Hany Farid, a UC Berkeley professor who specializes in the field of digital forensics, sits at his home in Vermont.
John Tully/For the S.F. Chronicle
Inside the windowless operations room of what appears to be an Iranian military base, officers in camouflage pace and shout into headsets.
Suddenly, the overhead lights flicker. The ground trembles, and a pressure blast sends plumes of thick black smoke billowing into the room. Officers fly from their seats. Shooting sparks ignite into an inferno.
When this video — purporting to show an Israeli missile strike — began circulating on X last December, it looked plausible and devastating. The footage quickly went viral, catching the attention of reporters and editors across major global news outlets. But one editor had some doubts. So she sent the footage to the world’s premier expert in deepfakes, Hany Farid.
Article continues below this ad
It took Farid just a few minutes to confirm the video had been made using artificial intelligence. “Looking at videos like this is sort of my life,” he said. “Some mornings I’m watching videos of people getting their heads chopped off before I’ve even rubbed the sleep from my eyes.”
Farid, a UC Berkeley computer science professor who recently departed to teach at Dartmouth, is best known for his work as a deepfake detective, analyzing videos and images for traces of manipulation by digital or AI tools.
San Francisco Chronicle Logo
Make us a Preferred Source to get more of our news when you search.
Add Preferred Source
It’s a daunting task. Over the past year, deepfake content has grown at a rate of roughly 900%, according to the cybersecurity firm Deepstrike, driven by the rise of generative AI apps like ChatGPT, Gemini, ElevenLabs and Google’s Veo 3. Together, these tools have made it easy to create fake videos and audio so realistic they’re exceptionally hard to detect.
The consequences are profound. Fake footage of military conflicts is reaching newsrooms and defense officials. Corporations are losing tens of millions of dollars to impersonators on Zoom calls who digitally disguise themselves as bosses and demand wire transfers. Predators on the internet are mining the social media accounts of ordinary women and minors to create hyper-realistic porn in their image. According to Farid, nearly half of what we encounter online may already be synthetic media, generated or altered using AI.
Article continues below this ad
To combat this, Farid serves as an expert witness in court cases and has repeatedly testified before Congress. He’s a leading advocate for the creation of digital watermarks that would identify AI-generated content. And he continues to pioneer ways to detect digital deception. Using techniques like shadow calculus — the mathematics of how light and shadows behave — and his own AI models, he searches images and videos for invisible artifacts like pixel anomalies, light sources that don’t perfectly obey physics and perspective lines that fail to converge.
The cybersecurity firm he co-founded in 2022, GetReal, holds contracts with U.S. and foreign governments, defense contractors, tech companies and news outlets including Reuters and the Associated Press.
Lily Xu, an assistant professor at Columbia who has known Farid for nearly a decade, describes him as “probably the best-known and most important person on the front lines of the deepfake war.”
Professor Hany Farid teaches aclass at UC Berkeley’s School of Information in 2022.
Brittany Hosea-Small for UC Berkeley
Vanishing points
While it’s long been possible to alter digital photos, manipulating audio and videos was more difficult — until recently. AI text-to-video applications, in which a user can merely type instructions to create videos, and the explosion of platforms that allow people to pay for custom deepfakes, like Xanthorox, now use technology once reserved for specialists at Hollywood studios.
Article continues below this ad
Scammers, meanwhile, can commission synthetic audio or video by using just three to five seconds of clean audio or a few publicly available photos — at a cost of about $5, according to experts on biometrics.
The leap in realism between first-generation AI videos and today’s deepfake content is making detection harder for experts. Generative AI footage produced before 2022 was riddled with inconsistencies: faces morphing between frames, warped jawlines, phantom limbs appearing and then vanishing.
Newer AI footage holds identities stable across hundreds of frames with no perceptible degradation. It can even replicate the texture of authentic video — the tremble of a handheld phone, the chaos of a crowd.
“People used to need to know things like how to use a darkroom to manipulate media, which took years of training,” said Siwei Lyu, a professor at the University at Albany and director of its media forensics lab. “But now with AI, you don’t need to know anything.”
At his Berkeley office on a recent afternoon, Farid clicked the “media requests” folder on his desktop, containing all the footage that’s ever been sent to him for verification by partnering media outlets. To demonstrate, he pulled up the video showing the apparent Iranian bombing. Scrolling through, he began describing how he detects inauthentic footage. He paused on a frame that seemed to show computers shattering and flying off the table.
Article continues below this ad
“This is exactly what you’d expect a strike to be like,” he said, pointing at the screen. “It just obliterates the place, it’s so violent. Even this” — he hovered his mouse over a ray of light in the corner of the frame — “This is called lens flare. It’s exactly what you’d get when a light is shining directly into the camera.”
Still, he noted, there were hidden clues that identified the video as a fake. Dragging his mouse over the scene, he began to identify the scene’s “vanishing point”: the single dot in space where the lines of all the physical objects would converge if the image were a real three-dimensional scene, captured by a camera. Using his cursor, Farid drew a number of straight red lines emanating from each object, then extended their endpoints outward to see whether they would intersect. They didn’t.
“This shows an anomaly,” Farid said. “A physical inconsistency.”
To confirm his findings, he dragged the screenshot to what he calls his “forensic workbench” where a software program runs a slew of measurements. It searches for information on the image’s history, the identity and location of whoever created the original footage and the path it has traveled since creation.
