Fake videos of real people — and how to spot them

Key claims and methods

Face synthesis requires only publicly available data. The core technical innovation is that you do not need to 3D-scan a person or have special hardware. Given a collection of ordinary photos of a person (searchable from Google), the method:

1. Reconstructs a high-detail 3D face model, capturing fine facial geometry, creases, and wrinkles
2. Synthesizes realistic facial texture and color using a blending technique that works across expressions
3. Creates a controllable model that can be animated using static photos, video input, or audio

The 3D model is derived from a large photo collection by analyzing expression variation. For George W. Bush, the researchers downloaded photos from Google Images. By comparing a large set of photos of the same person, the algorithm identifies which pixels correspond to which 3D geometry, then separates static facial structure from dynamic expression. The iterative, refined model recovers fine details that a single photo cannot capture alone.

Facial texture is synthesized with a blending technique superior to simple averaging. A naive approach—averaging all pixels from all photos—produces a blurry result. Instead, the method selects and blends pixels from the highest-quality source photos, producing sharp textures and colors. This works for arbitrary expressions.

The model can be driven by video input, transferring motion from a source actor to the synthetic face. In a demo, Suwajanakorn shows transferring motion of Daniel Craig and George W. Bush speaking into models built from their photos, creating videos where the synthetic person says something authentic (or not) while exhibiting their natural mannerisms.

Audio-to-mouth synthesis creates lip-synced fake speech. A separate pipeline uses a neural network to:

1. Convert audio input into a sequence of mouth landmark points (how the mouth moves)
2. Synthesize the mouth texture (lips, teeth, gums)
3. Enhance fine details (teeth) and blend the mouth region into the face and background from a source video

The example: given only Barack Obama's audio from 14 hours of video addresses, the system generates a video where synthetic-Obama's mouth moves in perfect sync with new speech audio.

Positive applications are genuine but require human intent. The talk emphasizes:

• Preserving Holocaust survivor testimonies as interactive holograms ("New Dimensions in Testimony" project)
• Resurrecting historical figures like Richard Feynman to give new lectures in new languages
• Allowing deceased authors to read their own books aloud
• Enabling educational and creative use at scale

Risk and dual-use concern

Deepfakes enable misrepresentation at scale. A synthetic video of any public figure can be created with enough photos and videos. The potential for political manipulation, defamation, non-consensual intimate imagery, and misinformation is immediately obvious.

Suwajanakorn is transparent about his concern. He states: "I think these results seem very realistic and intriguing, but at the same time frightening, even to me. Our goal was to build an accurate model of a person, not to misrepresent them. But one thing that concerns me is its potential for misuse."

Public awareness is urgent. He emphasizes: "Despite all this, though, fake videos could do a lot of damage, even before anyone has a chance to verify, so it's very important that we make everyone aware of what's currently possible so we can have the right assumption and be critical about what we see."

Detection countermeasures (briefly mentioned)

AI Foundation and Reality Defender. Suwajanakorn notes he is part of an ongoing effort at AI Foundation, which combines machine learning and human moderation to detect fake images and videos. One planned tool is Reality Defender, a web-browser plug-in that can flag potentially fake content automatically in the browser.

Note: The talk does not provide deep technical detail on detection, reflecting that in 2018, robust detection was not yet mature. The emphasis is on making the capability known, not on assuring the audience that detection will solve the problem.

Context and significance

This is an early, high-visibility statement on deepfake capability—made at TED in 2018, before the term "deepfake" had fully entered public discourse and before the technology had been weaponized at scale in elections. Suwajanakorn's framing is notably balanced: he celebrates the potential for legitimate use (educational, memorial, creative), but does not minimize the risk. His explicit statement that the technology is "frightening" to its own creator carries weight.

The technical depth is accessible: the talk explains the pipeline (3D modeling → texture synthesis → animation → audio-driven synthesis) without requiring advanced ML background, making it suitable for a TED audience while remaining technically honest.

The talk also prefigures later discussions about the asymmetry of detection: creating convincing deepfakes is now trivial; detecting them reliably remains hard. Early deployment of Reality Defender as a browser plugin reflects a 2018-era hope in technical detection—a hope that was later tempered by the difficulty of scaling detection to billions of videos.

Notes

This talk is historically significant as an early, clear statement from a creator of the technology itself about deepfake capability and risk. Unlike much technical research, which is presented at academic conferences to a specialist audience, Suwajanakorn's TED talk reaches millions and frames the technology in terms a general audience can grasp.

The positive applications (educational resurrections, memorial testimonies) are presented not to minimize risk but to establish that the technology has genuine value. This makes the risk discussion more credible: the author is not claiming the technology should not exist, but that its existence demands vigilance and awareness.

The transparency about "concerning" the creator is notable. In 2018, many technologists were still optimistic that better detection or content-authentication standards would adequately address deepfakes. Suwajanakorn's emphasis on awareness and critical consumption prefigures later recognition that detection alone cannot scale and that the problem is as much social and institutional as technical.

The talk makes minimal claims about detection efficacy, perhaps wisely. Reality Defender is mentioned but not promoted as a solution. Instead, the core message is: "Everyone needs to know what's possible." This acknowledgment that awareness ≠ immunity is more honest than many contemporary talks on the subject.