How to tell if an image is AI-generated or a real photo

Can you know for sure whether an image is AI-generated?

No. There is currently no method, manual or automatic, that always separates a real photo from a generated image. What you can do is collect signals and weigh each one.

Think in terms of likelihood. One weak signal is not enough to claim anything. Several independent signals that point the same way, especially when they include the image's origin, support a reasonable conclusion.

Whether an image is AI-generated is not always the right question. A real photo can be edited, cropped, or posted with a false date and location. Checking the source helps in all of those cases.

What visual clues give away an AI-generated image?

Visual errors are the best-known clue and also the least reliable. Zoom in and check the areas where generators tend to slip:

• Hands and fingers: too many or too few fingers, impossible joints, hands that blend into objects.
• Text and signs: warped letters, nonsense words, signs that look like writing but say nothing.
• Backgrounds: blurred or distorted faces in a crowd, repeating patterns, objects that melt into each other.
• Shadows and reflections: shadows falling in different directions, reflections in mirrors, glass or water that do not match the scene.
• Accessories: mismatched earrings, glasses with uneven frames, necklaces that vanish into skin.
• Skin and texture: skin that is too smooth, with no pores or blemishes, or hair that looks painted on.
• Impossible physics: unsupported objects, structures that could not stand, proportions that do not add up.

Two caveats apply. Newer generators make these mistakes less and less, so a flawless image is not necessarily real. And a real photo can look odd because of the lens, compression, a beauty filter or retouching, so one strange detail does not prove an image is synthetic either.

How do you find where an image came from with a reverse image search?

A reverse image search is usually the strongest check, because it does not depend on how the image looks but on where it came from. Tools such as Google Lens and TinEye let you upload an image and see where else it appears.

1. Upload the image, or paste its address, into a reverse image search tool.
2. Look for the earliest appearance you can find and check who posted it.
3. See whether the original account or site describes it as an illustration or a generated image.
4. Check whether established news outlets carry the same image and who they credit.
5. For a news event, look for other photos or videos of the same moment taken from different angles.

A real, significant event almost never leaves a single image behind. If the only copy comes from an anonymous account and nobody else recorded the scene, that is a serious warning sign, though not proof.

Can metadata tell you if a photo is real?

It helps when it is there, but its absence proves nothing. EXIF metadata is information the camera or phone stores inside the file: device model, date, settings and sometimes location. Some software also records there that an image was edited or generated.

It has two limits. Many social networks and messaging apps remove metadata when an image is uploaded or forwarded, so most images in circulation arrive without it. Metadata can also be edited: it is something the file claims, not a guarantee.

If you have the original file, you can inspect what it contains with the metadata viewer. If you care about the opposite problem, protecting your privacy before you share, read how to remove metadata from photos.

What are Content Credentials (C2PA) and how do you check them?

Content Credentials are a digitally signed provenance record that some cameras, editing apps and image generators attach to a file. They follow the open C2PA standard and can show which device or software created the image and how it was edited.

To check them, load the file into a compatible verifier, such as the one offered on the Content Credentials site. Some platforms also display a label when they detect this data.

When credentials are present and the signature is valid, they are a strong signal about origin. But as with metadata, their absence means nothing: most cameras and apps do not add them yet, and they are lost when someone takes a screenshot or when the file passes through a platform that does not preserve them.

Some generators also embed invisible watermarks in the pixels. They exist, but each one can only be read with its maker's own tool, so there is no universal way to check for them.

Do AI image detectors work?

They offer a useful hint, not proof. A detector analyzes the image and estimates whether it looks more like a photograph or a generated picture. It is wrong in both directions: it can flag a real photo as AI and pass a generated image as real.

Compression, cropping, screenshots, filters and forwarding through messaging apps all degrade the signals a detector relies on. The more an image has traveled, the less dependable the result.

Our AI image detector works within the same limit. It is an automatic estimate made by a general-purpose AI model, not a forensic system. It answers in three levels (probably made by a person, inconclusive, or probably AI-generated), with a confidence level, the signals it observed and the limits of the analysis. Use it for guidance, not to accuse anyone or to prove a point.

What should you do before sharing or acting on an image?

Work through this list. It takes a few minutes and prevents most mistakes:

1. Get the best-quality version available, not a screenshot.
2. Run a reverse image search and find the earliest post.
3. Identify who published it and whether that source has a motive or a record of misleading people.
4. Look for independent images or accounts of the same event.
5. Check metadata and Content Credentials if you have the original file.
6. Examine the visual details slowly, zoomed in.
7. Run the image through a detector and read the result as one more clue.
8. If you still cannot confirm it, do not share it as real, or say clearly that it is unverified.

If you need a record of what you examined, for a news story or a complaint for example, the authenticity report puts the file's fingerprint, the date and the estimate in one document. It shows which file was reviewed and when; it does not certify that the image is true.

What mistakes do people make when checking an image?

• Relying on one signal. Neither fingers nor a detector settle the question alone.
• Assuming an image is real because it looks perfect. Visual quality no longer separates anything.
• Deciding it is AI because it has no metadata. Almost no image from social media keeps it.
• Analyzing a screenshot. Quality, metadata and credentials are all lost.
• Treating a detector result as proof. It is an estimate that can be wrong.
• Ignoring context. A genuine photo with a false caption misleads too.

The same caution applies to writing. For that case, see whether AI text detectors work.