Deepfakes

Deepfakes are images, videos, or audio recordings in which deep learning (DL) replaces or fabricates a real person's face, voice, or body so convincingly that viewers take it as authentic. The term, a blend of "deep learning" and "fake", dates to 2017, when a Reddit user named "deepfakes" began posting face-swapped videos made with open-source tools. Deepfakes are a form of generative AI, and the same techniques drive film effects and dubbing, as well as fraud, harassment, and disinformation.

How Deepfakes Are Made#

The classic face swap uses a pair of autoencoders that share one encoder. The shared encoder learns to compress any face into a compact latent space code describing pose, expression, and lighting, while each decoder learns to rebuild one specific person from that code. After training on thousands of face crops of both people, the system encodes a frame of person A and decodes it with person B's decoder, producing person B's face wearing person A's expression:

Other methods followed:

• Generative adversarial networks (GANs): a generator creates images while a discriminator tries to tell them from real photos, and the contest pushes the generator toward photorealism. GANs such as StyleGAN produce faces of people who do not exist.
• Diffusion models: generate images and video by removing noise step by step, guided by text, a reference photo, or a source video. They drive current photo-editing and text-to-video tools, which can generate a realistic video of a person speaking from a single reference photo.
• Voice cloning: speech models reproduce a person's voice from a short recording, and lip-sync models then reshape the mouth in an existing video to match the new audio.

Uses and Harms#

Deepfake techniques have legitimate uses. Studios de-age actors and dub films into other languages with matching lip movements. The 2020 documentary Welcome to Chechnya replaced the faces of persecuted interviewees with those of volunteers, protecting their identities while preserving every expression. Patients who lose their voice to disease can bank recordings and speak through a clone of it, and generated faces serve as synthetic data for training models without collecting real people's photos.

The harms are larger in scale. Most deepfakes online are non-consensual intimate imagery targeting real people, usually women. Fraudsters clone the voices of relatives or executives to request urgent payments, and in early 2024 an employee of the engineering firm Arup transferred about US$25 million after a video call in which the company's chief financial officer and other colleagues were all deepfakes. Fabricated clips of politicians spread during elections, and the existence of convincing fakes lets people dismiss genuine recordings as fabricated.

Detecting Deepfakes#

Detection systems look for traces that generation leaves behind: blending seams around the face, mismatched lighting and reflections in the eyes, unnatural blinking or head motion, frequency artifacts invisible to the eye, and lip movements that drift out of sync with speech. Most detectors first locate and crop each face, then pass the crop to a classifier trained on real and manipulated examples. Ultralytics YOLO26 fits this second stage: its classification models train on a folder-per-class dataset, so "real" and "fake" folders of face crops are enough to train a real-versus-fake classifier.

Detectors struggle to generalize to generation methods they never saw. In Meta's Deepfake Detection Challenge, the winning model scored 65.18% average precision on a hidden test set of new videos, against 82.56% for the top model on the public test set. Because detection lags generation, the industry also verifies provenance: Content Credentials (C2PA) attach a cryptographically signed record of how media was captured and edited, and AI watermarking embeds a hidden signal in generated content that detectors can read later.

Deepfakes vs. Related Concepts#

• Synthetic data: any artificially generated data used to train or test models, from simulated driving scenes to rendered product images. Deepfake techniques are one source of it when generated faces fill a training set.
• Computer-generated imagery (CGI): digital characters and effects that artists model, rig, and animate by hand. Visual effects studios increasingly combine CGI with learned face replacement in the same shot.
• Shallowfakes: misleading edits made without AI, such as slowing a clip to make a speaker sound impaired or captioning a real video with false context.

Regulation#

Laws now target deepfakes directly. Article 50 of the EU AI Act, which applies from August 2, 2026, requires deployers who use AI to create deepfakes to disclose that the content is artificially generated or manipulated, and requires providers of generative systems to mark their output in a machine-readable format. In the United States, the TAKE IT DOWN Act, signed in May 2025, made publishing non-consensual intimate deepfakes a federal crime, and since May 19, 2026, covered platforms have been required to remove such content within 48 hours of a valid request.