How to detect AI-generated images
last updated 7 june 2026
To detect an AI-generated image, check three things in order: visual tells (hands, teeth, text, warped backgrounds, too-perfect lighting), embedded provenance (C2PA Content Credentials and SynthID watermarks), and a detector's confidence score. No single signal is proof — combine them.
AI image generators like Midjourney, DALL·E and Stable Diffusion now produce photorealistic results, so the old advice of "just look at the hands" is no longer enough on its own. The reliable approach is to layer three independent checks — your own eyes, the file's provenance metadata, and an automated detector — and treat agreement between them as your signal.
verifai automates the provenance check for you, in the browser: click scan and it reads the content credentials embedded in each image and flags the ones credentialed as AI, right on the page — no right-click, no upload. For text, it adds a transparent on-device heuristic score.
Spot an AI-generated image in four steps
- 1
Scan the obvious anatomy and text
Zoom in on hands, fingers, teeth, ears, eyes and any text in the image. Generators still struggle with finger counts, symmetrical jewellery, and legible signage — garbled text on a sign or storefront is one of the strongest tells.
- 2
Check lighting, reflections and backgrounds
AI images often have lighting that doesn't match across the scene, reflections that don't correspond to objects, and backgrounds that melt or repeat. Look for blurred or nonsensical detail away from the focal point.
- 3
Inspect provenance metadata
DALL·E (OpenAI) and Adobe Firefly attach C2PA Content Credentials, and Google's Imagen/Gemini images carry an invisible SynthID watermark. Verifying Content Credentials confirms origin — but absence proves nothing, since screenshots and re-saves strip metadata.
- 4
Run a detector for a confidence score
Use an automated check as a tie-breaker. verifai reads each image's embedded content credentials and flags those marked as AI; text gets a 0–100 on-device heuristic score. Images with no credentials are shown as "unverified" — not a clean bill of health.
The most reliable visual tells in 2026
As models improve, the tells move from obvious to subtle. These are the ones still worth checking first:
- hands and fingers — wrong counts, fused or bent digits
- text — menus, signage and logos that look right but read as gibberish
- symmetry — mismatched earrings, eyes, or background architecture
- skin and texture — waxy, over-smoothed, or uncanny perfection
- physics — shadows, reflections and scale that don't agree
Why metadata alone isn't enough
Content Credentials (C2PA) and SynthID are the future of provenance, but adoption is partial and metadata is fragile. A genuine AI image loses its credentials the moment someone screenshots it or a platform re-encodes it on upload. So a missing watermark is not evidence the image is real — it's just missing information. That's why combining checks beats trusting any one of them.
frequently asked
Can you detect AI images with 100% accuracy?
No. No method — visual inspection, metadata, or automated detector — is 100% accurate, and accuracy drops as generators improve. The practical goal is a confidence level from multiple independent signals, not certainty.
Does verifai detect Midjourney and DALL·E images?
verifai reads the content credentials embedded in images, so it reliably flags credentialed output from tools like DALL·E, Firefly and Gemini. Midjourney embeds none, and screenshots strip them — so those images show as "unverified" rather than "human." A hosted deep-scan to catch uncredentialed images is planned.