How to tell if an image was made with AI
Start with what the file itself declares: the tool above reads its Content Credentials, IPTC source type and software fields.
Next, check for a watermark with the AI company’s own checker: Gemini for Google’s AI tools, and the web tool OpenAI provides for checking files. Then weigh what’s missing. Editing, converting or sharing a file can remove its metadata, so a missing signal says nothing either way.
What this checker reads
Content Credentials are a signed record, based on the C2PA standard, of how a file was made and changed. From them, the checker reads the software that wrote them, the signer named in their certificate, and the actions and ingredients recorded. An action can carry a digital source type from IPTC’s vocabulary, such as trainedAlgorithmicMedia for media made by a trained AI model.
IPTC’s Photo Metadata Standard defines the same field for XMP, where it isn’t signed. Some local image generators save their generation settings as text inside the PNG: AUTOMATIC1111 in a “parameters” chunk, ComfyUI in “prompt” and “workflow” chunks. The checker reads those settings too, and says when the Software field in Exif, the Creator tool field in XMP or the software named in Content Credentials mentions a known AI tool. See every metadata field in the file in the viewer.
Why there’s no score
This checker doesn’t analyze the pixels, so it has nothing to estimate. It reports statements the file carries, which are there or not, leaving nothing to put a number on. A declaration is only what the software that wrote it put there, and fields other than Content Credentials carry no signature.
What the results mean
This file declares it was made with AI and This file declares AI was used to edit it appear when an action or ingredient in the Content Credentials, or the IPTC field in XMP, uses trainedAlgorithmicMedia, compositeWithTrainedAlgorithmicMedia, compositeSynthetic or trainedAlgorithmicData, or when an action is described as “AI Generated Image”.
This file contains AI generator settings means it holds settings in a known AI generator’s format.
This file names an AI tool in its metadata means a software, creator tool or credit field, or the software named in its Content Credentials, mentions a known AI tool, or its IPTC fields naming an AI system or prompt are filled in, and nothing this page read declares AI.
This file has Content Credentials, and they don’t declare AI can describe a camera photo: Google’s Pixel Camera adds Content Credentials to photos taken in some modes on Pixel 10 phones. Credentials giving the source as algorithmicMedia land here too, though Google Merchant Center lists that value among those for generative AI images.
This file has Content Credentials, and the parts this page could read don’t declare AI means a part that could hold a declaration couldn’t be read.
This file has Content Credentials this page couldn’t read covers credentials that are damaged, present more than once, compressed in a way this page can’t unpack (Brotli in some browsers, or too large once unpacked), or found but not read because reading stopped early.
This page couldn’t read all of this file’s metadata means the read stopped early or skipped a part that could hold a declaration.
No AI signal found. That doesn’t mean the image is real means none of the above turned up.
Why “no AI signal” isn’t proof
Editing, converting or sharing a file can remove its metadata. In IPTC’s 2019 test, images saved from Twitter had their metadata stripped and images saved from Facebook had no XMP, while Google Drive, Dropbox and OneDrive passed the original file through. Platforms might also remove C2PA manifest data when their software doesn’t support the standard yet, as the documentation for the open-source Content Credentials tools notes.
What survives isn’t proof either: fields other than Content Credentials carry no signature, so any of them can be written by software that has nothing to do with how the image was made.
What “Named signer” means
The named signer is the organization, or failing that the common name, in the certificate the credentials name for their signature, and this page doesn’t check that the signature matches it. Content Credentials carry a signature from the software or service that wrote them, so the named signer is usually the company behind that software, such as Microsoft Corporation for Azure OpenAI images, rather than the person who made the image.
OpenAI, Adobe (for Firefly), Microsoft (for Azure OpenAI) and Google (for Nano Banana Pro images in the Gemini app) say their generated images carry Content Credentials. Checking their signatures takes a validator, such as the Content Credentials Verify page.
SynthID and other invisible watermarks
SynthID is a watermark embedded in the content itself rather than stored as metadata, so this page doesn’t check it or any other invisible watermark. Gemini checks images for SynthID and Content Credentials once you’re signed in, and Google says that while other companies have started to adopt SynthID watermarks, Gemini can currently only recognize content created by Google AI tools. OpenAI provides a web tool for checking files in the browser.
Who else reads these fields
Meta has said it labels images on Facebook, Instagram and Threads when it finds AI indicators from the C2PA and IPTC standards. TikTok has said it uses Content Credentials to label AI-generated content uploaded from certain other platforms. Pinterest says the labels it puts on generative AI content draw on a Pin’s metadata following the IPTC standard, among other signals. Google Merchant Center requires the IPTC digital source type on AI-generated product images.
Why nothing is uploaded
This page’s script reads the file in the browser tab, never sends it anywhere and makes no network request, and the tool keeps working with the network off once the page has loaded. If the image is yours, you can also remove location and camera details before sharing.
