How to Tell if Photos Are AI-Generated (2026 Guide)

AI image generators have improved fast, and the tells that worked in 2023 — six fingers, melted text, obviously waxy skin — are far less reliable now. Faces, specifically, are usually the strongest and most convincing part of a modern AI-generated image, because that is what these models have been trained hardest on.

That means the smartest approach in 2026 is not staring at the face looking for a giveaway. It is checking the areas generators still struggle with — edges, backgrounds, small objects, and consistency across an entire set of photos — and combining several weaker signals into one confident judgement rather than hunting for a single smoking gun.

This guide covers the tells that still hold up, what to check on a live video call, why metadata and provenance tools matter more than people realise, and how to read an 'AI detector' percentage sensibly rather than treating it as a verdict.

Where generators still struggle: hands, ears, teeth, jewellery

Even strong models still make mistakes in specific, predictable places. These are worth checking first because they are quick to spot and still fairly reliable in 2026.

• Look closely at hands and fingers for extra digits, fused fingers, or fingers that bend in an anatomically odd way, since hands remain one of the harder structures to generate consistently.
• Check ears for asymmetry, missing detail, or an ear shape that does not match between two photos supposedly of the same person.
• Zoom in on teeth for unnatural evenness, blurring at the gumline, or teeth that seem to merge into one shape rather than separating cleanly.
• Examine jewellery, especially earrings and necklaces, for chains that break, warp, or don't connect logically around the neck or ear.
• Look at glasses frames for arms that vanish behind the ear incorrectly or lenses that distort the eye shape unnaturally.

Background and environment tells

Generators put most of their effort into the subject's face and much less into the surrounding scene, so backgrounds are often where the illusion breaks down.

• Look for background text — signage, book spines, labels — that turns into meaningless squiggles rather than real letters.
• Check for architectural elements that don't make physical sense, like railings that end mid-air or windows with impossible geometry.
• Notice if a crowd or background group of people has faces that are blurred, repeated, or subtly warped compared with the sharp foreground subject.
• Watch for objects that blend into each other at the edges, such as a hand appearing to merge with a drink glass or a bag strap disappearing into clothing.

Hairlines, lighting, and skin texture

Hairlines and flyaway hair

Individual strands of hair, especially where hair meets the forehead or background, are still a common weak point. Look for hair that fades into a soft blur rather than showing individual strands, or a hairline that looks painted on rather than growing naturally from the scalp.

Lighting inconsistency

Check whether the direction of shadows on the face matches the direction of light in the rest of the scene. AI-generated images frequently light the subject's face from one angle while the background implies light coming from somewhere else entirely, which is a strong and often-overlooked tell.

Uniform, poreless skin

Real photos, even filtered ones, usually show at least some texture — pores, fine lines, minor blemishes — especially in a well-lit close-up. Skin that looks completely uniform and smooth across the whole face, with no texture variation at all, is worth treating as a signal, though it is not decisive on its own since heavy filters can produce something similar.

The composition giveaway: too-perfect, too-similar photos

One of the most reliable modern tells has nothing to do with pixel-level detail. It is noticing that every photo on a profile shares the same lighting style, the same slightly-too-perfect symmetry, or nearly identical framing, as if they were all produced by the same process rather than taken on different days with different cameras.

• Compare the lighting style across all the photos on a profile — real photo sets usually show natural variation from different days, times, and locations.
• Look for a repeated, subtle 'model shoot' quality across casual-looking photos that are supposedly candid snapshots.
• Check whether facial proportions stay exactly identical across photos taken at supposedly different angles, since real faces shift naturally with angle and expression.
• Notice if the profile has an unusually small number of photos, all looking like professional headshots, with none of the ordinary mess of a real phone camera roll.

Video calls and deepfake tells

Live video is harder to fake convincingly than a still photo, which is exactly why asking for a video call remains one of the best verification steps available. Real-time deepfake and face-swap tools do exist, but they still show specific strain under certain conditions.

• Ask the person to turn their head fully to the side during the call, since face-swap overlays often glitch, blur, or lag at extreme angles.
• Ask them to hold a hand up in front of their face briefly, since overlapping objects frequently cause visible warping or flickering in real-time deepfakes.
• Watch the edges of the face and jawline for a faint shimmer, blur, or colour mismatch against the neck and background, especially when they move quickly.
• Pay attention to audio-lip sync, since even good face swaps can drift slightly out of sync with speech under a poor connection.
• Ask an unexpected, simple question mid-call and watch for an odd delay, since some live-swap setups need a moment to process before responding naturally on camera.

Metadata and provenance: C2PA and what it can tell you

Visual inspection has limits, so it helps to know that a growing number of cameras and editing tools now embed provenance information directly into image files under a standard called C2PA, which records whether an image was captured by a camera or generated or edited by AI.

• Check if an image carries C2PA 'Content Credentials' where supported, since this can directly show whether a tool recorded the image as AI-generated.
• Remember that most images shared through messaging apps and social platforms have metadata stripped out during upload, so an absence of metadata proves nothing either way.
• Treat metadata as a bonus signal when it's present and reliable, not as something you can expect to find on every photo.
• Be aware that metadata can, in principle, be stripped or altered, so it should support other checks rather than replace them.

Reading AI-detector percentages sensibly

Automated AI-image detectors are useful, but a score like '78% likely AI-generated' should be read as a signal that points you in a direction, not a courtroom verdict. These tools can be fooled by heavy compression, filters, or re-uploading, and they can also misjudge unusual but genuine photos.

• Use a detector score as one input alongside visual inspection and reverse image search, not as the only check you run.
• Be more confident in a strong result — very high or very low — than in a score that sits close to the middle.
• Re-run a detector on more than one photo from the same profile before trusting the result, since a single unusual photo can skew a single check.
• Remember heavily compressed or re-saved images (common after passing through messaging apps) can confuse detectors in either direction.

Our guide to reverse image searching a dating profile pairs well with this one — an AI-generated face often produces a clean, empty reverse image search precisely because it has never existed anywhere else, which is itself worth noticing.

Judging the whole profile, not one photo

The strongest, most reliable approach in 2026 is to stop looking for one decisive tell and instead build a set-level judgement. Look across every photo on the profile together: does the lighting style repeat suspiciously, do facial proportions hold up across angles, does at least one photo show ordinary real-world imperfection, and does a live video call match the photos convincingly. Any one weak signal is easy to dismiss; several weak signals pointing the same way are not.

Quick reference: AI photo checklist

• Zoom in on hands, ears, teeth, and jewellery for the classic structural mistakes generators still make.
• Check background text, architecture, and crowds for warping that the sharp foreground subject doesn't share.
• Compare lighting direction on the face against lighting in the rest of the scene.
• Look across the whole photo set for repeated, too-perfect composition rather than natural variation.
• On a video call, ask for a side profile, a hand across the face, and an unscripted question.
• Check for C2PA content credentials where available, and treat their absence as inconclusive rather than suspicious.
• Treat AI-detector percentages as a supporting signal, most useful when the score is clearly high or clearly low.