“Trump Announces North Korea Nuclear Pact”

I settled on the kind of emotional and SEO hook that could be unleashed this Fall as Americans go to the polls in a critical mid-term election. This post is really about the AI tool known as “Deep Fake:” what it is, why it’s important, and the opportunities and challenges that it presents in the future. 

“Well, who you gonna believe? Me, or your own eyes?”

Chico Marx in “Duck Soup.”

Deep Fake is technically elegant, commercially intriguing, and socially challenging. With it you can create exceptionally high quality video of any person saying anything. Yup: from Kim Jong Un announcing unilateral disarmament to Jeff Bezos vlogging that HQ3 will open in Caracas in 2020 it will all be possible. Just as it may be possible to recreate seminal events like that Lincoln address, film new Westerns starring John Wayne, or choose just the right visage for a remote instructor or therapist. If you haven’t seen a DeepFake video here is Barak Obama, err, Jordan Peele, to tell you about it. 

DeepFake is an application of Deep Learning; check out this 8-minute video tutorial. Keep in mind is that it’s just a bunch of math. There’s no magic or mysterious “we don’t know what it’s doing or how it’s doing it.” It’s just a computer doing what computers have always done: a ton more calculations with way more speed and accuracy than humans can conceivably achieve with paper and pencil. As a digital technology it’s growing exponentially. From the first presentation of the underlying technique at CVPR 2016  to the Fall 2017 appearance of Reddit user DeepFakes’ celebrity porn (banned in February this year) to the January 2018 release of an open-source application, FakeApp, and plenty of good how-to videos and blogs,  the technique has gone from university research to a do-it-yourself home project in two years. Exponential technologies don’t slow down: last month saw the release of full-on body-swapping with Everybody Dance Now. Where is it all going? 

Given the doubling of GPU performance roughly every 2.5 years and continuous application optimization it’s reasonable to conjecture that in about 20 years we’ll be able to run highly convincing person-swapping apps in real time on our phones. How about having your ten-years-younger-self show up in that beach video, or the “perfect you” in the video conferences that we’re all doing from home? Already today teams with powerful desktop GPU hardware and access to enough images can, as the Jordan Peele video proves, develop compelling content. Granted you currently have to be dedicated to slog through lot of steps and tweaking to make it work. Powerful results are achievable and as with all powerful technologies we will see the productive, the unproductive, and the downright stupid. 

I can’t resist starting with the latter: remember iFart? OK, iFart’s revenue was not stupid. The combination of human creativity and bad taste has boundless potential with DeepFake. (Is anyone else imagining notable politicians literally talking out of their derrieres?) Personally, I’ll settle for a video of me playing in the Warriors’ backcourt with Steph Curry. Entertainment will be a major DeepFake application. 

Balmain’s Margo, Shudu, and Zhi

Today Dwayne “The Rock” Johnson commands nearly $20 million up-front for a feature film plus part ownership of the content. The Rock could have limitless opportunities to license his image for multiple simultaneous movies. Studios will be massively incentivized to create proprietary digital human stars, perhaps using GANs to generate imaginary faces that are then DeepFaked onto anonymous minimum contract body-actors. Balmain’s new campaign will feature three CGI “supermodels.” The question is not whether digital human talent will be pervasive in entertainment, it’s which generation methods will prevail in a 21st-Century platform standards war. I can hardly wait to see the competing product marketing pitches: “The DeepRunway suite crushes any ray-tracing CGI methodology by delivering real wardrobes and FIMPS (facial imperfections): Real(tm) moles, Real(tm) dentals, and Real(tm) blemishes from our 5th-generation GAN algorithm.” It’s going to be great. 

While the public has accepted CGI supermodels it will not readily accept, but is  highly susceptible to, fake vignettes of elected officials. The incentive to create counterfeit politicos is irresistible. Fake information is as old as politics. Boomers well recall the forged Canuck Letter that destroyed Ed Muskie’s 1972 presidential campaign, and Jonathan Swift noted 300 years ago, “Falsehood flies, and the truth comes limping after it.” We’ll see more than one “secretly filmed” DeepFaked private fundraiser talk given by leading candidates this Fall and beyond.

Demand for knowing what is and is not DeepFaked will create an attractive video authenticity market. Competition has already started. Factom proposes to use a blockchain-based method to track the generation and ownership of video content. Researchers at SUNY-Albany have published detection methods based on facial actions such as blushing and eye-blinking that are difficult to map because of a relatively small quantity of training material. Just as with security breaches in computer networks I’m confident that we will see an ongoing cat-and-mouse game between the fakers and the authenticators. That, ladies and gentlemen, is what I call recurring revenue. 

Ultimately individuals will have the ability to create and project their own realities, as well as tools for detecting what is real and what is an illusion. I have no doubt that the opportunities for creating tools to create fakes, to detect fakes, and the value of the human skill of deciding between what makes sense and what does not will explode in the next 10 years. 

Leave a Reply

Your email address will not be published. Required fields are marked *