How to tell if an image is AI generated: 5 checks, ranked
To tell if an image is AI generated, read its Content Credentials, look for a maker's watermark and trace its source first; your eyes and detectors only hint.
8分で読めます
目次

To tell if an image is AI generated, look for evidence in this order: a label inside the file, a hidden watermark its maker can read, other copies of the picture online, the picture itself and finally a detector's estimate. The first two can settle the question. The other three only lean one way or the other.
Each check looks for signs of AI, and none of them can show that a picture is real, so an empty result means you don't know yet. Every outside source linked below was checked in November 2026.
How do you check if an image is AI generated?
Here's how to spot AI images, strongest evidence first. Watch the last column, because many pictures you'll want to check reach you as a screenshot or as a copy a chat app has already re-saved.
| Check | If it finds AI | If it finds nothing | On a screenshot or chat-app copy |
|---|---|---|---|
| 1. A label in the file (Content Credentials or an AI label in the metadata) | A record signed by a known company settles it. An unsigned label is strong but can be typed in by hand. | Nothing. Labels get lost easily. | Usually gone |
| 2. The maker's hidden watermark (Google's SynthID and others) | That company's AI made or edited it. | Only that this company's AI didn't. Any other AI is still possible. | Often still there, Google says |
| 3. Reverse image search | An earlier copy or a source that says how it was made | Nothing either way | Works on any copy |
| 4. Looking closely | A reason to keep checking | Nothing. Good generators leave few slips. | Works on any copy |
| 5. An AI detector | An estimate that leans AI | A low estimate, which can be wrong | Still works; it reads only the pixels |
Work down the table and stop at the first real hit. Checks 1 and 2 look for marks that the app, camera or company behind a picture put there on purpose, whereas a detector's model reads the pixels and its number is an estimate. So it comes last, once the others have come back empty.
Can you tell if a screenshot is AI generated?
Partly. A screenshot is a brand-new file, so whatever the original carried in its metadata stays behind. Uploading does much the same: EXIF metadata, C2PA tags and visible watermarks commonly get stripped when images go up on social networks, messaging apps or marketplaces, as Sightengine's detector page notes.
Every check except the file's label still works on a copy. Google's SynthID watermark sits in the image itself, and Google says it "will usually still exist" after an image is re-scaled, re-colored or compressed; for a screenshot, crop tight around the picture before you check it (Gemini Apps Help). A reverse image search can lead you back to an earlier copy, and a detector still reads the pixels, which is all a screenshot has left to give. If you can, ask whoever posted it where the picture came from.
How do you check an image's Content Credentials and metadata?
Content Credentials are a signed record of how a file was made, written into it by the camera or app that produced it. The standard comes from the Coalition for Content Provenance and Authenticity (C2PA), and contentcredentials.org names Adobe, Google, Microsoft, OpenAI and Sony among the companies behind it. A record shows how the file was created and its editing history. To read one, open the file on the verify page at contentcredentials.org.
A simpler kind of label also exists. The IPTC Digital Source Type list has a code for "Created using Generative AI" and another for "Edited using Generative AI". That label is a plain metadata field with no signature behind it, so it can be added or removed by hand. Our AI image detector reads this label and Content Credentials before anything else.
Camera credentials need a second look. Content Credentials are "used for both AI and non-AI content" (Gemini Apps Help), so read what a record says before you count it as a sign of AI. A camera's record describes how the photo was taken.
| App | Images downloaded | With a trusted signature |
|---|---|---|
| ChatGPT | 19 | 19 |
| Gemini | 20 | 20 |
So a ChatGPT or Gemini picture you download yourself carries its signed record, at least until an app or a screenshot saves a new file without it.
Can you find a hidden AI watermark like SynthID?
Yes, if you're using the right checker. SynthID is Google's invisible watermark, and Google DeepMind says it's added the moment an image is made and is designed to stand up to cropping, filters and lossy compression. Google's SynthID Detector also recognizes images from partners including OpenAI, NVIDIA and Kakao. Pick a checker by who you think made the picture.
| Checker | Whose AI it recognizes | What it asks of you | Where your image goes |
|---|---|---|---|
| Gemini app | SynthID from Google's AI only; Content Credentials from products on the C2PA Conforming Products List | Sign-in, one file up to 100 MB, about 10 image checks per 24 hours; on the web and in the Android app so far | You upload it to Gemini |
| SynthID Detector | Google's AI and partners including OpenAI, NVIDIA and Kakao, per Google DeepMind | One file (image, audio or video); it opens a Google sign-in | You upload it to Google |
| OpenAI's checker | ChatGPT, the OpenAI API and Codex | One file at a time, no sign-in on the page; don't crop or convert it first | You upload it to OpenAI, which says it isn't stored unless legally required |
| LeanImg AI image detector | Our own pattern check for the ChatGPT and Gemini marks, separate from Google's and OpenAI's checkers | Nothing, no account | Stays in your browser tab |
| Files | Mark found |
|---|---|
| ChatGPT originals | 19 of 19 |
| ChatGPT copies, re-saved | 73 of 76 |
| Gemini originals | 19 of 20 |
| Gemini copies, re-saved | 0 of 60 |
That last row is a limit of our check. Google's own checkers are built to read SynthID, which Google says usually survives compression, so take a re-saved Gemini picture to the Gemini app.

