On the banality of our deepfake future

What's real? What's fake? Can you tell the difference? If you can today, you won't be able to tomorrow.

We're on the brink of a new phase of human technology in which anything can and will be faked so well that no human will be able to tell the difference. 

This will be a catastrophe, some say. Others claims that it will be a huge boon. Both are true. But the most shocking emergent fact is that deepfake technology is becoming… a banality. It's boring. It's ubiquitous. And it's not really changing much for good or ill. 

Deepfakes use a revolutionary machine learning model called Generative Adversarial Networks (GANs) in which data (video, pictures, audio, line patterns — just about any digital data, really) is fed into the system. One model, called the generator, seeks out patterns in the data and attempts to create new fake version. Another model, called the discriminator, attempts to guess if an example presented to it is real or fake, and then is fed the answer after guessing, thereby improving its ability to judge if examples are real or fake. This process is repeated thousands of times and, in the process, the generator gets better at creating plausible fakes, and the discriminator gets better at telling the difference between the real and the fake. Eventually, the generator can reliably produce fake data that looks just like real data. 

In my Fast Company piece, How Holograms, Deepfakes, and AR Are Raising the Dead, I touched on the problematic ethics of using Deep Fake technology to create content using a dead person's face, body and voice. This, to me, is a clear ethical violation — to put words into the mouths of any dead person without their prior approval is simply the exploitation of that person's memory. 

A new documentary about food-and-travel media personality Anthony Bourdain, called Roadrunner, tests the limits of this ethical question. The film uses a deepfake recreation of Bourdain's voice to read his own written private emails. It's murky because they're his words. It's just that he never actually recorded them. 

Deepfake voice cloning is not only being used for social engineering attacks, but also for the creation of AI-generated songs by dead singers. 

Or live ones. Holly Herndon created (in partnership with a startup called Never Before Heard Sounds) a so-called "digital twin" — an AI-based tool that enables other people to make songs and other content using a deepfake of her voice. 

Holly Herndon @hollyherndon

Ok wow this is finally happening 🤯 Meet Holly+ my digital twin 👭 I built an instrument with @HeardSounds where anyone can make music with my voice 🗣 and I am starting a DAO to govern my digital likeness🪞 I can’t wait to hear what y’all make! 🌹 Holly.plus

Three TikTok deepfakes of Tom Cruise appeared a few months ago on an account called @deeptomcruise, which were created by visual effects specialist Christopher Ume with the help of Tom Cruise impersonator Miles Fisher. In fact, deepfakes of celebrities number in the thousands on YouTube. The Tom Cruise deepfakes made news because they were so good, and represent what's coming. 

We used to fear our deepfake future because nefarious evil-doers would create fake video "evidence" of politicians and others saying things they never said and doing things they never did. 

I've always feared the opposite: Knowledge of the existence of deepfake represents the end of video evidence as proof that something really happened. For example, former President Donald Trump actually made a concession speech, which some of his supporters dismiss as a deepfake, because they believe some false conspiracy or another about the 2020 election. 

The most present and damaging harm, of course, is to the victims of deepfake porn. In fact, the overwhelming majority of deepfake videos are pornography, mostly combining the faces of celebrities and politicians with existing porn videos. 

Like the Bourdain doc, other new uses for deepfake technology are ethically murky. Hollywood, for example, is excited about improving the visual experience for movies dubbed into other languages. Instead of words that don't match the mouth, deepfake tech promises a future where actors visually appear to pronounce Mandarin or Spanish or Tagalog, when in fact the actor originally spoke English. 

Deepfake technology is even being used to generate fake satellite imagery. 

Amazon uses deepfake technology to recreate the voices of celebrities for optional voices for its Alexa virtual assistant. They have permission. But creating deepfake versions is easier for the celebrity than recording all the responses.

Apps that do deepfake AI for the vague amusement of the masses are now a dime a dozen. 

Flying Dog Brewery even created a non-alcoholic IPA called Deepfake. 

Deepfake media that impersonates real humans, that fabricates fake humans and that generally tricks the eyes and ears, is here to stay. We simply have to adust to the fact that many of the things we think are real are — or could be — machine-made simulations.