How to spot an AI-generated image

This is the visual layer — the things your eyes CAN see with careful inspection. These tells complement SafPix's automated signals (metadata, noise patterns, compression) which reveal what's invisible to the naked eye.

None of these alone is proof. But when two or three show up together, you're almost certainly looking at a synthetic image.

01 Hands, ears, and readable text

This is the classic. Generative models have always struggled with the fine structural details of hands (fingers bending the wrong way, six fingers, fingers fused), ears (asymmetric, warped, or merged with hair), and any text that appears in the image (jumbled characters, nonsensical signage, garbled logos).

New-generation models have improved dramatically — Flux and the 2026 Midjourney releases produce hands that look correct at a glance. But under scrutiny, you'll still find issues: fingernails that blur into adjacent fingers, a pinkie at an impossible angle, a watch face with characters that aren't quite numbers. Text on shirts, signs, or menus in AI images almost always degrades into pseudo-letters on close inspection.

Zoom in on hands. Zoom in on any text. If you see something that doesn't resolve into a real word or a plausible finger, that's your first signal.

02 Skin that's unnaturally smooth

Real camera sensors capture micro-texture at the pixel level: skin pores, peach fuzz, tiny hair strands, the grain of fabric, pencil-stroke imperfections. When you look at a real phone photo at 100% zoom, you see stuff in every pixel. It's not flat.

AI-generated faces — especially on the first-pass output from a generator — often have perfectly flat skin that looks like it's been heavily retouched. No pores. No faint shadows from fine hair. No micro-variation. It's as if the entire face is one smooth surface painted on.

The opposite failure also happens: some generators over-compensate by adding exaggerated texture that looks like low-grade film grain pasted onto the image.

• Real photo skin: irregular pores, faint peach fuzz, tiny asymmetric blemishes, hair strands that pass in front of the face
• AI skin: porcelain smoothness, symmetric pseudo-pores that repeat, weird uniform lighting with no micro-shadow

03 Backgrounds that blur into nonsense

Modern AI generators put enormous compute into the foreground subject — the face, the product, the central object — and comparatively less into the background. That means backgrounds are where you'll find the most revealing failures.

Look for: doorframes that bend at impossible angles, two objects fusing into each other (a coffee cup growing out of a chair), geometry that changes direction mid-wall (a window frame that starts square and ends tilted), shadows that don't match the light source, and "nonsense writing" on background signs or posters that looks like letters but isn't.

In real photos, backgrounds are boring but internally consistent — a chair is a chair, a wall is a wall, a sign reads what it reads. In AI photos, backgrounds are where the model gets tired.

When you suspect AI, stop looking at the subject and stare at everything behind it. That's where the cracks show.

04 Missing, stripped, or AI-branded metadata

Every time a real camera takes a photo, it embeds metadata into the file: camera make, model, the exact date and time, GPS coordinates, exposure settings, software version. This metadata is called EXIF, and it's the fingerprint of a legitimately-captured photo.

AI-generated images don't have camera EXIF because they weren't captured. Instead, they either:

• Have no EXIF at all — the metadata block is empty or absent, which is suspicious for any image that claims to be a personal photo
• Have AI-generator EXIF — the Software field says Midjourney, Stable Diffusion, DALL·E, or another generator, which is effectively a signed confession
• Have editor EXIF — Adobe Photoshop 25.0 indicates the image has been edited (which could be legitimate retouching or AI-to-JPG post-processing to hide generator fingerprints)

You can check EXIF in a few ways. On most desktop operating systems, right-click the image file → Properties or Get Info. On iOS, the Photos app shows camera info. Or — easiest — drop it into SafPix and the Digital Origin signal reads it for you.

Catch: many social platforms (Instagram, WhatsApp, Facebook, Twitter) strip EXIF on upload to protect user privacy. So a photo received via these platforms will typically have no EXIF even if it's 100% real. Missing EXIF is a weak signal on its own — but combined with any of the other four, it strengthens the suspicion.

05 The "one photo only" rule

This is the least technical tell and maybe the most powerful one. AI profiles — dating, marketplace, fake LinkedIn, scam accounts — almost always have a very limited photo set. One great hero image, maybe two. And crucially, they can't give you a new one on demand.

If you're talking to someone and have any doubt, ask: "Can you send me a quick photo with a peace sign right now?" or "Send me a pic of your dinner" or "Selfie with your left hand on your ear." A real person can do this in 10 seconds. A scammer running an AI profile cannot — because generating a specific new image that matches the old ones, on demand, in the right pose, is still beyond most generators' ability to stay consistent across outputs.

This works for images you suspect are stolen (reverse image search finds them elsewhere), images you suspect are AI (can't produce a new one), and images that are heavily filtered (real person but won't drop the filter for a specific request).

A real photo has a real person behind it who can take another one. That fact alone is the deepest forensic signal in the world.

Why these five, and not others?

There are dozens of subtler AI tells — pupil reflection inconsistency, jewelry that "melts" where it should have hard edges, teeth that don't line up, weird hair transitions. But the five above are the ones that work in the real world because:

• They're fast to check — you can evaluate all five in under a minute
• They compound — two or three together is a near-certain verdict
• They survive format changes — a screenshot of an AI image still shows the smooth skin and nonsense backgrounds
• They work on cartoonish AI — even obvious stylized AI art typically has at least hand/text or background issues
• They resist model improvements — each of these is tied to fundamental generator limitations, not a specific architecture

When your eye isn't enough

Some AI images genuinely look real at a casual glance — perfect hands, natural skin, coherent backgrounds. That's where SafPix comes in. 18 forensic signals reveal what your eyes can't see: sensor noise patterns, compression integrity, texture micro-structure, and more.

Think of visual inspection as the first layer and SafPix as the second — the X-ray that reveals hidden data. When the stakes are high — a romantic partner, a business deal, a news story — SafPix Pro adds professional tools: PDF reports for documentation, batch processing for multiple images, and comparison mode to cross-reference sources.

Your eyes catch visible anomalies. SafPix reveals the invisible ones. Together, you see everything.