Check if image is AI from metadata, not pixels

Check if image is AI, on this page, means read the file’s own account of itself. Press Inspect only. If the report shows a C2PA Content Credentials manifest, an IPTC or XMP source type, or generator text such as a prompt, those records are in this file. If it shows none of them, the check is over. This is not a pixel-level detector, and it will not invent a confidence score to fill the gap.

What people mean by the question

The phrase covers at least four different jobs, and only one of them is something a metadata tool can finish. The first job is “does this file contain a generator’s paperwork?” That is a read of C2PA, IPTC, XMP, and text chunks. The second is “do the pixels carry an invisible watermark?” That is a different system, and this site does not run one. The third is “will Instagram, LinkedIn, or another app label the upload?” A platform can use the paperwork, a watermark, or a classifier you never see. The fourth is “did a person hold the camera?” No file record answers that once the paperwork is missing, and a pretty photograph is not evidence either way.

This page does the first job and says so in the report. It names the other three so you do not treat a quiet inspect as a certificate of authenticity. If you already know you only want the manifest line, the narrower page is check for C2PA Content Credentials. If you want the walkthrough that pairs the AI rows with GPS and camera EXIF before a strip, use how to check a photo for AI metadata.

How to check if image is AI with this viewer

The card above is the same inspect the rest of the site uses. The default mode is AI only, so a later clean targets C2PA and known AI records on JPEG, PNG, and WebP and can leave camera EXIF and GPS. You do not have to clean to finish the check. Inspect is enough.

1. Drop the original still. A screenshot is a new file. It often lacks the manifest the original had. Inspect the file you will actually publish.
2. Press Inspect only. The field report groups GPS, camera and EXIF, XMP, IPTC, C2PA, and AI markers. For this question, read C2PA and the AI rows first.
3. A C2PA row means a signed manifest is in this file, with the size the reader measured. It does not name every action inside the manifest in the way a full C2PA validator would.
4. An AI row means the text matched a known source type or generator marker: DigitalSourceType, a trained-algorithmic flag, or PNG keys such as parameters, prompt, and sd-metadata.
5. Camera EXIF and GPS do not prove a human took the picture, and they do not prove the opposite. A generated file can be saved with a camera profile. A real photo can carry a manifest after an editor export.

The list is not an ExifTool -a dump. MakerNotes stay out of it. For every tag on a local machine:

exiftool -a -G1 -s photo.jpg

Public C2PA verifiers answer a related, narrower question: is there a manifest, and does its signature check out? That overlaps the C2PA row here. It still does not classify pixels, and it still does not tell you what a platform will print on the upload. Use one or both reads. Do not add their results together into a score this site does not compute.

Metadata check versus a pixel detector

A pixel detector, if you use one elsewhere, looks at the picture. A metadata check looks at the records beside the picture. They fail in different ways. The table is the honest split. This site only does the left column.

We do not publish a detection rate, a false-positive rate, or a count of images tried against a pixel model. Those numbers are not a product this site has.

What a generator may leave in the file

Generators do not write the same paperwork, and a single product can change what it writes. The notes below match the limits already stated on each generator page. They are not a promise that today’s download matches last month’s.

• ChatGPT and DALL·E downloads can carry a C2PA manifest. The ChatGPT page also says SynthID, when it is in the pixels, stays after a metadata clean.
• Gemini and Imagen newer exports can carry C2PA. SynthID is not metadata. This check will not show it.
• Adobe Firefly downloads can carry Content Credentials. A cloud copy of that credential, if Adobe published one, is not inside the file this reader scans.
• Photoshop Generative Fill can put credentials on the export. Stripping the file does not delete a cloud record.
• Midjourney exports sometimes show IPTC and a prompt. A C2PA manifest is often absent. Absence is not proof the file is a photograph.
• Stable Diffusion PNGs can store a parameters chunk, and ComfyUI can add a prompt and a workflow. Those text chunks are metadata. They are not a watermark.
• A real photo edited in Lightroom, Canva, or a similar tool can pick up a manifest or an AI source type without being a fully generated image. See Lightroom and Canva.

