What Do AI Generated Faces Look Like?

What's changed — and what hasn't

Two or three years ago, this guide would have been straightforward. AI-generated faces had consistent, obvious tells: hands with the wrong number of fingers, teeth that fused into a smooth surface, backgrounds with warped geometry near the subject's edges. You could often tell at a glance.

That's no longer the case for modern AI generators. The obvious tells have been largely corrected in recent model versions. What's left is subtler — patterns that require closer inspection and that are sometimes impossible to identify by eye alone.

Researchers at Northwestern University found that people perform close to chance when asked to identify AI-generated faces — meaning visual inspection alone has become an unreliable method. At the same time, detection tools that analyze the pixel-level statistics of images are still quite good at catching modern AI faces, because the mathematical properties of generated images differ from photographed ones even when the visual result is convincing.

This guide covers what your eyes can still catch — and where you need a tool to pick up the slack.

Visual tells that AI faces still produce

Hands and fingers

Still the most reliable tell in images where hands are visible. AI models have improved significantly at generating hands, but they still struggle with complex hand positions — particularly when fingers are interlaced, when a hand is gripping something, or when fingers are partially obscured. Count the fingers when hands appear. Look at the proportions between knuckles and finger length. The contact points where fingers meet an object (a phone, a cup, a steering wheel) are still frequently wrong.

Teeth in open-mouthed smiles

Real teeth have variation. They're slightly different sizes. Some are more prominent than others. There are small imperfections — slight staining, chips, real three-dimensionality. AI-generated teeth are frequently too uniform: identical in shape, too white, and too numerous. The edges where the top and bottom rows of teeth meet can also look slightly wrong — blended rather than distinct. Look closely at the gum line too.

Ear structure

Human ears are complex three-dimensional shapes with significant individual variation. AI models generate them plausibly from a distance but tend to flatten or simplify the inner ear — the helix, antihelix, and tragus often look smoother than a real ear at close inspection. The back edge of the ear, where it meets the skull, is frequently slightly wrong. On profile-view photos, this is one of the easier tells.

Hair at the edges

Look at where the hair meets the background — especially in photos with complex backgrounds. Real hair has individual strands that interact with real light and behave in front of a real backdrop. AI-generated hair often has an unnaturally smooth transition at the edges: the outer boundary of the hair blends into the background in a way that looks slightly like a cutout, particularly with fine or flyaway hair.

Jewelry and accessories

Rings, earrings, necklaces, glasses frames. AI generates these as part of the overall image rather than as separately rendered objects — which shows at close inspection. The metallic quality of jewelry is often slightly off: too uniform a reflection, slightly blurred where it contacts skin. Glasses frames sometimes clip into the face at the temples. Earrings can lose their shape at the edges.

Skin texture

This one is subtle and less reliable as a stand-alone tell. AI-generated skin is frequently smoother than real skin — not obviously airbrushed, but missing the fine-grain texture of actual skin (pores, fine lines, variation in tone across the face). High-resolution crops of the skin area can reveal this, but it requires the original image at reasonable resolution and isn't a reliable check on compressed or thumbnail-size images.

Eye symmetry and light reflections

The light reflections (catchlights) in both eyes should be consistent with each other and consistent with the apparent light source in the photo. AI-generated eyes are usually good now, but mismatched catchlights or slight differences in iris pattern between left and right eyes still appear occasionally. Look at both eyes carefully and compare them directly.

Background logic

Step away from the face and look at the background. Bookshelves with titles, signs with text, clocks showing time — AI generates these as plausible-looking but often incoherent. Text is frequently garbled or follows letter shapes without being readable. Architectural elements in the background sometimes violate physics (shadows going wrong directions, perspective inconsistencies). This is one of the most reliable tells in photos where the background has detail.

The "too perfect" quality

Beyond specific anatomical tells, there's a harder-to-define quality that trained observers notice: AI-generated faces are optimized for a certain kind of attractiveness that makes them look slightly unlike real people's photos. Everything is in proportion. The skin is consistently good across the whole face. The lighting is flattering in a controlled, specific way. There's a uniformity to it that's different from the randomness of real life.

This isn't a definitive tell on its own. Real people take professional photos. Real people have good skin days. But combined with any of the specific tells above, this general quality matters.

It also explains why fake profile photos tend to look like they were professionally photographed: the underlying optimization is for images that look like high-quality portraits, because the training data includes many of them. An AI face generator asked to produce a dating profile photo doesn't produce a casual selfie — it produces something that looks like a headshot.

When visual inspection fails — use a detection tool

Current AI generators have gotten good enough that the visual tells above aren't sufficient for catching all AI-generated faces. A well-rendered image from a recent model may pass every visual check and still be AI-generated. The pixel-level statistical signature of the image is detectable even when the image looks completely real.

Faux Spy analyzes these statistics — noise patterns, color gradients, frequency distributions across the image data — and returns a confidence score. The check takes a few seconds and doesn't require uploading the image anywhere. Hover over the image in Chrome, click Investigate, and you have a verdict.

For dating profile verification specifically, see the catfish detector guide and the platform-specific guides for Tinder, Bumble, Hinge, and Facebook. For understanding AI images more broadly, see the full guide to how to tell if a photo is AI generated.