Author details how he created a deepfake video over two weeks at a cost of $552 using Faceswap, with an explanation of the tech and some of its limitations
I learned a lot from creating my own deepfake video. — Deepfake technology uses deep neural networks to convincingly replace one face with another in a video. Tweets: @mathewi , @yoda , @aprilwright , @arstechnica , and @binarybits Tweets: Mathew Ingram / @mathewi : Here @binarybits describes how he created a deepfake video of Mark Zuckerberg for $552 by mapping video of the character Data from Star Trek — of course, it helps that Mark actually looks a lot like Brent Spiner: https://arstechnica.com/... Drew Olanoff / @yoda : i'm gonna keep warning y'all. this is going to be the technology that does the most damage. https://twitter.com/... April C. Wright / @aprilwright : A *really* good article about the hows and whats and current state of #deepfakes And I agree with the conclusions, particularly: inevitable “reality apathy” https://arstechnica.com/... @binarybits @arstechnica : “I started with a video of Mark Zuckerberg testifying before Congress and replaced his face with that of Lt. Commander Data from #StarTrekTNG. Total spent: $552.” https://arstechnica.com/... https://twitter.com/... Timothy B. Lee / @binarybits : There's been a lot written about the social implications of deepfakes, but less about how they actually work. Here's a thread about that. Read my article here for the full details. https://arstechnica.com/...
Context & Ripple Effects
Timothy B. Lee's $552 Faceswap experiment lands a month after the viral Zuckerberg and Boris Johnson deepfakes made by Bill Posters' team, but it answers a different question: not what a well-resourced activist can do, but what one journalist with an open-source tool and two spare weeks can do. The choice of subject is deliberate — swapping Data onto Zuckerberg exploits a real facial resemblance, which is exactly why the result reads as a proof of accessibility rather than a political stunt.
The piece also documents the limitations honestly, which matters because the field splits two ways from here: commercial vendors like Synthesia are packaging the same underlying tech into corporate training videos ([[a:955505]]), while detection work like DARPA's fake-spotting research races to keep up. A published cost-and-time baseline sits right at the hinge between those two trajectories.
First-order effects
- Readers now have a concrete floor for entry-level deepfakes — $552 and two weeks on open-source Faceswap — replacing vague fears with a reproducible recipe that names its own failure modes.
- Mark Zuckerberg becomes the de facto test subject for the technology twice in one year, following the earlier viral fakes, keeping his likeness at the center of the synthetic-media debate.
Second-order effects
- A mainstream explainer with a published bill of materials invites imitation: within months, online creators were publishing YouTube instructions for their own deepfake memes ([[a:957391]]), turning a one-off experiment into a self-teaching ecosystem.
- Detection efforts like DARPA's gain urgency and a clearer benchmark, since every documented cheap forgery defines exactly what detectors must catch.
Third-order effects
- If hobbyist-grade tools keep closing the quality gap, the binding constraint shifts from skill to intent — pushing platforms and regulators toward provenance and authentication layers rather than tool bans.
- The same pipeline has a legitimate twin: startups like Hour One now pay people to license their likenesses for AI-voiced video ([[a:970128]]), suggesting synthetic faces split into a consented commercial market and an unconsented fringe sharing identical infrastructure.
The trend: Deepfake creation is collapsing from a specialist capability into a documented weekend project, forcing the industry's response to move from detecting individual fakes to building trust infrastructure for all video.