AI Image Detection — what this tool checks, what it can't see, and how to read the score

Midjourney v6, DALL·E 3, Stable Diffusion XL and Flux can all make a fake portrait good enough to fool most people on first glance. Veo and Sora can fake a 10-second clip the same way. This page sends your image to a vision model that returns a yes/no with a short reason for its answer. Below is a plain-English walk through how the check is done, what makes it fail, and the small visual tells you can still spot with the naked eye.

Try the spot-the-fake test before reading on

Pick the next portrait you see on LinkedIn or X. Now ask yourself three questions before scrolling past:

• 1.Do both earrings match? (Most AI portraits get one right and one wrong.)
• 2.Where does the hair meet the background — is the boundary clean, or does it melt?
• 3.Are the irises perfectly round and centred? (StyleGAN faces almost always centre the eyes; real photos rarely do.)

If one of those three is off, run the photo through the detector above. If all three look fine and the photo still feels off, run it anyway — the things you can see are only the ones that haven't been fixed yet.

What the detector actually does to your image

The image is sent to a vision model that has been shown a large set of real photos and a large set of AI-generated photos and learned the difference. Three kinds of clue carry most of the weight:

Camera noise that should be there

Every camera sensor leaves a faint speckled pattern on every photo it takes — a fingerprint that is unique to that sensor. AI images have no sensor and no fingerprint. A clean, noise-free photo with sharp edges in every region of the frame is the single strongest sign of generation.

Frequency patterns the eye misses

Diffusion models build images out of noise in steps. That process leaves regular ripples in the high-frequency parts of the image — the very fine detail in skin, fabric and grass. The detector runs a frequency analysis (a 2D Fourier transform) and looks for the grid-like pattern these ripples create. You won't see it; the model does.

Things that don't make physical sense

The vision model also checks that shadows point the same way, reflections match the scene, hands have five fingers, and any text in the photo is real text rather than letter-shaped shapes. It returns a percentage, a confidence band, and a one-paragraph note that names the things it noticed.

Tells you can still spot by eye

Even with Midjourney v6 the following errors slip through often enough to be worth checking:

• •Hands. Count the fingers. Look at how the fingers wrap around an object. A coffee cup with a thumb on both sides is a giveaway.
• •Text inside the photo. Look at any sign, book cover or t-shirt slogan. AI writes letter-shaped doodles, not real words. If the photo claims to be from a specific place, the language on the signs should match.
• •Earrings and shirt collars. The two sides are almost always slightly different. Real people wear matching pairs.
• •The background. Look at the parts of the photo that aren't the subject. AI backgrounds tend to dissolve into vague shapes — windows that don't quite line up, stairs that lead nowhere, leaves that morph between two plants.
• •Skin on older people. AI smooths skin into a kind of polished plastic. A real 70-year-old has pores, freckles, and fine lines that aren't symmetric.

What this tool will not catch

Honesty matters here more than marketing. The detector misses three types of image often enough that you should know:

• ⚠A real photo that has been heavily edited. Strong noise reduction, beauty filters and aggressive cropping can strip the same sensor noise the detector looks for. The detector then leans toward "AI-generated" — a false positive. If a photo has been through an Instagram filter, take the score with a pinch of salt.
• ⚠A screenshot of an AI image. The act of taking a screenshot adds the device's own pixel grid back on top of the AI image. That grid masks the giveaway frequency patterns, and screenshotted Midjourney portraits slip through more often than you'd expect. Always check the original file if you can.
• ⚠A real photo with one AI element pasted in. A real wedding photo with the bride's face swapped out, for example, will often score as "real" because most of the pixels are. The detector flags whole-image generation better than localised edits.

Reading the percentage

• ✓90–100% real or AI: safe to trust. Both extremes have strong sensor-noise or frequency-pattern evidence backing them up.
• ~60–90%: the model leans one way but the photo has been edited or compressed enough that some signals are degraded. Use the written reason rather than the number.
• ?40–60%: the detector can't tell. Either the image is from a newer model the detector hasn't seen, or it's a hybrid. Look for a second opinion — the visual tells above, the Content Credentials metadata (see below), or a reverse image search to find the source.

Content Credentials — the metadata signal worth knowing about

Some recent images carry a small cryptographic tag inside the file called C2PA Content Credentials. Adobe Firefly, Microsoft Designer, Leica M11-P cameras and OpenAI's tooling all write this tag, which says who made the image and whether AI was involved. You can read it on contentcredentials.org/verify. If a photo has the tag and it says "captured on a camera", that's a stronger signal than any AI detector. If the tag is missing — which is still the case for most photos on the web — fall back to the detector and the visual tells.

Frequently asked questions