The word "deepfake" has been doing a lot of work since 2017, most of it badly. It gets attached to anything from a Hollywood face-swap to a poorly captioned screenshot. So before we talk about what deepfakes do, it helps to be precise about what one actually is, because the precise version is less cinematic and more useful than the panic version.
A deepfake is media in which a person's face, voice, or whole performance has been synthesized or altered by a machine-learning model to depict something that didn't happen. The category breaks into a few distinct techniques, and the distinctions matter because each fails differently.
The original, and still the most recognizable. The model lifts one person's face and maps it onto another's head in existing footage. This is the "celebrity in a video they never made" version. It's also the easiest to catch, because the seam between the swapped face and the real head (at the jaw, the hairline, in fast motion) is where artifacts pile up.
Subtler and, for that reason, more dangerous. Here the target's real identity stays put; the model only puppets the mouth, expression, or head pose to match a new audio track. Researchers studying detection have noted that lip-synced fakes are harder for people to spot than face swaps, precisely because almost everything on screen is genuine. You're being asked to mistrust one small region of an otherwise real face.
The newest tier: a person (or a convincing approximation of one) generated more or less from scratch, often from a text prompt or a few reference clips. This is where the diffusion-era video models live, and where the old "look for the seam" advice starts to run out, because there's no original footage to seam against.
Two families of models matter. The first wave, roughly 2017 to the early 2020s, ran on generative adversarial networks (GANs) and autoencoders: two neural networks competing, one generating fakes and one trying to catch them, each making the other better. That arms race is why early deepfakes improved so fast.
The current wave runs on diffusion models, the same architecture behind modern image and video generators. They start with noise and refine it into a coherent frame, which produces cleaner, more temporally stable video than GANs managed. This is the practical reason the 2026 tells are weaker than the 2020 tells: the technology that generates the fakes is the same technology being used to generate everything else, and it has gotten good. If you want the specifics of what still holds up on video, that's its own piece; see how to spot AI video. The audio side, which fuels most of the fraud below, has its own playbook in voice-cloning scams.
In early 2024, an employee in the Hong Kong office of the London-based engineering firm Arup transferred about US$25.6 million (200 million Hong Kong dollars) across 15 transactions after joining a video call on which the chief financial officer and several colleagues were all deepfakes. The employee had been suspicious of the initial email; the live video call is what dissolved the doubt. Hong Kong police said the fraudsters had built the fakes from publicly available footage of the real executives, the kind of recorded meetings and conference clips anyone can find online. As of early 2025, the money had not been recovered.
That case is the vivid one, but the broader pattern is steadier. The FBI's Internet Crime Complaint Center added a dedicated "AI-related" section for the first time in its report covering 2025, logging more than 22,000 such complaints and nearly $900 million in losses. Business email compromise (the broader category these executive-impersonation scams fall under) accounted for about $3 billion in reported losses that year on its own. Deepfakes don't need to fool the world. They need to fool one finance employee on one Tuesday.
The most widespread harm, and the least discussed in the "is democracy ending" coverage. A frequently cited 2019 survey by the detection firm Sensity (then Deeptrace) found that 96% of the deepfake videos it catalogued online were non-consensual pornography, and that essentially all of the targets were women. That figure is old now and the denominator has changed as general-purpose video generation exploded; treat the exact percentage as a snapshot, not a current census. But every serious analysis since has found the same shape: this is overwhelmingly where the volume is, and it overwhelmingly targets women and girls.
Here's where honesty cuts against the headlines. Going into the 2024 global election year, the forecast was apocalyptic. The reality was muted. Researchers at the Knight First Amendment Institute examined 78 election-related deepfakes and found that roughly half had no deceptive intent at all: they were satire, commentary, or campaign material that everyone understood to be synthetic. Meta reported that less than 1% of the misinformation its fact-checkers reviewed during the 2024 cycles was AI-generated. And cheap fakes (real footage relabeled or miscaptioned, no AI involved) were used several times more often than deepfakes.
That does not make deepfakes harmless. Political misinformation is mostly a demand problem, not a supply problem. People share what flatters what they already believe, and a slowed-down real clip does that job just as well as a synthesized one, for less effort.
"You won't be able to believe anything you see." The more durable risk isn't mass deception; it's the liar's dividend: the person caught on real video who now gets to claim it was a deepfake. The erosion of trust cuts in the direction of denial as much as belief.
"Detectors will save us." Detection software is a useful signal, but it degrades badly on compressed, re-uploaded clips and produces both false alarms and misses. It's a clue, not a conclusion.
"There's an obvious tell." There used to be: bad hands, dead eyes, garbled text. The leading models cleared most of those. Treat a clean-looking video as unproven, not verified.
The thing that holds up in 2026 is provenance, not a visual trick. Industry standards like C2PA Content Credentials and watermarking systems such as Google's SynthID attach signed, tamper-evident history to media at creation, and since 2 August 2026, the EU AI Act has required machine-readable disclosure on AI-generated content. The catch worth memorizing: a missing credential proves nothing. Plenty of real footage carries no provenance at all. The signal is positive confirmation (a verified source, a second independent outlet, an original you can trace), not the absence of a red flag. For the harder question of whether disclosure even makes synthetic media acceptable, see what counts as real.
So: a deepfake is a specific thing, it does specific damage, and the worst of that damage is financial and personal long before it's political. The defense is a slower eye, not a sharper one, and the habit of asking where a thing came from before you decide what it means.
That habit is, more or less, the whole game we play here every day.
What's the difference between a deepfake and a 'cheap fake'?
A deepfake uses AI to synthesize a face, voice, or whole performance. A cheap fake just edits real footage: slowing it down, splicing it, captioning it dishonestly. During the 2024 elections, cheap fakes were used roughly seven times more often than AI ones. The boring trick still works better than the fancy one.
Can deepfake-detection software just tell me if a video is fake?
Not reliably. Detectors are useful as one signal among several, but they degrade on compressed, re-uploaded clips and they produce both false positives and false negatives. Provenance (a signed C2PA Content Credential, the original source, a second outlet) settles questions that a detector score only gestures at.