Spot AI Images & Deepfakes

Why AI images and deepfakes are hard to catch now

For a couple of years, spotting an AI image was almost a party trick. Count the fingers, read the text on the sign, notice the plastic skin, and you had your answer. That era is closing. The image and video generators of 2026 have quietly fixed most of the crude giveaways: they render six-fingered hands far less often, keep lighting more consistent, and produce skin and hair that survive a second glance. Deepfake video has improved in lockstep, with tighter lip-sync and more natural motion than the uncanny early demos.

The result is a moving target. A checklist of visual tells is still worth having — plenty of images in the wild come from fast, cheap or older tools that still slip up — but you have to hold it loosely. The absence of a tell no longer means an image is real. That is exactly why the second half of this guide matters more than the first: when the pixels can lie convincingly, you win by checking where the image came from, not by squinting harder at it.

The visual tells, in detail

Here is what each item on the checklist actually looks like, and a note on which ones are fading as models improve.

• Hands & fingers. The classic tell — wrong finger counts, fused digits, joints bending impossibly. Still worth checking, but the leading models now get hands right most of the time, so this one is fading fast.
• Teeth. Generators often render an implausible number of teeth or an unnaturally uniform row that blurs into a single band when you zoom in. A weakening tell, but still common in group shots and wide smiles.
• Eyes & pupils.Look for mismatched or misshapen pupils, irises that differ between the two eyes, and catch-lights (the little reflections) that don’t match the scene’s light sources.
• Ears & jewellery.Faces are roughly symmetrical; AI sometimes isn’t. Watch for one earring, mismatched ears, or jewellery that changes shape, doubles, or fuses into the skin.
• Text in the image. One of the most durable tells. Signs, logos, labels and book spines often come out as gibberish or almost-right lettering. Newer models handle short words better, but longer text still tends to break down.
• Skin texture. A waxy, airbrushed, pore-less smoothness paired with a flawless studio glow. Improving quickly, and easily masked by a grainy filter, so this one is increasingly unreliable on its own.
• Hair strands. Individual strands dissolving into the background, clumps with no clear ends, or a hairline that blurs into the forehead. Fine detail is hard to fake, so hair remains a useful place to look.
• Lighting & shadows. Physics is unforgiving. Shadows that fall in different directions, a face lit from one side while the body is lit from another, or objects that cast no shadow at all. Global consistency is genuinely hard for generators.
• Reflections.Windows, mirrors, glasses and even eyes should reflect what is actually in front of them. AI frequently invents reflections that don’t match the scene — a strong, durable tell.
• Background. The model spends its effort on the subject and cuts corners behind it: warped objects, impossible geometry, repeating textures, and background people or railings that bend and merge. Often the richest hunting ground.

Notice the pattern: the tells that survive are the ones rooted in physics and global consistency — lighting, shadows, reflections — and in fine, repetitive detail like hair and dense text. The tells that are disappearing are the ones models were explicitly trained to fix, like hands and smooth skin. Weight your attention accordingly.

The verification workflow that beats eyeballing

When it actually matters — before you believe, share, or act on an image — stop staring at pixels and run these steps in order.

1. Reverse image search. Drop the image into Google Lens or TinEye. This is the single highest-value move: it often finds the original, an earlier date, the real context, or a fact-check debunking it. A photo that first appears only on an anonymous account is a warning sign in itself.
2. Check provenance & metadata. Look for C2PA Content Credentials (inspect them at contentcredentials.org/verify) and scan the EXIF data for a camera make, model and timestamp. Present credentials are a strong positive signal; but remember that missing metadata proves nothing, because a screenshot or re-upload strips it.
3. Trace the source and cross-check. Where did this actually come from? Follow it upstream to a named account or outlet, and check whether reputable, independent organisations are reporting the same thing. A real, newsworthy image rarely lives in a single unverifiable post.
4. Look for the improbable. Ask whether the event even makes sense. If a genuinely dramatic thing happened, credible outlets would be covering it. If the only evidence is one striking image and silence everywhere else, treat that silence as the answer.

Deepfake video specifics

Video adds motion, and motion adds new tells. Watch for lip-sync that drifts out of time with the words, blinking that is too regular or eerily absent, a faint flicker or seam around the edge of the face where the swap was blended in, skin tone that shifts as the head turns, and audio that sounds flat, clipped, or dotted with artefacts. Playing the clip at reduced speed and full screen makes these far easier to catch.

The same caveat applies, only more so: real-time and pre-rendered deepfakes are improving fast, and a well-made one can pass every one of these checks. So video verification follows the same rule as images — trace it to a reputable outlet and ask whether the moment appears anywhere credible. If you want to sharpen the related instinct for machine-written words, our companion tool Spot the AI text trains you on human-versus-AI writing samples.

The bottom line

Pixel-peeping is a fun skill and a useful first filter, but it is a losing arms race — every tell you learn is one the next model is being trained to erase. The durable defence is a shift in habit: assume a striking image might be synthetic, and make provenance and source-checking your default before you trust it. Reverse image search, Content Credentials, and a quick sanity check against reputable outlets will catch far more fakes than the sharpest eye. Use the checklist to raise your suspicion; use verification to settle it.