TOKYO - A cosplayer dressed as a popular female character became the target of a malicious deepfake—his image was manipulated by generative AI to depict him in underwear. The victim, who is actually a man, said he was mistaken for a woman and found images of himself altered into R18 content circulating online.
As generative AI rapidly evolves, the boundary between real and fake continues to blur. Concern is growing over the proliferation of highly convincing yet harmful fabricated images and videos.
To explore ways to identify these fakes, Fuji TV’s “It!” program interviewed two companies with proprietary AI technologies. Their insights reveal new methods for detecting AI-generated forgeries that are spreading at unprecedented speed.
Among the examples examined were videos created using “Sora,” the new AI model released this month by U.S.-based OpenAI. While some clips can be recognized as AI-generated upon close inspection, many are so realistic that viewers may easily be deceived. One video, for instance, showed a heated exchange between two men, entirely fabricated by AI. The only input required was a short text prompt—meeting certain conditions, anyone can create such content with ease.
These technologies have already sparked global controversy. A recent fake photo posted on social media appeared to show Prime Minister Koichi and opposition lawmaker Kiyomi Tsujimoto shaking hands and smiling in the LDP president’s office. Tsujimoto later clarified on X (formerly Twitter) that the image was fabricated, noting that even newspapers had inquired about its authenticity.
Meanwhile, the male cosplayer’s fake image spread online without his knowledge. “It was just a normal photo of me in a school uniform, but it was altered to look like I’d taken my top off. Using someone’s image without permission is simply wrong,” he said. The image was discovered only because someone happened to report it to him. He warned that such misuse could soon threaten ordinary people’s daily lives.
According to police, reports and consultations involving AI-generated sexual images of minors exceeded 100 cases last year. In one case, a male student was referred to prosecutors for creating and sharing fake nude images of a female classmate. Authorities believe most of these malicious fakes are produced by individuals known to the victims.
To understand how easy it is to create such content, “It!” visited a company that launched Japan’s first AI video talent service. Using a single photo, its system can generate a 3D model with realistic movement in about two hours. “You can even change the clothing,” explained the engineer, as a test image of a director was transformed into a 3D avatar in a suit, complete with natural gestures and facial expressions.
The company also demonstrated AI-generated voices nearly identical to the original person’s speech. When compared side by side, even the director admitted, “It sounds just like me. The tone and breath at the end of sentences feel authentic.”
Developers say their goal was to make professional-quality video production accessible to the general public as social media demand for short, high-quality videos grows worldwide. However, preventing misuse has become a pressing challenge. “We’ve implemented automated detection systems and reporting mechanisms to prevent unauthorized use of celebrity likenesses,” the company said.
Abroad, copyright and ethical issues are intensifying. In the United States, AI-generated videos depicting late celebrities—such as Michael Jackson appearing to rap—have sparked fierce debate. One fake video of civil rights leader Martin Luther King Jr. making animal sounds during a speech drew sharp criticism from his relatives, who called it “deeply offensive.” OpenAI responded that it prohibits the deliberate recreation of real individuals and will suspend accounts that violate this policy.
AI experts warn that the spread of such technology is forcing both media professionals and the public to reconsider how authenticity can be verified. To help, Japanese firms are now commercializing AI-based fact-checking systems. These can automatically analyze videos and detect signs of manipulation, such as mismatched lip movements, unnatural body motion, or inconsistencies between voice and environment.
One company demonstrated how its system analyzed a 10-second fake video in just three minutes, concluding that it was “highly likely to be synthetic.” The analysis flagged the subject’s hand movements as particularly unnatural—joints moved independently when they should have moved together.
Experts recommend several practical tips for spotting deepfakes. First, check for visible logos such as “Sora,” though many can now be removed with editing tools. If no logo is visible, examine the video’s length—current AI models struggle to generate long, coherent clips. Finally, look for subtle irregularities in body motion and synchronization between audio and visual cues.
While such detection tools are currently available only to corporations, experts say public awareness will be vital as generative AI becomes increasingly advanced and accessible. As one developer put it, “Every new technology brings both innovation and abuse. The key is to learn how to recognize and respond to it.”
Source: FNN
