For years, one of the unspoken rules of the internet urged users not to believe everything they read. The boom in artificial intelligence, in particular generative AI, now means internet users can’t believe everything they see online, either.
Thanks to everything from advancements in machine learning to rapidly improving deepfake technology, computers are now so good at producing imitations of the offline world that the naked eye alone can’t always separate the real from the synthetic.
Use Cases
AI has a variety of use cases, from helping with data analysis to aiding in medical diagnoses. Deepfakes, meanwhile, represent the flipside of AI—producing images, audio, and video with the underlying purpose of deception.
The vast majority of AI-produced audio and visual content does not qualify as deepfakes. For example, if a user uses a text-to-image generator to create a woman playing a guitar, that’s simply an artificial creation. Some companies, like Meta, make it clear the imagery is original and made via AI.
Experts have zeroed in on intent to deceive as the defining feature of what qualifies as a deepfake. For example, a user would then need to superimpose, say, Taylor Swift’s head onto the body of the woman playing the guitar. Manipulation and intent to fool people are at the heart of this technology—and, given how sophisticated AI has become, the danger here is particularly acute.
Deepfakes aren’t just starting to proliferate; they’ve gotten very, very convincing (take this quiz to see if you can spot them).
How it Works
The word deepfake is a portmanteau, combining the computational concept of “deep learning” with the word “fake.” Deep learning is an AI approach wherein computers are taught to recognize complex patterns in things like pictures and sounds to produce realistic creations of some sort (like an insight or prediction).
Deep learning is a subset of machine learning, which allows algorithms to learn and adapt with a bare minimum of human intervention required. The technique is built upon neural networks—systems with interconnected nodes in a layered structure, similar to the neural pathways of the human brain. “Deep” refers to the multiple layers in this structure.
The second ingredient is simply human intent.
Deepfakes use that technology to create something incredibly realistic with the sole purpose of trickery. This is generally done in one of two ways—existing content can be manipulated to include artificial elements (such as by superimposing someone’s face into a photo or video) or AI tools can produce realistic content entirely from scratch.
The latter is generally done by software trained on an extensive corpus of written, pictorial, or video data. Training the system allows it to establish relationships between words, images, and objects. The output can be surprisingly realistic facsimiles of the real world.
The Danger
In their worst form, these technologies can be used to make the general public believe a wide variety of untrue scenarios.
Examples include attempts to influence election outcomes (listen to a fake robocall of President Joe Biden), attempts to steal financial information from unsuspecting people, and more. After seeing a deepfaked video of himself (watch here), legendary investor Warren Buffett said he felt like a “genie” was being let out of a bottle.
In nefarious cases, creators range from organized campaigns to ad hoc bad actors. It’s believed that elections will be an increasingly tempting target for state-backed deepfake efforts, while smaller-scale targeting will continue to proliferate. Experts say the technology has democratized and accelerated the spread of hypereffective disinformation.
An AI developer explains why we should be worried about deepfakes
13:07
AI-generated media that looks and sounds exactly like the real world will soon permeate our lives. How should we prepare for it? AI developer Tom Graham discusses the extraordinary power of this rapidly advancing technology, demoing cutting-edge examples—including real-time face swaps and voice cloning—live from the TED stage.
More on Deepfakes6 Videos
Dive Deeper
Relevant articles, podcasts, videos, and more from around the internet — curated and summarized by our team
With current deepfake technology, we can create footage of an alternate reality where humankind’s giant leap beyond Earth’s edge fell short. As Apollo 11 astronauts prepared for their journey to the moon, two versions of a speech were drafted for President Richard Nixon; one to deliver upon the astronauts’ success and another to comfort listeners should the mission fail. You can now listen to the contingency speech thanks to deep machine learning.
Generative adversarial networks (GANs) are getting better at making faces (and other images). GANs pit two neural networks against one another, one creating images and the other trying to detect whether they’re real or fake. As technology advances, it’s grown harder for humans to tell the difference. This article provides clues revealing an image is fake, such as indecipherable text, surreal backgrounds, smeared hair, and weird teeth.
Bad actors generally use deepfakes to trick people, but there are instances where the technology can be used for good. Case in point: In China, deepfaking your dead loved ones has become a growing industry. Read here about how people are using the technology to keep their loved ones alive via avatars that use the preserved voices and shared family memories.
Deepfakes are synthetic AI creations that can seem very realistic. They’re so realistic, in fact, that their function is generally to trick someone—whether by, say, attempting to embarrass an individual or by disseminating disinformation within the context of an election. This podcast takes a deep look at the dangers of deepfakes, and how to fight them.
Deepfaked versions of humans that look, act, and sound indistinguishable from the real thing are no longer relegated to science fiction. Case in point: LinkedIn cofounder Reid Hoffman recently created an AI doppelgänger of himself that was fed a sample of his voice and trained on his books and other writings and speeches. See if you can spot which Reid is the real one in this video.
Tom Cruise isn’t on TikTok. But you might think that he actually does have an account on the service if you came across the TikTok deepfakes of the actor that have gone viral. They’re the work of a visual and AI effects artist named Chris Umé and a Cruise stand-in, actor Miles Fisher, who already bears a vague resemblance to the actor. Watch what happens after Cruise’s face is superimposed onto Fisher’s.
Explore all Deepfakes
Search and uncover even more interesting information in our vast database of curated Deepfakes resources
