Farid Argues for Presumption of Fake in Age of Generative AI
In a column for IEEE Spectrum, computer science Hany Farid argues that users must shift from a presumption of authenticity to a presumption of fake when viewing online media.
The Tech TL;DR:
- Generative AI and real-time deepfakes have made visual evidence unreliable for human analysis alone.
- Experts propose shifting the burden of proof to creators by requiring cryptographic provenance signals at capture.
- Real-time video calls are increasingly vulnerable to live face-swap impersonations requiring cryptographic stream verification.
In March, a flurry of emails and text messages from a half dozen reporters around the world disrupted a quiet Saturday morning for Farid, who has spent three decades developing mathematical and computational techniques to analyze and detect manipulation in digital media.
News reports alleged that more than 175 people, predominantly children, perished in an attack after an Iranian news channel broadcast a week-old clip depicting what appeared to be a school struck by a U.S.-built Tomahawk missile. The U.S. president and military denied responsibility.
Farid initially approached the footage with suspicion due to its low quality, the delay of weeks before its emergence, and its distribution by an Iranian news agency with a less-than-stellar record of accurate reporting. Given the potential implications if the video was real, however, he spent his Saturday morning analyzing it.

Aligning Video Frames Reveals a Linear Missile Trajectory
Farid reasoned that if the Tomahawk missile was digitally added to each frame without considering camera motion, the missile path would appear physically implausible. By aligning the video frames using a stationary background landmark, he observed that the missile moved in a nearly linear trajectory.
He estimated that the missile corresponded to a 6-meter Tomahawk at a distance of approximately 100 meters, which aligned with the scene and a sound delay of 0.3 seconds from visible impact to impact sound, given that sound travels at 343 meters per second.
Despite this mounting evidence, Farid struggled to finalize his assessment. His skepticism was fueled by a study he conducted with colleagues at the University of California, Berkeley, which showed how difficult it is for observers to distinguish real from fake videos when general suspicion is high. After hours of additional analysis, he concluded the video was authentic and shared his findings with reporters.

Cryptographic Provenance Signals
Because forensic analyses do not operate at internet scale and speed, Farid argues that provenance signals must step in to establish authenticity. These signals include content credentials and watermarks embedded at the moment of creation, such as content credentials following the C2PA standard that attach a cryptographically signed manifest recording the device, time, location, and subsequent edits.
Google’s SynthID embeds an imperceptible signal directly into pixels or audio samples, while Apple allows camera sensors on newer iPhones to cryptographically sign raw pixel data.
Farid notes that provenance signals are often criticized as vulnerable to attack or only a partial solution for catching bad actors. Their true purpose, however, is to let good actors prove their recordings are authentic.
“These signals are part of a new reality in which consumers, journalists, and the courts must adapt to a presumption of fakeness,” Farid writes. By empowering good actors rather than just trying to catch bad ones, verified media can begin to drown out fakes.
Real-Time Video Call Impersonations
This widespread distrust of digital media now extends to routine video calls. Dmytro Kuleba, who had recently concluded his service as Ukraine’s top diplomat, reached out via a video conference request to Senator Ben Cardin, a Maryland Democrat who at the time led the Foreign Relations Committee, in September 2024.
Cardin became suspicious when Kuleba pressed him on whether the United States should allow Ukraine to fire long-range missiles into Russia. The State Department later confirmed Kuleba never requested the meeting, and the Senate security office warned staff of potential foreign adversary impersonation attempts.
Born February 10, 1966, Hany Farid is an Egyptian-American university professor who serves as Dean and Head of School for the UC Berkeley School of Information. Often called the father of digital forensics, he is one of the designers of the PhotoDNA algorithm and author of the book Photo Forensics (2016).
His academic path includes a computer science and applied mathematics undergraduate degree from the University of Rochester in 1989, a computer science M.S. from SUNY/Albany in 1992, a computer science Ph.D. from the University of Pennsylvania in 1997, and a post-doctoral program in Brain and Cognitive Sciences at the Massachusetts Institute of Technology completed in 1999.
Face-swap avatars and voice conversion tools can execute real-time deepfake impersonations using commercial products and open-source libraries running on standard laptops. While live calls leave no time for manual forensic analysis, real-time software must generate video at roughly 24 frames per second, forcing shortcuts that introduce slight drift or simplified facial interactions.
Farid suggests that video-calling software could eventually apply cryptographic signing to live streams, verifying whether incoming frames originated from a physical camera sensor.