Is This Image AI Generated? How to Check

A photo's own metadata can show that an AI tool made or edited it, when a C2PA Content Credential or an IPTC tag says so. Missing tags prove nothing, because metadata is easy to strip and plenty of apps never write it. Invisible watermarks can only be read by their owners' checkers, and detectors that scan pixels give a probability, not proof.

Check and scrub your file here

It runs in your browser. Nothing is uploaded, and the result is lossless.

Choose photos or videos

Drop files here or paste an image. Photo on your phone? Open scrubphoto.com there.

No upload. Works offline once the page loads.
Lossless. Every pixel kept, byte for byte.
Free. No sign-up, no limits.

No photo handy?

Is this image AI generated? What the file itself can tell you

Two parts of a file's hidden record matter most here.

The first is C2PA Content Credentials, a signed block that names the app or device that wrote it and can list what was done to the file. The C2PA's explainer says AI actions are identified through a field called the digital source type. The Content Authenticity Initiative says the signature makes credentials tamper-evident, which is why a valid manifest from an AI service is the strongest evidence a file can carry.

The second is the IPTC digital source type, a plain field in the file's XMP data. In May 2023 the IPTC published guidance asking software that generates images and video to write the value trainedAlgorithmicMedia there. It is a strong sign when an AI app wrote it, but any metadata editor can change a plain field, so treat it as a claim rather than a seal.

Platforms read the same tags. Meta labels a post when it finds C2PA or IPTC AI indicators, which on Instagram shows as the AI info label (Instagram has also shown it as AI content, and first called it Made with AI). The Instagram page covers how that label works.

How to check if a photo is AI, step by step

Get the closest thing to the original file. OpenAI's developer docs say editing, converting or sharing a file can remove Content Credentials, so a re-shared copy may say less.
Open it in the checker. It reads the file in your browser, uploads nothing and changes nothing.
Read the C2PA line: whether Content Credentials are present, and which app or device signed them.
Read the IPTC digital source type, using the table below.
Look at the XMP creator tool and any PNG text. The metadata viewer shows every field.
If the tags are missing, try the official watermark checkers listed further down.
Treat pixel scores and visual hunches as hints, not answers.

AI metadata checker results: what each value means

The C2PA signer

The signer tells you who vouched for the record. If it is an AI service like OpenAI, the image came out of that service. If it is an editor like Photoshop, the editor saved the file, and the source type and action list show whether AI was used. If it is a camera or phone, the file claims a capture. Google said in September 2025 that the Pixel Camera app on Pixel 10 attaches Content Credentials to any JPEG photo it captures.

The checker reads the signer's name but does not validate the signature. For that, use the CAI's Verify site, which the CAI describes as an inspection tool. The C2PA guide explains what a manifest holds.

The IPTC digital source type

These values come from the IPTC's digital source type list, with the IPTC's own names:

Value	IPTC name	What it means
trainedAlgorithmicMedia	Created using Generative AI	An AI model made the image
compositeWithTrainedAlgorithmicMedia	Edited using Generative AI	AI changed part of it, like inpainting
compositeSynthetic	Composite including generative AI elements	A mix where at least one piece is AI
digitalCapture	Digital capture sampled from real life	A camera or recorder captured it

The first three point to AI, and the last describes a capture, if the field is telling the truth.

Creator tool and PNG text

XMP CreatorTool names the app that saved the file, which hints at a workflow but says nothing about AI on its own. PNG text is more direct. Stable Diffusion WebUI saves a "parameters" key, and ComfyUI saves "prompt" and "workflow" keys, often with the full prompt inside.

Why a missing AI tag proves nothing

Metadata is the easiest part of a file to lose. Google DeepMind wrote in 2023 that metadata can be removed by hand or lost when files are edited. The C2PA explainer asks whether provenance data can be removed and answers that it can.

Many tools also never write it. Google Photos' help page says Content Credentials are not available for every photo or video. A Google post on Pixel names the risk: research shows that when only synthetic content is labeled as AI, people falsely believe unlabeled content is not AI. So "no metadata trigger found" means the file declares nothing, not that a person made it.

Check for watermarks with Google and OpenAI's own checkers

Some AI makers hide a watermark in the pixels, out of reach of any metadata reader (ScrubPhoto included). Only the company behind each watermark can check for it.

Gemini. Sign in on the web or the Android app, add the file, and ask "Was this image created or edited by Google AI?" Google's help page says it checks images, video and audio for SynthID, one file at a time, and summarizes any Content Credentials. Detected means all or part was made or edited by Google's AI. Not detected rules out Google AI, not other AI.
SynthID Detector. Google DeepMind's SynthID page calls it a verification portal, launched in 2025 for journalists and media professionals, with access expanding to everyone. It names OpenAI, NVIDIA and Kakao as partners whose AI content it can check.
OpenAI's checker at openai.com/verify. OpenAI's developer docs say it looks for SynthID and for a trusted C2PA manifest naming OpenAI. A not-detected result, OpenAI says, does not establish that a person made the content, and other companies' models are not covered.

Meta AI images also carry an invisible watermark alongside IPTC metadata, so stripping their tags does not remove every trace. The SynthID guide explains why watermarks sit outside any metadata tool.

What AI image checkers that scan pixels can and cannot tell you

Many tools sold as an AI image checker score the pixels instead of reading the file. The percentage they return is a probability from a trained model, not a reading of anything the file declares. A real photo can score as AI, and an AI image can pass.

The Content Authenticity Initiative says detection results can be unreliable, and that detectors are in a permanent arms race with bad actors. If a detector quotes an accuracy figure, look for the test data behind it.

ScrubPhoto is not one of these tools. It reports what a file says about itself, and the limits page lists what that leaves out. If you post AI work yourself, Meta requires you to disclose photorealistic video and realistic audio that was digitally created or altered, whatever the file says.

Questions about checking if a photo is AI

Can metadata prove a photo is AI-generated?

It can show that an AI tool made or edited it. A valid C2PA manifest signed by an AI service is the strongest case, and an IPTC value like trainedAlgorithmicMedia is a strong but editable sign. Metadata can never prove the opposite.

Is the ScrubPhoto checker an AI image detector?

No. It is a file parser that reads C2PA, IPTC, XMP, EXIF and PNG text and reports what they say. It never judges the pixels and cannot see watermarks.

Why does a real photo say edited with AI?

The app that edited it wrote compositeWithTrainedAlgorithmicMedia or a matching C2PA entry, for example after an inpainting edit that filled in part of the picture. The photo is still a real capture underneath, and the tag records the AI step.

Does a screenshot keep a photo's AI metadata?

No. A screenshot is a new file that the phone or computer makes, so it carries none of the original's tags. Google's Gemini help page asks for screenshots to be cropped tightly when you check them for a watermark.

Can I check if a photo is AI for free?

The checker on this site is free and needs no upload or account. Google's Gemini check needs a signed-in account and has a daily limit, and OpenAI's verify page covers OpenAI images.

Sources

Official pages this page was checked against in October 2026: