You Don’t Own Your Face
An introduction to deepfake technology and its near-term implications.
Deepfakes are manipulated images, videos or audio that appear to show a person saying or doing something that they never said or did. Deepfakes are a form of synthetic media: visual trickery that can seamlessly overlay a recognizable face onto an impersonator and turn a mimicry into a fully-fledged forgery.
You may have seen rapper Kendrick Lamar transform his face into celebrities like Kanye West and Will Smith in a recent music video. Jordan Peele’s eerie PSA deepfake of President Obama went viral in 2018. Or, you may have watched a politician making a shocking statement in front of a motionless camera while the edges of their face shake and blur. Welcome to the world of deepfakes.
Hollywood has been the forerunner of digital manipulation for years: using CGI to replace costumes, de-aging celebrities, building characters entirely from scratch. Today, sophisticated video manipulation is increasingly accessible thanks to open source code, lighter algorithms, free apps, and deepfake services.
Deepfakes borrow their name from a combination of the AI term ‘deep learning’ and the word… fake. It’s a clumsy portmanteau that hints at the larger technology at play here. Deepfakes are more complex than a high-quality Photoshop job. They’re constructed using algorithmic models that systematically replace the subject’s face in every frame of the video. In many cases, deepfakes are created using a generative adversarial network: a machine-learning system that pits two models against each other. The generator model creates suggested face replacements that are constantly evaluated by the second model, a discriminator, that provides feedback to the generator. The result is a program that teaches itself how to manipulate video until it can sufficiently fool itself.
Creepy, right? Imagine seeing a video of yourself emphatically speaking about a subject that you’ve never once mentioned. With only a reference dataset of images featuring your face (cough, social media), a deepfake model could easily create a video that features you saying anything that the programmer chooses.
Deepfakes are still an emerging technology. Not every manipulated video is going to deceive its audience. Manipulated videos often appear on social media and play for laughs. Plenty of others, though, aren’t so benign. Synthetic media, including deepfakes, have already been used to commit business fraud and share misleading information about politicians. A well-constructed deepfake is challenging to spot. You can explore this MIT site to test your ability to identify a manipulated image.
The technology reliably makes headlines for its potential political implications, but the most popular use of deepfakes to date comes as no surprise to any internet historian: pornography. A 2019 study estimated that approximately 96% of all deepfakes are used in pornographic videos. Most frequently, adult actors have their faces swapped with famous celebrities. It’s a horrifying trend that has now reached teenaged TikTok stars. As deepfake technology continues to both improve in quality and grow more accessible, it won’t be long before all women have reason to fear that their public images may be used to create sexual images without their consent.
Thankfully, there are also less disturbing applications of deepfake technology. Movie studios will salivate at the potential to hire A-list actors without ever requiring them to step onto a set. For amateur content creators, too, a new world of possibility is opening for parody and visual effects. Companies may utilize deepfakes as customer service representatives, allowing users to interact with an automated smiling face instead of a disembodied chat bot. Visitors to the Salvador Dalí museum in Florida are now greeted onscreen by the artist himself, at least until further inspection reveals the deepfake ruse.
None of these applications come without their own set of challenges. Deepfake performances in film will ruffle the feathers of purists who celebrate the craft of acting. Using the likeness of deceased celebrities invites ethical questions about whether we own our own image. Deepfakes intended to deceive will have to be countered with reliable accounts of provenance. New dimensions of copyright and privacy laws will be forged in courtrooms across the country.
Deepfakes are a tantalizing tool that should excite proponents of innovative AI systems. They are also plainly suited for fraud, deception, and sexual abuse. There’s no doubt that deepfakes will continue to proliferate and, soon enough, will be available to the average internet user. Responsible growth of the technology requires two components: (1) sufficient awareness that viewers will cast a critical eye on the media they encounter and (2) legal and ethical frameworks to guide the usage of deepfakes and to regulate those bad actors who will inevitably employ them for nefarious ends. We don’t need to be pessimistic about deepfakes — we need to be cautious and intentional.
Written by James Ostrowski & Alec Winshel