AI image generators are good now. Scary good. And every week, fakes slip past millions of people on social media, in ads, and yes, even in the news. You don’t need to be a forensics expert to catch most of them. You just need a checklist and about 30 seconds.
Here’s the practical version, no PhD required.
The visual tells that give it away
You don’t need to squint at pixels. Most AI images break in the same handful of ways. Train your eye to look for these first.
1. Hands, fingers, and teeth
This is the oldest trick in the book, and it’s still reliable more often than not. Count the fingers. Look at the thumbs. Are the knuckles bending the right way? Six fingers, fused digits, or hands that look melted into a wrist are huge red flags. Teeth get the same treatment, too perfect, too uniform, or weirdly blended into the lips.
2. Eyes and catchlights
In a real photo, the tiny white reflections in the eyes (called catchlights) usually match. Same light source, same direction. AI often gets one of them right and the other wrong, or gives someone two pupils that don’t line up. If the eyes feel “off” but you can’t explain why, trust the feeling.
3. Text and signs in the background
AI models treat writing as a texture, not as language. So signs, labels, t-shirt slogans, and book spines come out as garbled nonsense that almost looks right. If you can’t actually read a word in the image, it’s probably fake.
4. Hair, edges, and the “halo” effect
Look at where the subject meets the background. AI often adds a soft, slightly glowing halo around hair, or strands that blend unnaturally into whatever’s behind them. Real photos have crisp, messy edges. AI images look airbrushed by default.
5. Asymmetrical details
Earrings that don’t match. A watch on one wrist and a bracelet on the other that don’t belong to the same outfit. Tattoos that morph halfway through. Patterns on clothing that warp or repeat weirdly. Real life is messy. AI defaults to symmetry even when it shouldn’t.
6. Lighting and shadows
Shadows should match the light source. AI sometimes gives you two shadows going different directions, or a face lit from the front with a shadow that suggests the sun is behind. Squint and check the geometry.
The investigation tools
Visual checks catch most fakes. For the rest, you have a few free weapons.
- Reverse image search: Right-click an image and run it through Google Lens, TinEye, or Yandex. If the same photo has been online for years, you’re done. If you can’t find a match and the image looks too clean, that’s a yellow flag.
- EXIF metadata: Real photos from a camera or phone carry metadata. AI images usually have none, or the metadata is suspicious. Tools like exifdata.com or the right-click “Properties” panel on most OSes will show you what’s there.
- AI detection tools: Hive, AI or Not, and Sensity all offer free checks. They’re not perfect and they get fooled by heavy editing, but they’re another data point.
Provenance, not vibes
For high-stakes verification, the modern approach is provenance over perception. The C2PA (Coalition for Content Provenance and Authenticity) standard and the Adobe-led Content Authenticity Initiative (CAI) embed cryptographic signing into images at capture or at edit time. When a verified C2PA manifest is present and matches a known publisher, the image is far more likely to be authentic. When no manifest is present, that is itself a signal — it does not prove the image is fake, but it removes one trust anchor.
Reverse image search (Google Lens, TinEye, Yandex) is the second most powerful tool. If the same image has been online for years, you are done. If you cannot find it and the metadata is gone, treat that as a yellow flag, not a verdict.
Detector caveats
AI detection tools (Hive, AI or Not, Sensity) are useful as one data point, but they have real false-positive and false-negative rates. As generative models improve, detector accuracy degrades. A negative detector result does not mean the image is real. A positive result does not always mean it is fake. Use detectors only as one signal among several — not as a verdict.
Social networks strip metadata
Facebook, X, Instagram, and most platforms strip EXIF on upload. The absence of metadata from a screenshot is not evidence of AI generation. It is evidence that the image was uploaded to a social network.
Context beats pixels
Most misinformation is real imagery with a misleading caption, or a real photograph presented as evidence of a different event. A real photograph from five years ago can be presented as evidence of a current event. The image is real; the story is wrong. Provenance, original source, and date of capture matter at least as much as the pixels themselves.
A 30-second workflow for 2026
- Look at hands, eyes, background text, and lighting. (10 seconds)
- Check provenance: is there a C2PA manifest? Is the source verifiable? (10 seconds)
- Run a reverse image search. (10 seconds)
That sequence catches the vast majority of fakes and misleadingly-captioned real images. The goal is not to be right every time. The goal is to slow down just enough that the fakers have to try harder.
Sources & Further Reading
All claims in this article are sourced from primary documentation, vendor advisories, and reputable security researchers.
Spotted an error? Email the editor. Corrections are issued with a visible correction note.
Editorial standards. Every article on humanrequired.org is reviewed by a human editor before publication. AI may assist with drafting or research; final editorial control is human. Read the full standards.
