Is This Real or AI? How to Check a Picture You Saw Online

To answer "is this real or AI?" about a picture, add it to the checker above: upload it, drag it in, paste a screenshot with Ctrl+V or paste an image link. Your browser first reads what the file says about itself, its Content Credentials and metadata, right on your device. If that doesn't settle it, a detection model judges the pixels: Sightengine, with Hive as a backup.

The answer is one of four: Likely AI-generated, Likely a real photo, Can't tell, or Edited with AI, each with a probability range and the reasons. The table above shows how far each answer could be trusted when we ran the same flow on 100 pictures whose origin we knew:

• Likely AI-generated came up 49 times and was right every time. None of the 50 real photos got it.
• Likely a real photo came up 50 times and was right 49 times. The one miss was a screenshot of an AI wedding photo.
• Can't tell came up once, for a screenshot of a real photo. It means the signals were weak, and it is an honest answer, not a failure.

Notice where the doubt sits: every answer that was wrong or unsure came from a screenshot. On the 86 files that weren't screenshots, the real or AI answer was right every time in this test. It is a small test, so read it as a good sign rather than a guarantee; the accuracy report has the method and every result. You get 3 free checks a day, no sign-up.

The strongest answers to "is this real or AI?" don't come from a model at all. Some AI tools sign the images they make with Content Credentials, a record in the C2PA format that names the company and says AI made the picture. Some cameras do the same for real photos. When the signer is on the C2PA trust lists and the file hasn't changed since signing, the checker answers from the record alone: that settled 14 AI images and 2 real photos in our test, all correctly.

The two real photos came from a Google Pixel 9 and a Google Pixel 10. Google says every photo taken with Pixel Camera on the Pixel 10 carries Content Credentials, so an unedited original from that phone is strong evidence that the camera captured it, though not that the scene is what it seems.

These records are fragile, though. They belong to one exact file, so a copy that has been edited or re-saved can lose them: when we cropped and brightened a copy of the Pixel 9 photo with an ordinary image tool, its credentials were gone. Google Photos, by contrast, adds Content Credentials when it edits a JPEG that already has them. Camera details such as the phone model and exposure settings are a weaker hint, because anyone can edit them. To see everything a file carries, use the Content Credentials checker or the AI metadata checker.

Most pictures that make people ask "is this real or AI?" have already been shared a few times, and a screenshot is a new file that carries none of the original's history. What is left is the pixels, which is where the detection model earns its place: it needs no metadata, and it still called 9 of the 10 AI screenshots in our test AI.

A screenshot is a weaker sample, though: the picture is smaller and wrapped in an app's interface, and shared copies are often compressed again. Both answers in the table that were wrong or unsure came from screenshots, so look for a better copy before you settle on one. The guide to viral pictures and screenshots shows how.

This checker reads pictures, not video, and we haven't tested how detection models do on single frames taken from a clip. If what you saw was a video, these are the checks that exist today:

• Ask the company that may have made it. Google's Gemini app can check a video for SynthID, the invisible watermark in content made with Google AI. You need to be signed in; upload one video at a time, under 90 seconds and 100 MB or less, with about 10 video checks in 24 hours. Gemini says which parts of the video carried the watermark. Its SynthID check only recognises Google's own AI; if the clip carries Content Credentials, Gemini also shows a summary of them.
• Look closely at the details. Google's own advice for images and videos is to look for garbled or inconsistent text, mistakes in small details such as hands, teeth or the background, inconsistent shadows and lighting, and patterns that repeat unnaturally.
• Find the first upload. Search for the earliest copy of the clip and read who posted it and how they describe it. People who make AI videos often say so.
• Check a still frame, but weigh it lightly. You can screenshot a sharp frame and check it here. A frame is a compressed screenshot and wasn't part of our test, so treat the answer as a weak signal, not a verdict on the whole video.

OpenAI's documentation for its own checker lists images and audio, not video; our guide to OpenAI image verification explains what it can confirm.

A real-or-AI answer isn't a true-or-false answer: a real photo can be old, from another place, or cropped to mislead, and an AI image can be an honest illustration that nobody claimed was a photo. Check the date, the place and the first post before you pass a verdict on (more on false captions). To train your eye first, real and AI pictures side by side show what gave each one away.

If you are on the other side of the "is this real or AI?" question, with a real photo of yours being called fake, your best evidence is the original file. An edited or re-saved copy can lose that evidence, as our retouched copy of the Pixel 9 photo did.

• Share the original, not a screenshot. The file from your phone or camera keeps its metadata, and its Content Credentials if your device adds them.
• Add the shots around it. Photos taken just before and after, each with its own metadata, back up the one in question.
• Show a check, with its reasons. None of the 50 real photos in our test was called AI, but no detector is perfect. A result that lists what the file and the model said gives people something to look at besides your word.

If it is your artwork rather than a photo, the guide for artists and judges covers proof of process.

Yes. The checker on this page gives you 3 free checks a day without signing up. Pictures settled by their Content Credentials, and pictures someone has already checked, don't use one.

Sometimes. Text, hands, reflections and lighting can give AI away, but newer generators make fewer of those mistakes, so your eyes are the weakest test. The visual checklist shows where to look and how much each clue still tells you.

No. It means neither the file nor the model found signs of AI. In our test that answer was right 49 times out of 50, and the miss was a screenshot. It also says nothing about when or where the photo was taken.

No, it checks pictures only. For a clip, Gemini can look for SynthID in videos made with Google AI, and the earliest upload and its caption often tell you more. A still frame checked here is only a weak signal.

A screenshot is a new file, so it carries none of the original's credentials or metadata, and the picture inside is smaller and often compressed again. In our test, every wrong or unsure answer was on a screenshot.