How to check if a photo was AI-generated: 6 methods that actually work

Published 20 August 2026 · 8 min read

You've seen the photo. It's shocking, perfectly timed, almost too good. Before you share it — or publish it — you want to know: is it real?

The uncomfortable truth is that no single method gives a guaranteed answer. Modern generators produce images that fool the eye, and social platforms strip the metadata that would tell the truth. That's why professionals combine several methods — and why honest tools say "unverifiable" instead of guessing. Here are six methods that actually work, from quickest to most thorough.

1. Look for visual artifacts (weakest, but instant)

AI generators still struggle with certain details. Zoom in and check:

• Hands and teeth — extra fingers, fused digits, asymmetric earrings, blurred teeth.
• Text — signs, logos and captions in AI images are often gibberish or subtly misspelled.
• Reflections and shadows — mirrors that don't match, shadows falling in two directions, glasses reflecting nothing.
• Backgrounds — repeating patterns, warped architecture, people melting into crowds.

Treat this as a first filter, not proof. The latest models pass the eye test easily — and a real photo can look "off" for innocent reasons.

2. Check the metadata (EXIF)

Every photo straight from a camera or phone carries EXIF metadata: camera make and model, lens, exposure settings, sometimes GPS and timestamp. Open the file's properties or use an EXIF viewer and ask:

• Does it name a real camera (Canon, Nikon, iPhone…)? That's a point for authenticity.
• Does it mention editing software like Photoshop? The image was processed — not necessarily maliciously, but it was.
• Is metadata completely absent? Common for screenshots and social media downloads — inconclusive, not damning.

Key rule: missing metadata never proves manipulation. Most platforms strip it automatically.

3. Check C2PA content credentials

C2PA (Content Authenticity Initiative) is the emerging industry standard: a cryptographically signed manifest embedded in the file, recording its history — which camera or AI tool created it, and what edits followed. Adobe Firefly, for example, signs its output as AI-generated with the trainedAlgorithmicMedia flag.

A valid C2PA signature is the strongest evidence you can get — in both directions. It can prove an image came from a specific camera, or that it was AI-generated. But remember: the absence of a signature means nothing, because most real photos don't have one yet.

4. Run a reverse image search

Google Lens, TinEye or Yandex Images can find where else the photo appeared. This answers a different but equally important question: context. An authentic photo presented with a false claim — an old conflict photo recycled as today's news — is one of the most common forms of visual disinformation. If the "breaking news" image first appeared in 2019, you have your answer.

5. Look at compression artifacts (ELA)

Error Level Analysis exploits how JPEG compression works. When a region of an image is edited and re-saved, its compression level differs from the rest. ELA visualizes these differences: spliced objects, retouched faces and added text light up as brighter patches. It's a genuine forensic technique — but interpreting the map takes practice, and it works poorly on images that were saved many times.

6. Use a verification tool that combines all of the above

Doing all of this manually takes 15–30 minutes per image. Media Verify runs the key layers automatically and in parallel:

1. C2PA provenance — reads and validates content credentials; a signed AI declaration is treated as hard evidence.
2. Metadata forensics — camera data, editing software, stripped or inconsistent EXIF.
3. ELA — compression anomaly map, scored automatically.
4. AI classifier — a machine-learning estimate of synthetic generation.

Instead of a black-box percentage, you get a verdict (authentic / AI-generated / manipulated / unverifiable), a confidence level and the per-layer evidence breakdown — so you can defend the conclusion to an editor or a client.

Common mistakes to avoid

• "No metadata = fake." Wrong — screenshots and messengers strip metadata from real photos every day.
• "The detector said 80% AI, case closed." Single-model detectors have significant false-positive rates, especially on edited real photos. Demand evidence, not just a score.
• "It looks real, so it is." Current generators are past that point. Visual inspection alone is no longer enough for high-stakes decisions.

FAQ