Scientist Jordan Harrod Builds an "AI Tom Scott" Deepfake — on a Mere $100 Budget

Scientist Jordan Harrod has had a shot at replacing well-known maker Tom Scott with a deepfake — building an "AI Tom Scott" capable of mimicking his appearance and speech in a video recording, using a budget of just $100.

"This isn't a face replacement on a body double," Scott explains. "That's mainstream now - Hollywood's doing it, anyone can do it. This is trying to create a cheaper complete replacement for me, an automated digital puppet — on a budget of just $100."

"Some of you are getting better and better at creating Tom Scott-style videos," Harrod explains, "and it seemed inevitable that someone might try to use machine learning next, given the amount of free training data Tom has posted on the internet over the years. Given that, I decided to get ahead of you and do it myself."

While Harrod hasn't shared the specific tools used, in order not to incentivize others to follow in her footsteps, Scott has linked to several sources: The text for a plausible Tom Scott video was generated using the popular GPT-3 network from OpenAI, while the audio was created using Resemble.ai to replicate his speech. The facial recreation, meanwhile, relies on a techniques taken from a range of papers on generative adversarial networks (GANs) for both face-swapping and face-generation.

"I know in a world where Star Wars is using deepfakes, there might be a thousand 'bUt ItS nOt PeRfEcT' comments here, so just to be clear: This is what AI can do with the work of one individual, no body or voice doubles, and $100," Scott explains. "If you think you (or your plausibly deniable nation-state actor) can do better... then please don't try it on me!"

The full video is available on Tom Scott's YouTube channel, with links to sources in the description. More videos on the state of the art in AI are available on Jordan Harrod's YouTube channel.