Guide · 9 min read

How to tell if an image is AI-generated

Last updated: May 2026

You see an image online. It looks plausible, maybe a bit too plausible. Was it taken by a real person, or generated by a model that's never seen the sun? The good news: in many cases, the image itself will tell you. You just have to know where to look.

There's a particular kind of anxiety that comes with looking at an image in 2026. Is the politician really doing the thing in the photo? Did the photographer actually visit that place? Is that artist's portfolio their own work or output from a model trained on it? A few years ago, you could mostly trust your eyes. Now you can't.

This guide is a practical walkthrough of what actually works for checking whether an image was made by AI. We'll cover the methods that are reliable (and how to use them), the methods that sound reliable but mostly aren't, and what to do when nothing definitive turns up.

The thing most articles get wrong about AI image detection

If you've read other guides on this topic, you've probably been told to look for distorted hands, garbled text, wrong shadows, or "that AI look." This was useful advice in 2023. It is mostly useless in 2026.

The pixel-level tells that early image models left behind have been engineered out of the newer ones. Hands look right. Text reads correctly. Shadows fall the right way. The visual tells that remain are subtle and inconsistent, and any human running their output through one more pass of image editing can clean them up.

Pixel-based AI image detectors have the same problem. They look for statistical patterns left in the image data, and they struggle for the same reason: newer models leave fewer of them, and a single crop or resize is often enough to confuse the detector entirely. The result is high false-positive rates on real edited photos and high false-negative rates on newer or processed AI images. We've written about why these detectors are unreliable in general — but the short version is: don't trust a percentage from a pixel detector with anything important.

There's a better approach, and most articles don't mention it.

The single most reliable check: read the metadata

An image file isn't just pixels. It also carries metadata — information about how, when, and by what software the file was created. Real cameras and phones record details like the camera make and model, the lens, the exposure settings, GPS coordinates, and a timestamp. AI generators don't have any of that. What they often have instead is their own name, baked into the file.

Here's what we mean. If a real photo from an iPhone passes through your hands, its metadata will probably include:

If it's an image from Midjourney, its metadata might include:

One look at those two metadata profiles and you don't need to squint at pixels at all. The file is telling you what it is.

C2PA: the cryptographic version of "the file is telling you what it is"

The big improvement here is something called C2PA, which stands for the Coalition for Content Provenance and Authenticity. It's an open standard developed by Adobe, Microsoft, Google, the BBC, Intel, and a long list of others. The user-facing name for it is Content Credentials.

What C2PA does is embed cryptographically signed provenance metadata into an image. Where regular EXIF data can be edited or forged, C2PA Content Credentials are signed by the tool that made them. Tamper with them and the signature breaks, which any reader can detect.

The companies that already embed C2PA Content Credentials in AI-generated images include:

If you can read the Content Credentials on an image and they say "Adobe Firefly" or "DALL-E," that's about as close to definitive as you can get. The image is announcing itself.

The five-step practical check

Here's the workflow we'd actually recommend, in order of effort:

1. Check the metadata

Drop the image into a metadata viewer like our image checker. Look for any of:

If any of these are present, you have your answer. If they're not, move on — absence doesn't prove anything.

2. Check for camera signals as counter-evidence

Now look for things that suggest the image came from a real device. Strong camera metadata — make, model, lens, exposure, GPS — is hard for a generator to fake convincingly. A complete set of camera fields points away from AI. The absence of camera data points… nowhere in particular, because most platforms strip it.

3. Reverse image search

Run the image through Google Images, TinEye, or Yandex. If the image appears in lots of places dated before late 2022, it predates the modern image-generation era. If it appears only in AI-related contexts (subreddits, generation-tool galleries, image-prompt sites), that's a strong signal.

4. Check the source context

Where did the image come from? An account that posts hundreds of similar images, all photorealistic but uncreditable, is almost certainly running a generator. A photographer with a verifiable portfolio, dated work, and a body of evidence behind the image is the opposite. The image alone often can't tell you, but the context around it usually can.

5. If it really matters, ask for the original

This is the step almost no online guide mentions. If you genuinely need to know whether an image is real — for a news story, a legal matter, an academic question — the right move is to ask the source for the original file straight off the device. A real camera or phone original will carry intact metadata. A re-shared image won't. If someone can't produce an original, that's information too.

Why social media images are usually unverifiable

The frustrating fact: most images you encounter online have been laundered through a platform that strips metadata. Instagram, Facebook, X, TikTok — they all re-encode uploads, and the re-encoding throws out the original file's metadata along with the pixel data. Even an image that started life with full C2PA Content Credentials usually arrives on your screen with nothing attached.

This isn't malicious. Platforms strip metadata partly for file size, partly for user privacy (people leak GPS coordinates without realising). But the side effect is that the verification trail snaps the moment an image touches a major platform. Even a screenshot of an AI-generated image destroys the C2PA manifest, because the operating system writes a fresh PNG with no inherited metadata.

This is why the answer "I checked the metadata and it was clean" doesn't prove the image is real. It just means the trail's been wiped — which could mean anything.

CHECK AN IMAGE NOW

Our free image metadata checker reads C2PA Content Credentials, EXIF generator signatures, and Stable Diffusion metadata. No signup, no AI required.

→ Try the image checker

Honest expectations

Here's the realistic state of AI image detection in 2026: when an image has its metadata intact, identifying AI generation is often easy. When metadata's been stripped, no detection method is reliable enough to bet on. Pixel-based detectors will tell you anything with confidence, but their confidence isn't worth much.

The best mental model is to treat detection like investigative work, not a yes/no question. A C2PA signature from DALL-E is definitive. The absence of one is not. Camera metadata weighs against AI. The absence of camera metadata weighs nowhere. A reverse image search showing the image predating ChatGPT is strong evidence it's real. Source context matters as much as any technical signal.

Anyone who tells you they can verify an arbitrary image as definitely real or definitely AI, just from looking at the pixels, is overstating what's possible. The honest workflow involves several signals and a willingness to say "I don't know" when none of them are clear.

Further reading

Published May 2026 · telltale-ai.com
All articles · Text checker · Image checker · Privacy · Terms