Does a reverse image search show if a picture is AI?
Sometimes, and it works on any copy. On a computer, go to google.com, click Search by image and upload the file, or in Chrome right-click the picture and choose Search with Google Lens. On an Android phone, open the page in the Google app or Chrome, touch and hold the picture and tap Search image with Google Lens. The results include websites with the same or a similar image (Google Search Help). That sends the picture to Google, so skip it for anything private. AI pictures tend to turn up in fewer places than real ones, and the search can lead you to the original poster, whose hashtags or name may give it away (BBC Bitesize). You're looking for the oldest copy and who put it up. A news agency's archive and an account that posts nothing but generated art point in very different directions.
What does an AI-generated photo look like?
Often like any other photo, which makes looking the weakest way to identify AI generated images. The slips worth hunting for are garbled text, shadows and lighting that don't agree, patterns that repeat unnaturally and errors in small details such as hands, teeth or the background (Gemini Apps Help).
BBC Bitesize put the catch plainly in its October 2025 update: the best AI images are now "almost indistinguishable from real photos". Treat a slip as a reason to keep checking, and keep in mind that a picture without any slips can still be AI generated.
How accurate are AI image detectors?
Useful as one input, and wrong often enough that it shouldn't decide anything alone.
The three detector sites in Google's top five US results (November 2026) don't say much about their own testing. DeepAI's detector says TruthScan's engine powers it and gives no test set or method, and WasItAI doesn't publish an accuracy figure at all. Sightengine points to an outside benchmark by researchers at the University of Rochester and the University of Kansas and says no detector is fully reliable.
In October 2026 we put our detector next to Hive, Sightengine and TruthScan's engine, file for file.
| Detector | AI images caught | Real photos it called AI |
|---|---|---|
| TruthScan's engine | 110 of 113 | 3 of 68 |
| Ours, same files | 109 of 113 | 3 of 68 |
| Hive | 56 of 58 | 0 of 26 |
| Ours, same files | 58 of 58 | 2 of 26 |
| Sightengine | 43 of 44 | 0 of 22 |
| Ours, same files | 41 of 44 | 0 of 22 |
There isn't a row in that table that's perfect in both columns, ours included.
| Test set | Result |
|---|---|
| 114 AI images | 110 caught; 4 missed, 3 of them deepfakes |
| 69 real photos | 3 shown as 50% or more likely AI |
Our AI image detector reads the file's labels before its model runs and does both on your own device, whereas DeepAI, WasItAI and Sightengine ask you to upload the picture.

Questions people ask
Can ChatGPT identify an image?
OpenAI offers a separate checker for that. OpenAI's verification page looks for C2PA metadata or SynthID watermarks in images made with ChatGPT, the OpenAI API or Codex. Google's chatbot isn't any broader: Gemini "can currently only recognize content created by Google AI tools" (Gemini Apps Help).
How can you tell if a picture was created by ChatGPT?
Check its Content Credentials. We downloaded 19 images from ChatGPT, and every one had a trusted signature, so an untouched download can be read on the verify page at contentcredentials.org. For a re-saved copy, try OpenAI's checker or Google's SynthID Detector, which recognizes OpenAI images according to Google DeepMind; our own pattern check found the ChatGPT mark in 73 of 76 re-saved copies.
Is there an app that detects AI images?
Yes, several. The Gemini app checks for Google's watermark and for Content Credentials once you sign in (Gemini Apps Help), while detector sites such as DeepAI, Sightengine and ours give an estimate from the pixels. The tables above show what each one covers. Every detector's percentage is an estimate, so it's best read after the file's own evidence.
Can you tell if a photo has been AI edited?
Sometimes, from the file. An app that follows the IPTC list can write "Edited using Generative AI" into the metadata. SynthID covers images "generated or edited specifically by Google's AI models", and some results come back unclear as "Likely too small an edit" (Gemini Apps Help). Without a label or a watermark, a small AI edit doesn't leave anything these checks can read.
How to check if a photo is real or fake?
You can't prove real. The most a photo can carry is a camera's signed Content Credentials, and the most you can find is an earlier copy from a source you trust. If neither turns up, the fair answer is that you don't know.


