Generative image models from 2023 to today have closed most of the obvious gaps. The cliché “just count the fingers” still works often enough to be useful, but it is no longer sufficient on its own. A 2024 NIST study on generative media notes that detection accuracy of even leading commercial tools degrades sharply on freshly-trained models. The honest answer to “is this image AI?” is rarely a single number — it is a weighing of signals.
The approach below combines three layers: human visual inspection, automated detection tools, and provenance via reverse image search. Use all three before you trust a photo, especially one driving an emotional reaction.
Layer 1 — Visual signs to check first
Zoom in to 100% on a desktop. AI artifacts collapse at thumbnail size and reveal themselves at full resolution.
Hands and fingers
Six fingers, fused thumbs, hands gripping objects with no contact points. Still the most common giveaway in diffusion-model output.
Eyes and pupils
Asymmetric pupils, catchlights pointing in different directions, irises that drift outside the eye shape on close inspection.
Plastic or waxy skin
Skin uniformly smooth across cheek, forehead, and neck — real skin has pores, blemishes, and variation under different light angles.
Asymmetric jewelry and accessories
One earring, mismatched earrings, glasses that bend through the face, necklaces that fade into clothing.
Background text and signage
Text in the background is the most reliable single tell. AI models produce text that looks like writing but spells nothing real on close read.
Repeating patterns and crowds
Same face replicated across a crowd, identical tiles in a floor pattern, fence posts that merge into one another.
Layer 2 — Free AI image detectors compared
No detector is a silver bullet. Run an image through two of these and weight the answers; agreement matters more than any single score.
| Tool | Type | Strength |
|---|---|---|
| Sightengine sightengine.com | Commercial API + free web demo | Strong on photorealistic portraits; trained on multiple generator families. |
| WasItAI wasitai.com | Free web tool | Fast, simple yes/no probability. Best as a first-pass sanity check. |
| DeepfakeDetection.io deepfakedetection.io | Free web tool | Focuses on face-manipulation detection (face swaps, GAN portraits). |
| Hive AI Detector hivemoderation.com | Free browser extension + API | Used by many platforms for content-moderation pipelines. |
| Google Lens (reverse search) lens.google.com | Free | Doesn’t detect AI directly, but finds the earliest known version of an image. |
Layer 3 — Reverse image search
Provenance often beats detection. If a “just-captured” news photo already exists on a stock site from 2021, the question of AI vs. real becomes moot — it’s recycled regardless. Run every suspicious image through Google Lens, TinEye, and Yandex Images. Yandex in particular is unusually strong on faces. Bellingcat’s open-source investigation handbook treats reverse search as step one for any image-based claim.
A note on false positives
Heavily edited real photos (skin smoothing filters, JPEG compression artifacts, beauty-mode portraits) sometimes trigger AI detectors. Treat a single “likely AI” score as a flag for further checking, not a verdict. A real photo with a verifiable photographer and clean reverse-search history is more trustworthy than a detector score in either direction.
Track confirmed AI fakes in circulation
FAXTR’s AI Fakes feed catalogs viral AI-generated images already debunked by fact-checkers worldwide.