Within seconds, a red rectangular sign began flashing in the app’s upper-right corner: “SYNTHETIC.”
Article continues below this ad
After he initially found the footage to be fake, Farid contacted the editor who’d sent it to let her know. She didn’t move forward with the story.
Hany Farid fields requests from the government and news outlets to evaluate videos for signs they are AI-generated.
John Tully/For the S.F. Chronicle
Out of the darkroom
When Farid was 2, his family moved from Germany to Rochester, New York, where his father took a job at Eastman Kodak as a photo chemist. Growing up, Farid remembers watching his father develop images in darkrooms, bewitched by the smell of the chemicals.
While at the University of Rochester, Farid’s mother encouraged him to take a computer science class. It felt like he had found something he was truly good at. He made it his major, and went on to earn his Ph.D in computer science at the University of Pennsylvania.
Later, while completing his postdoctoral work at MIT in 1998, he had a career-defining epiphany. Thumbing through a newly updated copy of “Federal Rules of Evidence,” Farid came across Rule 1001 of article X, stating that for the first time, U.S. courts were allowing digital images to be submitted as evidence.
“I just remember thinking,’‘that’s a really bad idea,’’ he said. ‘Because digital images are malleable.’ Until then, only 35 mm negatives had been categorized as ‘original evidence’ in court because physicality made them difficult to alter.
Farid went home and began exploring how to detect manipulation in digital images. Using Photoshop, a relatively new computer program at the time, he began altering photographs — swapping the heads of his friends — then writing software to detect markers of the manipulation and reverse-engineer what he had done.
In 1999, when Farid became a professor at Dartmouth, he continued his efforts, establishing a lab that worked on some of the first digital authentication cases for the Associated Press and the FBI. But trying to establish a new academic discipline, which he called digital media forensics, was difficult. Funding was scarce, and there were no journals or conferences at which he and others could present their findings. Even most computer scientists didn’t understand the problem.
“At the time he was working on these things, basically nobody was,” said Ted Adelson, professor of computer and cognitive sciences at MIT and Farid’s postdoctoral advisor. “Nobody had any idea how to approach it.”
In the two decades since then, the field has exploded. Experts like Farid are racing to keep up with users harnessing deepfake technology for malicious ends. Scammers have already begun using sophisticated audio and video deepfakes to defraud families and older people. Rather than using premade audio fakes, the current standard, impersonators will soon be able to perfectly mimic the voice of a distressed family member in an emergency, in real-time calls. According to the Deloitte Center for Financial Services, AI fraud could cost the U.S. $40 billion by 2027.
A similar boom is happening around unauthorized deepfake pornography. In December 2025, images began flooding X showing women and children whose clothing had been digitally removed or altered to depict them in suggestive poses — all generated by Elon Musk’s chatbot, Grok. The European Commission estimates that 98% of deepfakes are sexually explicit material.
“AI companies commercializing and monetizing these tools are not doing enough to keep them safe,” Farid said.
Hany Farid, a UC Berkeley professor, is a go-to source for governments and journalists seeking to verify whether a video is AI-generated or real. His techniques include analyzing shadows and light.
John Tully/For the S.F. Chronicle
The liar’s dividend
A few recent bills, like the Take It Down Act, signed into law in May 2025, and the DEFIANCE Act, passed by the Senate in January 2026 and now headed to the House, are meant to protect individuals from nonconsensual imagery. But they only apply after a person has already been harmed, Farid said –– and address only one small part of the problem.
Farid is working on one possible solution as part of the Coalition for Content Provenance and Authenticity, a project drafting standards for marking media as authentic as soon as a camera captures it by sealing elements of the media’s metadata immediately.
Despite these efforts, Farid remains deeply worried by new-generation deepfakes and their potential for abuse, including by politicians and governments.
In January, the White House posted a digitally altered photo on X of Minneapolis civil rights lawyer Nekima Levy Armstrong after she was arrested during an ICE protest.
In the altered image, Armstrong appeared visibly distressed, with tears streaming down her face. Her skin tone also appeared to have been darkened. By comparison, an image posted to X 30 minutes earlier by then-Homeland Security Secretary Kristi Noem depicted Armstrong looking notably calmer.
Having the White House publish synthetic content is deeply troubling, Farid said, because it muddies the historical record and makes it harder for the public to trust any visual evidence published by government officials.
It also contributes to what’s known as the “liars’ dividend” — making it easier for officials to dismiss real, incriminating footage as “AI” when faced with documented accusations of wrongdoing.
“You create instability in a society not by convincing people of a lie, but by making the signals so noisy that people don’t know what to believe,” Farid said. When people are exposed to generative AI content from state-sponsored actors, he said, it encourages them to “check out” and disengage from the democratic system.
Even history may be subject to manipulation. AI videos circulating online over the past year have begun showing historical figures like Martin Luther King Jr. and Albert Einstein behaving in ways they never did, or delivering speeches they never gave.
Farid worries about what will happen when younger people encounter these videos before being taught the truth about these monumental figures. What happens when fiction becomes the first impression?
“Our new reality is that we don’t have a shared reality,” Farid said. “And that’s exceedingly dangerous for us as individuals, institutions, societies and democracies.”
Fiona Ulrich is a freelance writer.