How to Spot Deepfakes: A Practical Guide

Published 4 July 2026 · 7 min read · By Richard Brereton

Deepfakes — synthetic media generated or manipulated using AI — have moved from a research curiosity to a mainstream threat. Politicians appear to say things they never said. Celebrities appear in videos they never filmed. Profile photos on LinkedIn, dating apps, and social media are generated by AI and represent people who don't exist.

Detecting deepfakes is getting harder as generation quality improves. But the techniques for identifying them — visual inspection, forensic analysis, metadata checking, and AI detection — still catch the majority of synthetic media in circulation.

Visual Tells: What to Look for

Face and skin

• Unnatural skin texture — too smooth, no pores, waxy appearance
• Inconsistent lighting — face lit differently from the background or neck
• Blurring around the face boundary — particularly around hair edges
• Asymmetry — faces are naturally slightly asymmetrical; deepfakes often over-correct to perfect symmetry
• Teeth — blurred, unrealistic, or inconsistently sized
• Eyes — unnatural gloss, inconsistent reflections, or eyes that don't track correctly in video

In videos specifically

• Unnatural blinking — early deepfakes rarely blinked; newer ones blink but at unnatural rates
• Head movement — jerky or unnatural movement when the person turns their head
• Audio sync — lips not quite matching the audio, or audio quality inconsistent with background
• Compression artefacts — square pixelation patterns appearing around the face area

Forensic Detection Methods

Error Level Analysis (ELA)

ELA re-compresses a JPEG at a known quality level and maps where the error levels are inconsistent. In genuine photographs, error levels tend to be uniform. Manipulated or synthetic images show bright patches where different compression histories meet. ELA is most useful for detecting composited or edited images.

AI Detection Models

Dedicated AI classifiers are trained on large datasets of real and synthetic images. They identify statistical patterns in pixel data that aren't visible to the human eye — generation artefacts, frequency patterns, and compression signatures left by AI models. These are significantly more reliable than visual inspection for modern deepfakes.

C2PA Content Credentials

C2PA is an open standard that embeds a cryptographically signed manifest into image and video files recording how they were created. Images from C2PA-aware cameras like Leica or platforms like Adobe Firefly carry a verifiable creation history. A valid C2PA manifest is the strongest positive proof of authenticity — though its absence doesn't prove a fake.

How to Check an Image Using Free Tools

Mutant Verify combines Hive AI detection with local ELA analysis and C2PA credential checking in a single free tool. Drag and drop any image to get a probability score for AI generation, an ELA forensic overlay, and a check for C2PA metadata.

Context Checks That Help

Beyond technical detection, context is often the fastest tell:

• Reverse image search — run the image through Google Images or TinEye. If it appears in an AI art gallery or stock image site, it's likely synthetic.
• Source verification — where did the image or video first appear? Who published it and when?
• Request the original — genuine photographs have RAW files with EXIF data (camera make, model, settings, GPS). AI-generated images don't.
• Check the EXIF data — images from real cameras include metadata. Most AI-generated images have none, or have metadata that doesn't match a real camera model.