Verdict → Guide

How to tell if an image is AI-generated

Last updated 17 August 2026

The visual giveaways people learned to look for in 2023 — six-fingered hands, garbled text on signs, melted jewellery — have largely been fixed by current image models. Advice built around them is now actively misleading: a clean pair of hands tells you almost nothing. What follows is what still works, in the order worth doing it.

1. Read the file's metadata first

This is the highest-value check and almost nobody does it. A photo taken by a real camera carries EXIF metadata written by the device: camera make and model, lens, focal length, aperture, shutter speed, ISO, and often GPS coordinates. Image generators do not produce those fields, because there was no sensor and no exposure.

You can check it without installing anything:

• iPhone: open the image in Photos, swipe up or tap the ⓘ button. A camera original shows the device name, lens and exposure values.
• Mac: open in Preview, then Tools → Show Inspector → the ⓘ tab → Exif.
• Windows: right-click the file → Properties → Details.

Present, coherent camera metadata is strong evidence the image is a real photograph. The fields are also internally consistent in ways that are awkward to fake — the focal length matches the lens, the exposure triangle makes physical sense.

Be aware that EXIF fields are editable — anyone determined can write plausible camera data into a generated file. It is a strong signal, not a signature.

2. Search for the image elsewhere

Reverse image search answers a different question: not "was this generated?" but "where did this come from?" That is often the more useful answer.

• Google Lens, TinEye, Bing Visual Search and Yandex all accept an uploaded image.
• If the image turns up in a stock library, a news archive, or a post from years before the current claim, you have your answer regardless of what any detector says.
• If a supposedly newsworthy image appears nowhere else, that absence is itself informative. Real events of consequence get photographed more than once, from more than one angle.

3. Look for physical inconsistencies, not anatomical ones

Generators model appearance, not physics. The artefacts that survive in current models are the ones that require a scene to be globally consistent:

• Light direction. Do all the shadows fall consistently with a single light source? Mixed shadow directions in a scene with one obvious sun are hard to explain.
• Reflections. Mirrors, windows, water, sunglasses and eyes. Reflected content that does not correspond to the scene in front of it is a strong tell — reflection requires a model of geometry the generator does not have.
• Background text. Not the foreground signage models now render well, but small, out-of-focus, incidental text — spines of books, distant shopfronts, labels.
• Repetition. Crowds, foliage, brickwork, textiles. Look for the same element copy-pasted at slightly different scales.
• Where things meet. Hair against a background, fingers around an object, the join between a person and what they are holding or sitting on.
• Depth of field that does not match. Blur that varies without following a coherent focal plane.

Zoom in. Most of these only become visible at 200–400%, and most people never zoom.

4. Check for content credentials

An increasing number of generators attach provenance data to their output — C2PA "Content Credentials", which record how a file was made and edited, and in some cases invisible watermarks embedded in the pixels themselves. Some camera manufacturers have begun signing files at capture for the same reason.

Where present, this is the single most authoritative signal available, because it is cryptographic rather than statistical. In practice coverage is still partial and the data is easily removed — screenshotting an image discards it entirely — so its presence is meaningful and its absence is not.

5. Run a detector — and know what it is measuring

Automated detectors combine several statistical signals. It is worth knowing that these signals differ enormously in quality, because a lot of published advice treats them as interchangeable.

The figures below are measured on Verdict's held-out validation set, reported as AUC — where 1.0 is a perfect separator and 0.5 is a coin flip:

Two of those four are barely better than guessing. Error Level Analysis in particular is widely recommended online and, on this data, carries essentially no signal for the AI-versus-real question — it was designed to spot local re-saving in edited JPEGs, which is a different problem. Sensor-noise analysis is only marginally better once an image has been through social-media re-compression.

This matters if you are reading someone's ELA heat map and drawing conclusions from it. On modern generated images, those maps are close to noise.

One important qualification, because the published literature looks like it disagrees. Several peer-reviewed papers report 89–100% accuracy for ELA-based AI detection — for example in Sensors (2023) and Scientific Reports (2023). Those studies are measuring something different: they feed the full ELA image into a convolutional network that learns whatever patterns it contains. The figure above is ELA reduced to a summary statistic and fed into a weighted score, which is how most detectors actually use it, and how a human reads a heat map.

Both results can be true at once, and the distinction is practical rather than academic. A CNN trained on ELA maps performs superbly on generators it saw in training and much worse on ones it did not — the 2026 IJET study that tested this explicitly found cross-generator performance was where these methods struggle most, and one ELA-based system that scored 96% fell to around 20–38% under added noise, cropping or histogram equalisation. So: ELA as a learned feature is genuinely useful in a narrow setting; ELA as a general-purpose "is this AI" signal is not. Verdict weights it at zero for that reason.

