Digital Forensics
When your eyes fail you, use tools. Digital files carry invisible fingerprints that can reveal their synthetic origins.
Context Verification
The most powerful tool isn't a complex algorithm—it's Reverse Image Search. When a controversial image appears (e.g., "Explosion at the Pentagon"), verify it immediately.
The "Context Void" Theory
Real major events generate thousands of photos from hundreds of angles. Deepfakes usually start as a single image with no corroborating angles. If only one picture exists of a massive event, it is almost certainly fake.
- Google LensGreat for finding the exact source or debunking articles.
- TinEyeExcellent for finding the oldest version of an image.
Metadata & EXIF
Real photographs contain EXIF data: Camera Model, Shutter Speed, ISO, Date Taken. AI images are born in a void.
However, be warned: Social Media strips EXIF data. A photo on Twitter will have no metadata whether it's real or fake. This technique only works on original files (sent via email, Discord, or direct download).
Typical AI Header Tags
Creator: "Flux.1 Dev"
Parameters: --ar 16:9 --steps 30
Software: "Adobe Firefly"Error Level Analysis (ELA)
When a JPEG is saved, it compresses the image. If you photoshop an object into an image, that new object will have a different "compression level" than the background because it was saved at a different time.
ELA Tools visualize this compression. In a real photo, the noise should be uniform. In a manipulated photo, the fake object will glow brightly (or be pitch black) compared to everything else.
Note: Fully AI-generated images (not edits) often pass ELA because the entire image was generated at once, creating uniform noise. ELA is best for spotting "bad photoshop" edits rather than full AI generations.
Frequency Fingerprints
Images are made of pixels, but mathematically they're also made of waves (frequencies). Camera sensors capture those waves naturally. Diffusion models build them from scratch by up-scaling a small noise map—and that process leaves an invisible, grid-like fingerprint.
You can't see it with your eyes, but a Fast Fourier Transform (FFT) heatmap often reveals a bright cross or star pattern at the center of AI images—the signature of the up-sampling grid. This is how automated detectors can flag an image that looks flawless to a human: they aren't looking at the picture, they're looking at the math.
Provenance: C2PA & Watermarks
Everything so far has been about catching fakes after the fact. Provenance standards flip the problem: instead of proving an image is fake, they let real images prove they're real.
C2PA / Content Credentials
A digital passport for images, backed by Adobe, Microsoft, and camera makers. A supporting camera cryptographically signs the photo at capture; every edit adds to the signed history. Look for the "Cr" pin on supporting sites, or check any file at verify.contentauthenticity.org. A broken or missing chain doesn't prove fakery—adoption is still growing—but an intact one is strong evidence of authenticity.
Invisible Watermarks (SynthID)
The supply-side counterpart: generators like Google's Imagen embed an invisible pixel-level watermark into everything they produce. It survives cropping, rotation, and recompression, and detector tools can read it even though you can't. Visible watermarks can be cropped out; these can't.
Field Checklist
Before you guess, run through these checks:
- 1Reverse-search the image first: a sensational photo that exists in exactly one place is almost certainly fake.
- 2Check whether the earliest appearance was in an AI art community rather than a news source.
- 3Inspect EXIF metadata on original files—look for generator tags like "Flux.1" or "Adobe Firefly", or a suspiciously empty profile.
- 4Remember that social media strips metadata: absence of EXIF on a downloaded post proves nothing.
- 5Run ELA on suspected composites, but not on suspected full generations—those pass with uniform noise.
- 6Look for Content Credentials (the "Cr" pin) and verify signed files at verify.contentauthenticity.org.