Deepfake researcher Hany Farid says media should be presumed fake until proven real

Hany Farid, a computer science professor at Dartmouth College who has spent three decades building techniques to detect manipulation in digital media, opens his IEEE Spectrum essay with a Saturday morning in March. A half dozen reporters around the world wrote to him about a week-old video, aired by an Iranian news agency, of what appeared to be a U.S.-made Tomahawk missile hitting a school. News reports said the strike killed more than 175 people, most of them children, and the U.S. president and military denied responsibility. Everyone, Farid included, was skeptical of the video.
His first answer to the reporters was suspicious: the quality was low, the video emerged weeks after the bombing, and it was distributed only by a news agency with a less-than-stellar record of accurate reporting. But because the stakes were high if the video was real, he analysed it. First he aligned the frames on a stationary background landmark. If the missile had been digitally added without regard for camera motion, its path would look physically implausible; instead it moved in a nearly linear trajectory. Then he checked the missile's size in pixels. Under some reasonable assumptions, it matched a 6-meter Tomahawk about 100 meters from the camera. That distance fit the scene and fit the 0.3-second delay between visible impact and the sound of impact, given that sound travels at 343 meters per second.
Still, he struggled to give the reporters an answer. A recent study he worked on with colleagues at the University of California, Berkeley, showed how hard it is to tell real videos from fake ones, and that observers are now so suspicious that they have trouble convincing themselves a video is real. After a few more hours and analyses, Farid decided the video was authentic and told the reporters so. The episode was his turning point: a lack of artifacts was not enough to convince him either, so he had flipped from a presumption of real to a presumption of fake.
He now argues that everyone should make the same shift. The burden, he says, is on the creator to prove content is real. That may seem unfair to legitimate creators, but he thinks it reflects reality and points to a future where we can again trust that most online media is what it appears to be.
The practical answer is provenance. Forensic analysis of the kind he does does not operate at internet scale and speed; provenance signals can. Credentials following the C2PA standard attach a cryptographically signed manifest to a photo or video at capture, recording the device, the time and location, and any later edits, so tampering with the content or its history can be detected by machine. Watermarks such as Google's SynthID embed an imperceptible signal in pixels or audio samples that identifies the content's origin. On some of Apple's newest iPhones, the camera sensor can sign the raw pixel data as it is captured; the signed data is developed into a reference image saved alongside the normal, editable photo, and anyone can compare the two.
Farid addresses the usual criticism that these signals are only partial or can be attacked. That is true, he says, only if they are meant to catch bad actors. Their purpose is to let good actors prove their recordings are genuine, so they need not be adversary-proof. If enough good actors adopt them, they can begin to drown out bad actors. The model shifts from catching the bad to empowering the good.
The same distrust now reaches routine video calls. In September 2024, Senator Ben Cardin (D-Md.), then chair of the Foreign Relations Committee, received a video meeting request from Dmytro Kuleba, who until a few weeks earlier had been Ukraine's foreign minister. Cardin grew suspicious when Kuleba, out of character, pressed him on whether the U.S. should let Ukraine fire long-range missiles into Russia, and he ended the call. The State Department later confirmed Kuleba never requested the meeting, and the Senate security office warned staff that foreign adversaries, possibly Russian, would likely try similar faked meetings again.
Farid says such impersonation needs only a face swap or avatar deepfake plus voice conversion, both working in real time, and both available in commercial products and open-source libraries that run on a modern laptop. A live call leaves no time for human analysis. But real-time fakes must be generated at around 24 frames per second, so the software takes shortcuts, such as simplifying faces and allowing drift between audio and video. The eye does not notice them, he says, but forensics can expose them. He adds that provenance signals should eventually play a role here: video-calling software could sign each frame of a live stream, as Apple now does for stills, and verify in real time whether frames came from a physical camera sensor or were altered in transit.
His conclusion: forensics alone was enough for the first two decades of the digital revolution, but AI has changed the scale and speed of manipulation. To believe a recording is real now, it must also carry a provenance signal from the point of recording. The two approaches work together, with provenance allowing good actors to be trusted and forensic analysis rooting out the bad ones.
Key facts
- Hany Farid, a Dartmouth computer science professor, says that after analysing a video of an alleged Tomahawk strike on a school and judging it authentic, he realised he had moved from a presumption of authenticity to a presumption of fake.
- His checks: a nearly linear missile trajectory after frame alignment, a size matching a 6-meter Tomahawk at about 100 meters, and a 0.3-second impact-to-sound delay consistent with that distance at 343 meters per second.
- He argues provenance signals (C2PA content credentials, watermarks such as Google's SynthID, sensor-level signing on some of Apple's newest iPhones) let good actors prove their recordings are real, and need not be adversary-proof.
- Real-time deepfakes already run on a modern laptop; the 2024 fake Kuleba video call to Senator Ben Cardin was ended only because Cardin grew suspicious.
- Farid proposes that video-calling software could sign each live frame and verify it against a physical camera sensor, but this is a suggestion, not a deployed feature.
Why it matters
Farid's claim comes from someone whose job is spotting fakes. He says the default for visual recordings has to flip: instead of the skeptic proving a video is fake, the creator must prove it is real. His own example is telling. A video he ended up judging authentic still took hours of analysis, and a lack of artifacts alone did not convince him. He ties this to a study with Berkeley colleagues in which observers were so suspicious that they struggled to accept a real video as real. He also argues forensics alone no longer scales, so trust now has to be built in at the point of recording.
Who it affects
Farid names consumers, journalists and the courts as the groups that must adapt to a presumption of fakeness. Reporters who sent him the Tomahawk video are one example; he is also a frequent expert witness in U.S. courts and advises law enforcement. Anyone on live video calls is affected too, as the Cardin case shows. Legitimate creators bear the burden of proof under his proposed stance, which he admits may seem unfair to them.
How to use it
This is an argument, not a product, but it points to concrete mechanisms. C2PA credentials attach a signed manifest at capture recording the device, time, location and later edits. Watermarks like Google's SynthID embed an origin signal in pixels or audio samples. Some of Apple's newest iPhones can sign raw sensor data and save a reference image next to the editable photo, so anyone can compare the two and see what changed. For video calls, Farid suggests software could sign each frame and check it against a physical camera sensor in real time; he frames that as something that should eventually happen.
How solid is it
It is a first-person opinion essay, and the central stance is explicitly the author's view ('I think we need to shift'). The Tomahawk analysis is Farid's own, with stated assumptions ('some reasonable assumptions'), and the essay describes no independent confirmation of the video's authenticity beyond his analyses. The Cardin incident, including the State Department's confirmation that Kuleba never requested the meeting, is presented as background. The essay does not give the Berkeley study's sample size, methods or numbers.
Risks and caveats
Farid concedes that provenance signals are often criticised as partial or vulnerable to attack, and answers that they are meant to empower good actors, not catch bad ones. His hope that enough adoption will let good actors drown out bad ones is stated as a possibility ('can begin'), with no timescale for when trust would be restored. Real-time frame signing for video calls is a proposal. The essay does not say who fired the missile or who is responsible; it concludes only that the video is authentic. A short closing paragraph on the page is in a different voice from Farid's essay and bio and should not be read as his; it says forensic analysis may still be needed for edge cases such as older hardware or unsafe situations.
“To believe a recording is real now, it must also carry a provenance signal from the point of recording.”
— Hany Farid, IEEE Spectrum