How AI Deepfakes Work—and Why They're So Hard to Stop
AI-generated deepfakes can swap faces, fabricate voices, and create nonconsensual intimate images of real people in minutes. Here's how the technology works, why detection remains an arms race, and what new laws are trying to do about it.
What Is a Deepfake?
A deepfake is a photo, video, or audio clip that has been fabricated or manipulated using artificial intelligence to make it appear real. The term blends "deep learning" — the AI technique behind the technology — with "fake." While early deepfakes required significant computing power and expertise, today anyone with a smartphone can generate a convincing fake image in seconds using freely available apps.
The technology has legitimate creative uses — from restoring aged film footage to dubbing actors in different languages. But it has also become a powerful tool for fraud, political manipulation, and sexual abuse.
How Deepfakes Are Made
Most deepfakes rely on one of two core AI architectures: Generative Adversarial Networks (GANs) or diffusion models.
A GAN pits two neural networks against each other. The generator creates fake images; the discriminator tries to spot them. Over thousands of training cycles, the generator learns to produce images convincing enough to fool its opponent — and, by extension, human eyes. Diffusion models, which power tools like Stable Diffusion and DALL-E, work differently: they learn to gradually remove random noise from an image until a realistic picture emerges, guided by a text or image prompt.
For face-swap deepfakes, the AI is trained on hundreds or thousands of images of a target person — often scraped from social media. It then maps their facial geometry onto a different body or replaces another person's face in existing footage. So-called "undressing apps" take a clothed photo of a real person and generate a nude image of them without their knowledge or consent.
The Scale of the Problem
The numbers are staggering. According to data compiled by security researchers, AI-generated child sexual abuse material reports rose 1,325% between 2023 and 2024, with over 67,000 reports filed to the National Center for Missing and Exploited Children in 2024 alone — up from just 4,700 the previous year. A study by the Center for Countering Digital Hate found that one AI image generator produced roughly three million sexualized images in under two weeks, including tens of thousands depicting minors.
The harms extend far beyond sexual abuse. Deepfake financial fraud surged tenfold in 2024, costing North American businesses over $200 million in a single quarter. Scammers can clone a person's voice from as little as three seconds of audio and use it to impersonate executives, parents, or government officials. A 2025 study found that only 0.1% of people could correctly identify all fake and real media shown to them in a test.
Why Detection Is So Difficult
Detection tools — which use AI to look for telltale signs of manipulation, such as unnatural blinking, lighting mismatches, or skin texture anomalies — are locked in a constant arms race with deepfake creators. As detectors improve, generators are retrained to defeat them. Researchers at the Alan Turing Institute warn that detection accuracy drops sharply when the deepfake is created using a different method than the one used to train the detector.
One promising approach is content provenance: embedding cryptographic metadata into images at the moment of creation, so viewers can verify whether a photo was taken by a camera or generated by AI. The Coalition for Content Provenance and Authenticity (C2PA), backed by Adobe, Microsoft, and major camera makers, is building a global standard for this. Google's SynthID invisibly watermarks AI-generated images. But watermarks can be stripped, metadata can be erased, and screenshots bypass the system entirely.
What the Law Says
Legislation is catching up — slowly. The US TAKE IT DOWN Act, signed into law in May 2025, criminalized the nonconsensual publication of intimate images, including AI-generated fakes, and requires platforms to remove flagged content within 48 hours. The DEFIANCE Act, passed by the Senate in January 2026, creates a federal civil right of action allowing victims to sue creators and distributors of nonconsensual intimate deepfakes for up to $150,000 in damages — or $250,000 if the abuse is linked to stalking or assault.
Enforcement remains patchy. Many deepfake apps operate across jurisdictions, platforms struggle to keep pace with the volume of uploads, and anonymous creators are hard to trace. Civil society groups argue that takedown timelines are still too slow, and that tech companies must do more to prevent generation in the first place — not just react after harm is done.
What You Can Do
Awareness is the first line of defense. Experts recommend limiting publicly available photos on social media, using reverse image search to check whether your photos have been misused, and reporting nonconsensual intimate imagery to platforms and law enforcement. Organizations like StopNCII help victims create digital hashes of images so platforms can proactively block their spread before they go viral.
The deeper challenge is structural: as AI image generation becomes cheaper and more accessible, the gap between what is real and what is fabricated will continue to narrow. Solving it will require not just better technology, but stronger legal frameworks, platform accountability, and a cultural shift in how synthetic media is treated online.
