How to Check if an Image Is AI-Generated: 7 Signals That Work in 2026

How to Check if an Image Is AI-Generated

There is no single trick that catches every AI image. The image generators are improving every quarter, and the detection tools are always a step behind. What you can do is build a habit — a quick, repeatable check that catches most synthetic images and surfaces a "this needs a second look" flag for the rest. The seven signals below are the ones that hold up on real-world output, not on cherry-picked examples.

A confident "AI" verdict is a hint, not a fact

Public AI detectors report false positives at rates that are uncomfortable for serious work. They are useful for triage. They are not useful as the final word. Use them to prioritize what to look at, then look.

Warning

Signal 1: Run a reverse image search first

Before you do anything else, run the image through TinEye, Google Images, and Yandex Images. You are looking for the earliest known version of the image.

If the earliest known version is older than the AI generator you suspect was used, the image is not AI — it is real, possibly recycled. The reverse-image test is not a synthetic-media test; it is a provenance test. Provenance is the only test that survives contact with better generators.

Signal 2: Read the small text in the image

AI image models still struggle to render legible text inside a generated image. Look at:

• Street signs, storefronts, book spines, product labels
• T-shirts with slogans, protest signs, license plates
• Screens in the background, monitors, phones

If a sign says "OPNE 24 HRS" or a t-shirt reads "STOP KIDING," the image is almost certainly synthetic. Real photos of real places rarely have these errors; AI models frequently do.

Signal 3: Look at reflections and shadows

Mirrors, windows, eyeglasses, jewelry, and polished metal reflect the world. AI models do not yet model reflections consistently. Look for:

• A reflection that does not match the scene
• A person wearing glasses whose lenses reflect a different room
• A watch face with reversed or garbled numerals
• Two earrings that do not match

Reflections are a free tell. A real photo has physics; an AI image has plausible-looking pixels.

Signal 4: Check the repeating details

Bricks, tiles, fence posts, columns of text, the pattern on a shirt, the buttons on a jacket — these are all small repeating details that AI models often get slightly wrong. Zoom in. Look for:

• A pattern that shifts subtly between repeats
• A line of buttons where one is missing or has a different number of holes
• A brick wall that gets denser or sparser as the eye moves across
• Eyelashes that merge or vanish at the edge of the frame

These are subtle. They are not a verdict on their own. But combined with another signal, they are damning.

Signal 5: Look at hands, teeth, and hair

The classic tells, even in 2026:

• Hands. Six fingers, merged fingers, hands with the wrong number of knuckles, hands that disappear into sleeves oddly.
• Teeth. Incisors that merge, a third row of molars, teeth that change size across the smile.
• Hair. Strands that appear and disappear at the edge of the silhouette, bangs that go behind an ear and reappear in front, frizz that ignores the wind direction in the rest of the image.

These are improving, but slowly. They remain reliable triage signals.

Signal 6: Cross-check the metadata

A JPEG or PNG can carry EXIF data: camera model, capture time, sometimes GPS. EXIF is easy to strip and easy to fake. Its absence tells you nothing. Its presence — a real camera model with realistic capture settings — is a small positive signal. A claimed iPhone 15 photo with EXIF from a Canon DSLR is more interesting than suspicious.

For a more rigorous check, run the image through FotoForensics for error-level analysis. The technique looks at how the file compresses and re-compresses, and it can show you spots that have been edited. A re-saved JPEG has different compression artifacts from a raw original. A screenshot has none. An AI image, since it was never a real photograph, has its own tell — uniform noise patterns across the image.

Signal 7: Look for the model fingerprint

Each major image generator leaves a faint, characteristic fingerprint in its output. Researchers publish these signatures. The tool Hive and a few academic projects maintain public detectors that catch many of the most common models. Treat their output as a hint, not a verdict, and combine it with the visual signals above.

The three signals that do not work

A few common tells are now unreliable:

• "It looks too perfect." Real photos are sometimes over-saturated, over-stylized, and over-processed. The aesthetic of "too clean" is no longer a synthetic tell.
• "The eyes don't match." Better generators now render symmetric, well-lit eyes reliably. Mismatched eyes are no longer a useful tell on their own.
• "There's a watermark in the corner." The major generators have removed visible watermarks. The presence of one in a "real" image is now itself a tell that something is off.

The test that still beats the detectors

When the visual signals are mixed and the detectors disagree, do the only test that actually scales: find the original.

If the image is from a real event, someone at the event took a photo, and that photo will eventually surface. If the image is from a real person, that person has other photos. If the image is of a real product, the product has a press kit. The provenance test — find the original, in the wild, taken by someone who was there — is the only test that has not been beaten by the next generation of generators. It is also the test that takes the most time, which is why you should reach for it only when the visual signals leave you uncertain.