Published: June 7, 2026
Read time: ~11 min
Tags: AI, Technology, Safety
Slug: deepfake-detection
In 2023, spotting an AI-generated face was straightforward — extra fingers, melting ears, that glassy thousand-yard stare. In 2026, it is not. The same tools that required a research lab two years ago now run in a browser tab, and the output is convincing enough that professional fact-checkers are routinely fooled on first inspection.
This guide covers what to look for, which free tools are worth using, and the behavioral signals that often catch fake accounts even when the technical tells do not. No single method is definitive — but combining visual inspection, detection tools, and behavioral analysis gives you a reliable filter.
Quick Visual Checklist
Zoom into the photo and look for:
- Ears: blurred, asymmetrical, or missing detail
- Teeth: uniform, blurred together, or artifacted
- Hair edges: unnaturally smooth or melting into background
- Eyes: mismatched catchlights or unnaturally symmetrical irises
- Background text: garbled, illegible, or missing
- Glasses: frames that intersect the face or show wrong reflections
- Skin texture: too smooth or repetitively patterned at full zoom
- No EXIF metadata in the image file
01 — What Deepfakes Actually Are and How They Are Made
The word deepfake covers two distinct types of synthetic media that are often confused. The first is AI-generated faces: images of people who do not exist, produced by generative models like GANs or diffusion models. The second is face-swaps: real video or photos with one person's face replaced by another's. Both are now trivially easy to create.
AI-generated faces (GAN / diffusion) — Models like Stable Diffusion and Midjourney generate photorealistic faces of people who have never existed. These are the profile photos most commonly used in fake social media accounts, romance scam profiles, and disinformation campaigns. The face is original — there is no real person to search for — which is why traditional reverse image search often fails to catch them.
Face-swap deepfakes — A real person's face is mapped onto another person's body or video. Most commonly used to create non-consensual explicit content, but also used in fraud — swapping a trusted person's face onto a video call to impersonate them.
Why detection is getting harder — Each generation of detection tools triggers improvements in generation models. Visual inspection combined with behavioral analysis gives a more durable signal than any single tool.
02 — Visual Tells in AI-Generated Photos
No generation model is perfect. Every synthetic image leaves artifacts — subtle inconsistencies that trained eyes can learn to spot.
Ears and earrings — Ears are structurally complex and asymmetrical. AI models consistently struggle with them — producing ears that are blurred, missing lobes, asymmetrical in unusual ways, or where earrings differ between sides. Zoom into the ears first.
Hair at the edges — Where hair meets a background, AI-generated images often show an unnaturally smooth or blended transition. Individual strands are either too perfect or incorrectly blended into the background.
Teeth — Teeth in AI portraits are frequently blurred into a single mass, unnaturally uniform in shape and size, or feature boundary artifacts where the gumline meets the lip.
Eyes and catchlights — AI faces often have unnaturally symmetrical pupils and irises. Catchlights are frequently mismatched between left and right eye, or are geometrically impossible for the lighting shown.
Background consistency — Backgrounds in AI portraits often feature repeating or warped elements — a bookshelf where books have no titles, a window reflection that does not match the room.
Skin texture at scale — At 100% zoom, AI skin is often either too smooth or features repetitive texture patterns. Compare forehead texture to cheek texture: real skin varies.
Accessories and text — Glasses frames frequently show inconsistent reflections or pass through the face. Any text in the background is almost always garbled or illegible. This is one of the most reliable tells in current models.
03 — Visual Tells in Deepfake Video
Video deepfakes are harder to sustain than still images — inconsistencies that can be hidden in a single frame become visible over time.
Blinking rate and pattern — Watch for blinks over a 30-second window. The pattern is often wrong — too regular, too infrequent, or the blink is incomplete. Natural blink rate is 15–20 per minute.
Face boundary flicker — Where the synthesised face meets the real neck, ears, or hair, look for a subtle flicker or shimmer — particularly in motion.
Lip sync under stress — Lip movements in deepfakes degrade under fast speech. Pause the video during fast sentences and check whether lip positions match the sounds produced.
Lighting direction shifts — When the subject moves their head, the lighting on a real face shifts naturally. Deepfake faces sometimes retain a lighting profile that does not change appropriately with head rotation.
Physiological signals — Real faces show subtle colour variation in the forehead and cheeks tied to pulse. This is the basis for several academic detection approaches and some automated tools.
04 — Free Detection Tools Worth Using
No detection tool is definitive. Use them as a second opinion after visual inspection, not as a first-line filter.
Hive Moderation (hivemoderation.com/deepfake-detection) — One of the most widely tested free tools. Upload an image or video and receive a probability score. Performs well on GAN-generated faces and current diffusion model output.
FotoForensics (fotoforensics.com) — Shows compression artifacts across an image. AI-generated faces often show uniform ELA levels where a real edited photo would show inconsistency.
Illuminarty (illuminarty.ai) — Trained specifically on Stable Diffusion and Midjourney output. Highlights which regions of an image it considers synthetic.
Reverse face search — If the photo is not AI-generated but a stolen real image, a face search engine will find the original. Upload to FaceSift (facesift.com) to check whether the face appears elsewhere under a different name.
EXIF metadata inspection — Real photos contain EXIF metadata: device model, GPS, timestamp. AI-generated images contain none. Use Jeffrey's Exif Viewer or Exiftool. Absence of EXIF on a photo that claims to be candid is a strong signal.
05 — Behavioral Tells — How Fake Accounts Act
Single or very few photos — A real person's account accumulates photos over time — varying angles, lighting, locations, and ages. A profile with one perfect headshot suggests generated or carefully selected stolen images.
No tagged photos from others — Real people appear in other people's photos. An account where the subject only ever appears in their own uploads has no genuine social graph.
Account creation date vs. activity — Scam accounts are often recently created or show a gap — created years ago but inactive until recently.
Inconsistency between photos and claimed backstory — AI-generated faces are often idealised — younger-looking, more symmetrical, and without environment-specific context.
Urgency to move off-platform — Deepfake video calls are computationally intensive and imperfect under scrutiny. Accounts using AI photos will often resist live video — claiming poor connection, camera damage, or work restrictions.
Protecting Your Own Face From Being Used
- Limit publicly accessible high-resolution photos — Face-swap models perform better with more training data. Keeping your highest-quality photos private removes the easiest source material.
- Use different photos across platforms — The same photo on LinkedIn, Instagram, and a dating app gives anyone a ready-made multi-angle dataset.
- Run a periodic face search on yourself — Check whether your face has appeared on sites you have not published it to. FaceSift scans for face matches rather than exact image copies.
- Watermark photos published for professional use — Semi-transparent watermarks at the corner are easily cropped; placing them across the subject's face is more effective.
- Know the reporting path — Document first (screenshot, URL, date), then report via the platform's impersonation or synthetic media policy. For non-consensual intimate deepfakes, contact the Cyber Civil Rights Initiative (cybercivilrights.org).
Detection Checklist — Run Through These in Order
- Zoom into ears, teeth, hair edges, and eyes at 100% — look for the visual tells above
- Check any text in the background for legibility
- Inspect EXIF metadata — AI images have none
- Upload to Hive Moderation for a probability score
- Run through FotoForensics for error-level analysis
- Run the face through FaceSift — if it appears elsewhere under a different name, the photo is stolen
- Check account creation date and whether the subject appears in other people's photos
- Request a live video call — insistence on avoiding one is a strong behavioral flag
Top comments (0)