How to Spot AI Images and Deepfakes

You spot an AI image the same way a photo editor checks a suspicious picture: a quick look for visual glitches, then a hard look at where it came from. The trouble is that the visual glitches are fading fast. The warped hands and melted faces that gave the game away a couple of years ago are mostly gone from the better generators, so eyeballing a picture is no longer enough. Verification is what holds up now: finding the original, checking its provenance, and asking whether the thing it shows happened at all.

The honest short answer

Trust the source, not your gut. For a while, spotting AI meant hunting for anatomical mistakes, and it worked because the models were bad at hands, teeth and text. That window is closing. The strongest image generators in 2026 clear most of the old traps, and the ones that slip through are getting subtler by the month. So the honest advice is to stop trying to win by staring harder. Use the visual tells to raise suspicion, then confirm or dismiss that suspicion with the checks further down this page. When it matters who made an image and whether it is real, verification is the part that actually decides.

The visual tells that still help

Some giveaways survive, at least for now, and they are worth knowing because they cost nothing to check. Run through this list on any image that feels off:

• Hands and fingers. Count them. You will still occasionally find six fingers, fused knuckles, or a finger that tapers into a smudge.
• Teeth. Too many, too even, or a row that blends into one continuous shape when you zoom in.
• Eyes and reflections. Mismatched catchlights, irises that do not match, or reflections in glasses and windows that show the wrong scene.
• Text inside the image. Signs, labels, logos and book spines often come out as garbled letters that look like writing from a distance and fall apart up close.
• Waxy skin and melting hair. Skin that looks airbrushed to plastic, and strands of hair that merge into the background or dissolve at the edges.
• Lighting and shadows. Shadows falling in different directions, a subject lit from a source that is not in the frame, or a face too clean for the scene around it.
• Warped backgrounds. Bent doorframes, railings that do not line up, patterns that repeat oddly, and buildings whose windows drift out of grid.

Here is the important caveat. Every one of these is something the top generators now handle well much of the time, and the models improve with each release. A picture that passes all seven checks is not confirmed real. It just did not fail the easy tests. Treat a clean pass as a reason to keep looking, not a reason to stop.

The verification steps that actually work

When the stakes are real, work the source instead of the pixels. These steps hold up whether or not the image has an obvious glitch, and they are what fact-checkers and newsrooms lean on. Do them in order:

1. Reverse image search to find the original. Drop the image into a reverse search and look for the earliest and highest-quality version. This often surfaces the real context: a photo from a different year, a scene from a film, or a stock image that has been re-captioned to mislead.
2. Check provenance and metadata. Look for Content Credentials, the signed history added under the C2PA standard. Many AI tools and some cameras now attach one, and a viewer will tell you whether the image was generated, edited, or shot on a camera. Absence of a credential proves nothing on its own, but a present one is a strong signal.
3. Trace the source and cross-check reputable outlets. Follow the image back to who first posted it and ask whether that account has any standing. Then check whether established news organisations are carrying the same image or event. A major moment that only lives on one anonymous account is a red flag.
4. Ask whether the event appears anywhere credible. If an image shows something newsworthy, that thing should leave a trail: multiple outlets, named witnesses, corroborating footage. Silence everywhere trustworthy is itself an answer.

None of these requires you to be an expert or to spot a single pixel out of place. That is the point. The source is far harder to fake than the image, so it is where your attention pays off.

Deepfake video specifics

Video adds motion, and motion adds tells that a still image cannot. When you are watching something suspicious, slow it down and watch a few specific things. Lip-sync drift is the classic one: the mouth and the words fall slightly out of step, most visible on hard consonants. Blinking often looks wrong, either too regular, too rare, or oddly timed. There is frequently a faint flicker or shimmer at the edge of the face, along the hairline and jaw, where the swapped face meets the real head. And the audio can betray it too, sounding flat, clipped, or missing the small room noise a real recording carries.

The same warning applies here as with images. Purpose-built video models are closing these gaps quickly, and a well-made deepfake in 2026 can pass every one of these checks. So the tells are useful for catching sloppy fakes, but they are not a guarantee. For anything that matters, fall back on the source: who posted it, where it first appeared, and whether any outlet you trust has verified it.

Why detectors are unreliable, and what to trust instead

It is tempting to paste the image into an AI detector and let it decide. Resist that. Detectors output a probability, not a verdict, and that probability is shakier than the confident percentage on screen suggests. They tend to miss images from newer generators they were not trained on, they get thrown off by cropping, resizing and compression, and they flag genuine photos as fake often enough to do real harm. A detector is one weak signal, not the answer.

What earns trust is the combination: a quick visual scan, a reverse image search, a provenance check, and corroboration from sources with something to lose if they get it wrong. If you want to sharpen the visual half of that, our Spot AI images and deepfakes tool lets you practise on real and generated images side by side, which trains your eye faster than any checklist. The same problem shows up in writing, and it is worth reading how to spot AI-written text and trying its tool so you can test yourself on prose the same way. For the wider set of skills that make all of this click, the AI literacy guide pulls it together.

The takeaway is not that images can no longer be trusted. It is that trust has moved. It used to live in the picture; now it lives in the picture’s history and the reputation of whoever is standing behind it. Learn to check those, and a fading set of visual tells stops being your only defence.