Deepfakes only affect video
Deepfake technology, using AI to generate or manipulate synthetic media, extends well beyond video to include highly convincing audio voice cloning and increasingly realistic still images, and documented fraud and disinformation cases have used all three formats, not video alone.
What we know
The term deepfake originated in association with AI-generated or manipulated video, particularly face-swapping techniques that emerged publicly around 2017 on online forums, and this video-centric origin has left a lasting public impression that deepfakes are primarily or exclusively a video phenomenon. Synthetic media generation technology has since expanded substantially across other formats, including audio voice cloning and static image generation, both of which have been used in documented real-world fraud and disinformation incidents that did not involve video at all.
Voice cloning technology, which can generate synthetic speech convincingly mimicking a specific person's voice from a relatively small sample of their real recorded speech, sometimes as little as a few seconds according to some commercial voice cloning tools, has been used in several notable fraud cases. A widely reported 2019 case involved criminals using AI voice cloning to impersonate a German company executive's voice in a phone call, convincing a UK-based subsidiary's chief executive to wire approximately 220,000 euros to a fraudulent account, as reported by the Wall Street Journal citing the firm's insurer. A separate widely covered 2023 case reported by the Washington Post and other outlets involved a mother in Arizona who received a call using cloned audio of her daughter's voice, apparently generated from social media video clips, as part of an attempted extortion scam falsely claiming the daughter had been kidnapped.
AI-generated still images have also become increasingly difficult to distinguish from real photographs, and have been used in political disinformation campaigns, including fabricated images depicting public figures in situations that never occurred, and in romance scams, where fraudsters use AI-generated images of people who do not exist, avoiding the risk of a reverse image search revealing a stolen photo of a real person, a tactic security researchers have specifically flagged as an evolving countermeasure to that particular detection method.
Generative AI image tools, including Midjourney, DALL-E, and Stable Diffusion, along with widely available voice cloning tools, have become considerably more accessible to non-technical users since around 2022, lowering the barrier for creating convincing synthetic audio and images without specialized technical skill, a trend cybersecurity researchers, including those at the Department of Homeland Security and various academic institutions studying synthetic media, have specifically identified as expanding the practical scope of deepfake-related fraud and disinformation risk beyond the video-focused threat model that dominated earlier public discussion of the issue.
Detection research has correspondingly expanded to address audio and image-based synthetic media alongside video, with organizations including the Content Authenticity Initiative, a coalition involving Adobe, Microsoft, and major news organizations, developing content provenance standards intended to help verify whether images and other media have been AI-generated or manipulated, reflecting an industry recognition that the deepfake threat landscape spans multiple media formats rather than being confined to video specifically. The public focus on video deepfakes, while understandable given the format's origin and often more visually striking demonstrations, has left many people less prepared to critically evaluate suspicious audio calls or still images, a gap security awareness campaigns have increasingly sought to address.
Common claims
- Deepfakes are only a problem for politicians and celebrities in videosFalse. Audio deepfakes have been used in documented corporate fraud cases costing millions.
- You can always detect a deepfake by looking carefullyIncreasingly false. Audio deepfakes require no visual component, and video deepfake detection is an active arms race.
- Deepfake technology is too expensive for most fraudstersFalse. Commercial voice cloning services are widely available and inexpensive or free.