If you need to check if image is AI only for one of those sources, open that page after the inspect. This page stays the general read: paperwork in, verdict out of scope.

When the metadata has already been removed

Once the records are gone, you cannot check if image is AI with this tool. A strip, a screenshot, a messenger re-save, or an export that never wrote the tags all look the same in the report: no row. The After list on a file you just cleaned will say the rows are gone. That confirms the clean. It does not become a new answer about the pixels.

A screenshot usually drops C2PA because it builds a new image. It does not remove SynthID. The limits of that shortcut are on does a screenshot remove AI metadata. Do not screenshot a file in order to “pass” a check. The new file is simply missing the records, and a platform can still label it.

If you cleaned with AI only on a JPEG, PNG, or WebP, camera EXIF and GPS can remain while the C2PA and AI rows go. That leftover camera block is not evidence the picture is a photograph. If you cleaned with All metadata, the camera block goes too. On HEIC, AVIF, GIF, TIFF, and BMP, every metadata field this cleaner lists is removed in any mode, so a “keep the camera” plan does not apply. AVIF comes back as a JPEG or a PNG.

SynthID and other marks in the pixels

SynthID is an invisible watermark in the picture, not a tag beside it. TrustMark is the same kind of object. The inspect report has no row for either. Cleaning metadata does not scrape them out. We do not offer a removal method, a script, or a sequence of edits that claims to, and this page will not estimate how often a crop or a re-save would damage a mark. That honesty is the whole point of can you remove SynthID.

A vendor may publish a detector for its own watermark. Using that detector is a separate decision on someone else’s site. We do not proxy it, we do not quote a result we did not produce, and a missing label on a social network is not a substitute for it. If the only thing you wanted was a watermark test, stop after Inspect shows no metadata rows, and do not treat the empty list as the test.

What to do after you read the report

The report is an inventory. The next step depends on which job you actually had.

• You wanted to know whether this file still carries generator paperwork. You are done. Keep the original if you need those records later. A clean destroys the evidence this page can read.
• You are about to publish and you do not want GPS or a manifest in the copy you upload. Switch mode if you need to, press Clean, and read the After list. One image is free. The paywall on the next file is the existing one.
• You hoped a clean file would force a platform to drop a label. It will not. Instagram AI info, the LinkedIn CR icon, and similar marks are explained on their own pages, starting at Instagram and LinkedIn Content Credentials.
• You need a disclosure because a platform or a law asks for one. Stripping the file is not that disclosure. This site does not tell you to hide that a picture was generated.

Related reads that stay in their lane: the metadata viewer for the full field table, what AI metadata is for the vocabulary, and remove C2PA if the manifest is the row you intend to take off the next copy.

Questions

Is this an AI image detector?

No. It reads metadata: a C2PA manifest, an IPTC or XMP AI source type, and generator text such as a prompt or a workflow. It does not score the pixels. It does not say “this is AI” or “this is a camera photo” as a verdict.

If the report is empty, is the picture real?

No. An empty report means this reader did not find those records. The file may have been exported without them, screenshotted, or cleaned already. Pixel watermarks and a platform’s own classifier are invisible here.

Can it see SynthID or TrustMark?

No. SynthID and TrustMark sit in the pixels. This page cannot display them, measure them, or remove them. A clean After list does not mean a watermark is gone.

Which files can it read?

JPEG, PNG, WebP, HEIC, AVIF, GIF, TIFF, and BMP, in the browser. PDF uses the PDF page. Word uses the Word page. Video and RAW are not opened on this card.

Does the check upload the image?

No. Inspect runs in the tab. It does not use the free clean. Clean one file free does use that one clean, which is shared with the other tools on this site.

Will Instagram or another app agree with the report?

Not necessarily. A platform can label an upload from metadata, from a pixel watermark, or from its own systems. A manifest in the file can explain a label. A missing manifest does not promise that the label will stay off.

What should I open instead if I only care about C2PA?

The C2PA checker reads the manifest line and stops there. This page is the wider question: generator metadata plus that manifest, and an explicit refusal to pretend the pixels were classified.