Is This Photo AI? Checking a Viral Picture or Screenshot

To find out whether a photo is AI-generated, add it to the checker above. Your browser first looks inside the file for Content Credentials and camera metadata, without uploading anything. A screenshot almost never has either, so in most cases the image then goes to a detection model that judges the pixels themselves.

If you are asking "is this photo AI?" about something that just went viral on X, Facebook, TikTok or a group chat, this is the same AI image detection as on our home page, set up for screenshots: paste one with Ctrl+V. You get one of four answers: Likely AI-generated, Likely a real photo, Can't tell, or Edited with AI. Each comes with a probability range rather than a single score, and a list of the reasons behind it. You have 3 free checks a day.

Content Credentials are a signed record that some AI tools and cameras attach to the file they create. Under the C2PA standard behind them, the signature is bound to a hash of that file's image data, so the record belongs to that exact file and nothing else.

A screenshot is not a copy of that file. Your phone draws whatever is on the screen into a brand-new image, and the original's credentials don't come along. In our test, 24 of the 40 original AI files carried credentials naming the company whose tool made them. We also made 10 phone-style screenshots of AI images, and none of them carried any credentials at all.

Uploads are the other leak. When OpenAI started adding credentials to its images, it warned that most social platforms remove metadata from uploaded pictures and that taking a screenshot can remove it too. By the time a picture has been shared a few times, the credentials are usually gone, whoever made it.

Detection models don't need metadata. Sightengine, the main engine behind our checker, says its model analyses only the visual content and uses no metadata at all, which is why a screenshot can still be caught.

It does get a little harder. In our 10 original/screenshot pairs, Sightengine recognised all 10 originals and 9 of the screenshots. Hive, our backup engine, recognised all 10 originals and 8 of the screenshots. A screenshot adds a status bar and text, usually shrinks the picture, and is often compressed again when it is shared. Don't be surprised by a wider range or a Can't tell on a screenshot that you wouldn't get on the original file. A screenshot doesn't stop you asking "is this photo AI?"; it makes the answer a little less certain.

The closer you get to the original file, the more any check can tell you: every copy, crop and re-upload throws away detail and metadata. Before you decide, try these four ways to find a better copy of the picture.

• Save the image, not the screen. Many apps let you long-press or right-click to save a picture. The platform may still have stripped its metadata, but you get the picture without your phone's interface wrapped around it.
• Search for the earliest version. A reverse image search such as Google Lens, Bing or TinEye can lead you to the first post, a larger copy, or the photographer's own page.
• Read the first post. People who make AI images often say so in the caption or post them on accounts dedicated to AI art. That tells you more than any score.
• Check the biggest copy you can find. Thumbnails and reposts lose the fine detail a detector relies on.

A detector answers one narrow question: were these pixels made by an image generator? Plenty of viral fakes are real photos with a false caption: an old flood shared as today's storm, a crowd from another country, a crop that hides what was happening just outside the frame.

A Likely a real photo answer says nothing about when or where the picture was taken. For news pictures, check the date and the place as well, ideally with the reverse image search above.

Once you have an answer to "is this photo AI?", how you pass it on matters as much as the result, because a wrong label can stick to a real person or a real event long after it is corrected.

• Quote the range, not a certainty. "The checker says likely AI" is honest. "This is 100% fake" is not, and no detector can back it up.
• Take Can't tell at face value. Don't keep re-uploading cropped versions until you get the answer you wanted.
• Sharing is optional. If you share a result, the link can show a small thumbnail of the picture; you can switch that off. If you don't, we keep only a hash of the image and the verdict. Result pages are not indexed by search engines by default.

Because a screenshot is a new file made by your phone. Credentials belong to the original file and don't carry over. None of the 10 screenshot versions of AI images in our test had any.

Often, yes. Detection models judge the pixels, not the metadata. In our screenshot pairs, Sightengine still recognised 9 of 10 and Hive 8 of 10. Some signal is lost, so give a screenshot result a little less weight than one from the original file.

Not necessarily. Can't tell means the signals were weak or pointed both ways. Look for the first post, a larger copy, or a statement from whoever made the picture.

No. Your browser reads credentials and metadata locally. If the picture needs a detection model, it passes through our server only for that check and is not written to disk. We keep a hash of the image and the verdict, plus a small thumbnail only if you share the result with the thumbnail option on. See the privacy page.