10-year statute of limitations. S.3696 (118th Cong.) passed the Senate by unanimous consent on 23 July 2024 but did not pass the House. Reintroduced as S.1837 (119th Cong.) and passed the Senate again by unanimous consent in January 2026 — still pending in the House.

S.146 (119th Congress), signed 19 May 2025. Criminalises knowingly publishing non-consensual intimate visual depictions of minors or non-consenting adults — including deepfakes — and requires covered platforms to remove such content within 48 hours of a valid report.

Probably yes for some scenario, but coverage is uneven. By mid-2025, roughly 30 states had specific deepfake-NCII statutes, nearly all states had broader NCII laws that may apply to AI content, and approximately 45 states had criminalised AI-generated CSAM. See the NCSL tracker or Public Citizen's tracker.

DEFIANCE addresses intimate forgeries specifically. For non-sexual deepfakes, common-law theories include false light (in jurisdictions that recognise it), defamation, right of publicity, intentional infliction of emotional distress, and copyright (where you hold rights in the underlying source material). Consult a lawyer in your jurisdiction; do not contact the perpetrator before doing so.

The Coalition for Content Provenance and Authenticity — an open standard for cryptographically signing the origin and edit history of digital content. Signed structures (Content Credentials) are embedded in the asset itself. Current spec: version 2.2 (1 May 2025).

Steering members include Adobe, Microsoft, Google, OpenAI, Meta, BBC, Sony, and Truepic. As of 2024, TikTok was the first major social platform to attach Content Credentials to AI-generated uploads; Google integrated Content Credentials into Search and ad systems; Amazon attached them to Titan Image Generator outputs; LinkedIn displays them on uploads.

ISO/IEC 21617 — the international standard for asserting media authenticity, provenance, attribution, IP, and integrity. The Core Foundation (Part 1) was approved for publication at the JPEG 105th Meeting in Berlin in October 2024. Complementary to C2PA, not a competitor.

Often, yes — though difficulty depends on perturbation budget. Saberi et al. (arXiv:2310.00076) showed a fundamental trade-off between evasion error and spoofing error for low-perturbation watermarks. Watermarking is a useful production signal for cooperating actors; not a closed defense against motivated adversaries.

The official public reader is at contentcredentials.org/verify. Drag a file in and the reader resolves the chain. Library implementations: c2pa-rs (Rust) and c2pa-js (JavaScript), Apache-2.0.

imadethisup.org is a public-education project of Global Cyber Institute, Inc., a 501(c)(3) nonprofit (EIN 84-2148770). It is anonymously run by thought leaders from across the cybersecurity, forensics, and policy industries. There is no advertising, no sponsorship, and no third-party tracking. See About.

Every claim links to a primary source — peer-reviewed paper (publisher canonical PDF or arXiv preprint), official standard, federal agency document, or originating news investigation. The full bibliography is at References. To report a broken link, email info@imadethisup.org.

Yes — content is released under CC BY-NC 4.0 and the underlying code is released under the MIT License. Attribution to imadethisup.org / Global Cyber Institute is required. For commercial reuse, write info@imadethisup.org.

Operating costs are covered by Global Cyber Institute's general nonprofit funding. There is no advertising, no paid placement, no sponsored content, and no cookie-based tracking. See the privacy policy for full data practices.

Email info@imadethisup.org with the URL of the page in question, the specific claim, and a primary source for the correction. Corrections are reviewed weekly.