Why detection fails, and when to distrust it

Every detector is weakest in the same places. If your image is one of these, lower your confidence in any score you get:

• Screenshots. The worst case. No metadata, fresh compression, and the original pixel statistics destroyed. A screenshot of a real photo and a screenshot of a generated one look far more alike than the originals did.
• Social-media re-uploads. Stripped metadata plus aggressive re-compression removes much of what a detector reads.
• Heavily computational phone photos. Night mode, Smart HDR and portrait modes stack and process many frames. The output is real but statistically unusual — smooth in ways that can read as synthetic.
• Small or heavily cropped images. Less pixel evidence, less signal.
• Real photos of AI-generated things. A photograph of a screen showing a generated image is a genuine camera original. Detectors answer "how was this file made", not "is what it depicts true".
• Newer generators. Detection models are trained on the outputs that existed when they were built. A model released after the detector will always be the hardest case.

The direction of error also matters. A detector tuned to never wrongly accuse a real family photo will let some generated images through, and vice versa. Ask which way the tool you are using is tuned.

How to check a video

Video is checked frame by frame. The practical approach is to sample frames spread across the clip, score each one as an image, and combine the results.

Use the median of the per-frame scores rather than the maximum. A single glossy or motion-blurred frame in a genuine clip can spike the maximum and produce a false accusation; the median is far more stable. On clip-level testing this separates cleanly, because a generated video tends to be generated in every frame rather than in one.

The manual checks also transfer, with additions: watch for flicker in fine detail between frames, features that drift or reset across cuts, and inconsistent motion blur. Audio is a separate problem with separate tools.

A practical order of operations

1. Check the metadata. If there is coherent camera data, it is very probably a real photograph.
2. Reverse image search it. Provenance beats analysis whenever you can get it.
3. Zoom to 300% and look at reflections, shadows and incidental background text.
4. Run a detector, and read its per-check breakdown rather than only its headline number.
5. Ask where the image came from and who benefits from your believing it. Context resolves more cases than forensics does.

Common questions

Can you still spot AI images by looking at the hands?

Rarely, as of 2026. Hands, teeth and foreground text were the reliable artefacts of 2023-era models and current generators mostly render them correctly. Checking hands now produces false confidence in both directions. Reflections, shadow direction and incidental background detail have held up much better.

Does missing EXIF data mean a photo is AI-generated?

No, and this is the most common mistake. Almost every social platform strips EXIF on upload, and screenshots never had it. Missing metadata is what a normal shared image looks like. The inference only runs one way: coherent camera metadata is strong evidence of a real photo, but its absence is close to uninformative.

How accurate are AI image detectors?

Good ones reach roughly 90–95% on clean inputs — unmodified camera originals and untouched generator output. Accuracy drops substantially on screenshots, social-media re-uploads and heavily processed phone photos, because those strip away the evidence. Any detector quoting a single accuracy figure without saying what it was measured on is not telling you much. Treat every result as a probability to weigh, not a verdict to act on alone.

Is Error Level Analysis useful for detecting AI images?

It depends entirely on how it is used. As a standalone signal — the ELA heat map a person looks at, or an ELA statistic fed into a weighted score — it is much weaker than its popularity suggests: it measured an AUC of 0.57 on our validation set, close to a coin flip. ELA was designed to reveal parts of a JPEG re-saved at a different quality, which is a different problem from synthesis.

Published research reporting 89–100% accuracy for "ELA detection" is measuring something else: full ELA images fed into a convolutional network that learns the patterns in them. That works well on generators present in training and degrades sharply on unseen ones, and under added noise or cropping. So ELA is a useful learned feature in a narrow setting, and a poor general-purpose signal — which is why Verdict displays it but gives it no weight in the score.

Can an AI-generated image pass every check?

Yes. Someone who generates an image, writes plausible EXIF data into it, and prints and rephotographs it will defeat all of the pixel-level and metadata checks described here. What survives is provenance: where the image first appeared, who published it, and whether any independent record of the event exists. Forensics narrows the question; it does not close it.

How do I check whether a video is AI-generated?

Sample frames across the clip, score each one, and take the median of the results rather than the highest. A single unusual frame in a real clip can spike a maximum-based score. Also watch for flicker in fine details between frames and for features that drift across cuts. On iPhone, Verdict does this automatically.

Is there a free way to do this without an app?

Yes, and it covers most cases. Check the metadata in Photos, Preview or File Explorer; run the image through Google Lens or TinEye; and zoom in on reflections and background text. Those three cost nothing and resolve the majority of everyday questions. A detector adds value mainly when the file has been stripped of metadata and does not appear anywhere else.