To tell if an image is AI generated, check its metadata for Content Credentials or an AI source label, look for real camera data, reverse search it and zoom in on text, pairs and reflections. No single test is reliable. Missing metadata proves nothing, so trust the answer when several independent signals agree.
Sometimes, with confidence. Often, only with a "probably." No test works on every image, and anyone selling a detector that claims 99% accuracy on anything you throw at it is overselling.
What you can do is stack evidence. Each check below either finds something or comes up empty, and they fail in different ways. Provenance data can prove where an image came from, but it's easy to strip. Visual tells are always there to look at, but newer models make fewer of them. When several independent checks point the same way, you can trust the answer much more than any single one.
Work through them roughly in this order, cheapest and most decisive first.
| Signal | What it can show | How reliable | How to check |
|---|---|---|---|
| C2PA Content Credentials | A signed record of which tool made or edited the file | Strong when present. Absence proves nothing | Upload to a Content Credentials verifier |
| IPTC "digital source type" field | A label like trainedAlgorithmicMedia written by the generator | Good when present, but anyone can add or remove it | Any metadata viewer that shows XMP |
| Invisible watermark (SynthID) | That a Google model, or another tool using it, made the image | Survives cropping and compression better than metadata | Only the vendor's own checker can read it |
| Camera EXIF (make, model, lens, GPS) | That a camera or phone probably captured it | Moderate. EXIF can be copied or faked | Our AI image detector or an EXIF viewer |
| Visual artifacts | Rendering errors typical of generators | Weak and getting weaker with each model | Zoom in and look carefully |
| Reverse image search | An earlier original, or the account that first posted it | Strong for recycled or edited photos | Reverse image search across several engines |
| AI-detector score | A statistical guess from pixel patterns | Inconsistent across tools and image types | Treat as one vote, never a verdict |
Start with the file itself, because it's quick and sometimes settles the question outright. There are two kinds of metadata to look for.
Provenance labels. Major generators now write labels into the files they produce. The photo industry's IPTC standard defines a "digital source type" field, and its value trainedAlgorithmicMedia means the image was fully generated by an AI model. Meta has said its image generator writes that field. OpenAI adds C2PA Content Credentials to images from ChatGPT and its API. That's a cryptographically signed manifest recording the tool and any edits. If you find either one, you have your answer about where the file came from.
Camera data. A photo straight off a phone or camera usually carries dozens of EXIF fields: make and model, lens, exposure, a capture timestamp, often GPS. AI generators don't produce this. So rich, consistent camera data points toward a real capture.
Our AI image detector reads EXIF and XMP metadata right in your browser, without uploading the file. It reports camera make and model, any software tag, a capture timestamp, GPS coordinates and the number of metadata fields present. It doesn't validate C2PA signatures. For that, use a dedicated Content Credentials verifier such as the one at contentcredentials.org.
Almost nothing on its own. That's the part most people get wrong.
Metadata disappears constantly for innocent reasons. Taking a screenshot creates a new file with none of the original's data. Most social platforms strip EXIF when you upload, partly for privacy. Messaging apps often re-compress images, and many editing apps drop fields when you export. People also strip it on purpose to hide their location, which is exactly what our image metadata remover is for.
So here's how to read what you find:
| What you find | Most likely explanation | Next step |
|---|---|---|
| C2PA manifest naming an AI tool | Made or edited with that tool | Check whether it says generated or just edited |
| IPTC source type: trainedAlgorithmicMedia | Fully AI-generated | Treat as AI unless you have strong evidence otherwise |
| Rich camera EXIF (15+ fields, make, lens, timestamp) | Real capture, possibly edited later | Check the software tag and look for edits |
| Software tag naming an editor only | Edited photo, origin unclear | Reverse search for the original |
| No metadata at all | Screenshot, social media download, stripped file, or AI | Visual checks plus reverse search |
The bottom row is where most viral images land. If you pulled it off social media or someone sent it in a chat, expect it to have nothing.
SynthID is Google DeepMind's invisible watermark. It's woven into the image pixels rather than stored in metadata, and it's built to survive cropping, filters and lossy compression that would wipe out an EXIF tag. It's used across Google's generative products, and OpenAI's help center says images from ChatGPT and its API now carry a SynthID watermark alongside the C2PA manifest.
The catch is that only the vendor can read it. Google runs a SynthID Detector portal for journalists and researchers and has been adding SynthID checks to its own products. No generic metadata tool can see it, including ours. A negative result also only means "not watermarked by a SynthID-using tool." It doesn't rule out Midjourney, an open-source model running on someone's laptop, or anything else.
Older generators were easy to spot from six-fingered hands and melted faces. Current models rarely make those mistakes, so the tells have moved to the details. Zoom in and check:
Every one of these can show up in a real photo too, thanks to motion blur, heavy compression or aggressive beauty filters. One odd earring doesn't prove anything. Three independent oddities start to.
Pixel-level detectors, the ones that give you a percentage, are the least dependable tool in the kit. They're trained on output from specific models, so a new generator can slip past them. Compression, resizing and screenshots all degrade what they're looking at. Real photos with heavy editing or smartphone "AI enhancement" can get flagged, too. Run the same image through three detectors and it's not unusual to get three different answers.
Use them as one vote. If a detector says 95% AI and the image also has no camera data, garbled background text and no earlier source online, that's a coherent case. If it's the only evidence, you don't know.
Check the file's metadata for C2PA Content Credentials or an IPTC source type of trainedAlgorithmicMedia, look for camera EXIF data, run a reverse image search, and zoom in on text, hands, jewelry and reflections. Free tools cover every step. No single check is conclusive, so look for several signals that agree.
No. Screenshots, social media uploads, messaging apps and many editing apps all strip EXIF data, so most images shared online have none. Missing metadata is consistent with AI but also with ordinary sharing. Rich camera EXIF is more informative, because generators don't produce it.
Metadata labels like C2PA and IPTC fields can be removed by a screenshot or a re-save. Invisible watermarks such as Google's SynthID sit in the pixels and are built to survive cropping, filters and compression, but no watermark is guaranteed against deliberate removal.
Not on their own. Pixel-based detectors are trained on specific generators, struggle with new models and compressed files, and sometimes flag heavily edited real photos. Treat a detector score as one piece of evidence alongside metadata, reverse search and a close visual check.
C2PA is an open standard for Content Credentials: a signed record embedded in a file showing which camera, app or AI tool created it and what edits were made. OpenAI and Adobe add it to AI-generated images. You can inspect it with a Content Credentials verifier, but it disappears if the image is screenshotted.