detect·deepfakesby Resemble AI
How to spot · image

How to Tell If a Photo Is AI-Generated

A practical guide to spotting AI-generated images in 2026 — what still gives diffusion models away, what doesn't, and when you should stop inspecting and run the detector instead.

Resemble AI··3 min read

In 2022, an AI image gave itself away in about two seconds — too many fingers, garbled text, melted jewelry. In 2026, most of those tells are gone. Frontier diffusion models ship photorealistic output that routinely passes untrained human inspection.

Here's what still works when you need to squint at a photo before trusting it.

The tells that persist in 2026

1. Impossible reflections

Mirrors, windows, sunglasses, polished surfaces. Diffusion models don't do ray-tracing; they approximate reflections based on what neighboring pixels "should" look like. This produces surfaces that reflect geometry that doesn't exist in the rest of the scene.

Check: does the reflection in the window match the room? Does the glint in the sunglasses match the actual light source?

2. Inconsistent shadows

One subject lit from the left, another from the right, in the same frame against the same background. Real photos have one sun. Composite AI images often don't.

3. Skin too smooth, too uniform

Real skin has pores, tiny asymmetries, stray hairs, slight redness in the cheeks, subtle sheen where moisture sits. AI skin tends toward beauty-filter-even. It's the "this looks like a render" feeling — even when you can't specifically name the tell.

4. Jewelry, watches, and earrings

Small-object geometry remains a weak spot. Check:

  • Earrings that don't match each other
  • Watch dials with three hands (one too many)
  • Chain links that don't quite interlock
  • Impossible fastenings

5. Text in the background

Signs, labels, book covers, logos. Real-world text has specific letterforms; AI text often produces plausible-looking but nonsensical characters, or letters that are slightly off from any real font.

6. Hands and fingers

Mostly solved by 2026, but not entirely — especially in complex poses (interlaced fingers, hand holding an object). Check finger count and joint geometry at least once.

The metadata check (before zooming in)

Before looking at pixels, check the file itself:

  • C2PA signature. If present, an image was signed by a known device (a Leica M11, a Sony Alpha) or a known generator (Midjourney, Firefly, OpenAI). A valid C2PA signature from a camera is a strong real signal; a signature from a generator tells you directly what produced it. Absence of metadata isn't evidence of fakery — metadata gets stripped during social-media uploads all the time.
  • EXIF data. Right-click → Properties on Windows, or look up the file in an EXIF viewer. Real camera photos usually carry camera model, aperture, exposure time. Absence is a weak signal.
  • File compression. Real photos usually have one JPEG save roundtrip. AI-generated images are often saved multiple times before they reach you (generation → service → platform → your device). Double-compression artifacts are detectable but only with tools.

When to reach for the detector

If you've done the inspection above and any of these is true, run the image through an automated detector:

  • The photo is compressed heavily (e.g., a screenshot of a social-media post)
  • You can't find a C2PA signature or EXIF data
  • The subject is unfamiliar and you have no mental baseline
  • The stakes justify a second opinion

Use our free image deepfake detector — it runs the image against a library of known generator fingerprints (Midjourney, DALL·E, Stable Diffusion, FLUX, Imagen, Firefly), inspects for diffusion-decoder tiling patterns, and returns:

  • A verdict + confidence score
  • The specific regions the model flagged
  • The closest generator match and its similarity
  • C2PA / PerTH watermark status

What doesn't work

  • Reverse image search catches previously-seen AI images. It won't catch a novel generation. Useful as a second check, not a first.
  • "It looks weird" is often correct but always unreliable. If you can't articulate the specific tell, you may be picking up on something or you may be wrong. Don't publish based on gut.