How to tell if an image is AI generated in 2026

September 27, 20268 min readBy the FreeImageTools teamMeasured and sourced

A grid showing ten common image edits and whether the C2PA and IPTC AI labels survived each one

The big AI image makers now label what they produce. OpenAI, Adobe, Google and Meta all write some kind of marker into their images, and there are standards for it. So why do most AI images you find online carry no label at all? We took two AI labelled files and ran them through ten ordinary edits to find out. The label usually did not survive the first one.

THE SHORT ANSWER

Can I tell from the file itself?: Sometimes. If a label is still inside, a checker such as our AI Image Detector will show it. Most images online have lost theirs.
Which AI tools label their images?: OpenAI and Adobe Firefly use Content Credentials, Google uses SynthID watermarks plus Content Credentials on some images, and Meta uses IPTC metadata and watermarks.
Why do so many AI images have no label?: Everyday edits delete it. In our test, the Content Credentials survived 1 of 10 common edits, a plain copy of the file.
What if there is no label?: Then the file cannot tell you. Check where the image came from, run a reverse image search, and look for the visual giveaways.

AI LABELS BY THE NUMBERS

1 of 10

edits that kept the C2PA Content Credentials, and it was a plain file copy

2 of 10

edits that kept the IPTC AI label in the file's metadata

edits that kept the camera data but dropped the AI label, making the file look like a normal photo

Nov 2025

when the Gemini app started checking uploaded images for Google's SynthID watermark

The three kinds of AI label

There is no single AI tag. There are three separate systems, stored in different places, and they fail in different ways.

Label	Where it lives	Who uses it	What removes it
Content Credentials (C2PA)	A signed manifest inside the file	OpenAI, Adobe Firefly, Google Nano Banana Pro, some cameras	Almost any re-save, and every screenshot
IPTC digital source type	A field in the file's XMP metadata	Meta AI, and software that follows IPTC guidance	Re-saving without keeping XMP, and every screenshot
Invisible watermark	Hidden in the pixels themselves	Google SynthID, Meta	Designed to survive edits, but only the maker can read it

C2PA, the Coalition for Content Provenance and Authenticity, defines a manifest that records which tool made a file and what happened to it, signed with a certificate so changes can be detected. The IPTC field is simpler: a single value such as trainedAlgorithmicMedia, the term IPTC uses for an image created by a trained AI model, or compositeWithTrainedAlgorithmicMedia for a real image with AI edits such as generative fill. Watermarks take a different route and hide a signal in the image data, which makes them far harder to strip but impossible to check without the maker's own detector.

Who labels their AI images in 2026

OpenAI has added C2PA metadata to images generated in ChatGPT and its API since February 2024, so that anyone can see an image came from its products.
Adobe automatically applies Content Credentials to images made entirely with Firefly, such as Text to Image. They name Adobe as the issuer and record the AI tool used, and may be stored in Adobe's public cloud, so some files carry only a link to them.
Google hides a SynthID watermark in images from its AI models. Since November 2025 the Gemini app can check an uploaded image for it, and images from Nano Banana Pro in the Gemini app, Vertex AI and Google Ads also carry C2PA metadata.
Meta marks photorealistic images from Meta AI with a visible label, an invisible watermark and IPTC metadata set to trainedAlgorithmicMedia.
Stable Diffusion interfaces such as AUTOMATIC1111 and ComfyUI are not labelling for the sake of transparency, but they save the full prompt, seed and settings into every PNG by default, which gives the image away just as clearly.

We tested what happens to the label

We made two AI labelled test files. The first carries Content Credentials signed with the C2PA project's own test certificate and declaring trainedAlgorithmicMedia, applied to a real photo that also has ordinary camera data. The second carries only the IPTC field. Then we put both through ten everyday operations and checked each result with our AI Image Detector.

What we did to the file	C2PA label	IPTC label	What the file looked like afterwards
Copied or renamed it, bytes unchanged	Kept	Kept	AI labelled
Took a screenshot of it	Lost	Lost	No record at all
Saved it from a browser canvas as JPEG	Lost	Lost	No record at all
Saved it from a browser canvas as PNG	Lost	Lost	No record at all
Saved it from a browser canvas as WebP	Lost	Lost	No record at all
Re-saved it with Python's Pillow library	Lost	Lost	No record at all
Resized it with Pillow	Lost	Lost	No record at all
Re-saved with Pillow, keeping EXIF	Lost	Lost	A camera photo, or no record
Converted to WebP with Google's cwebp	Lost	Lost	No record at all
Converted with cwebp, keeping all metadata	Lost	Kept	A camera photo, or AI labelled

The Content Credentials survived one operation out of ten, and it was the one that does not change the file. The IPTC label did a little better because a converter told to keep all metadata carried the XMP across. Every screenshot and every canvas save, which is how most web tools work, including the ones on this site, wiped everything.

The most misleading result

Two operations kept the camera data but dropped the Content Credentials. The AI label was gone, the camera model and exposure settings were still there, and the file now looked exactly like an ordinary photo. Partial metadata can make an AI image look more trustworthy, not less.

How to check an image, step by step

Read the file. Drop it into the AI Image Detector. It reads Content Credentials, the IPTC field, stored prompts and camera data in your browser, without uploading anything.
Verify any credentials. If Content Credentials are found, check the signature with the Content Credentials Verify site. Our detector shows who signed them but does not validate the certificate.
Ask Google about its own images. If the image may come from Google's AI, upload it to the Gemini app and ask whether it was made with Google AI. That checks for SynthID, which no metadata reader can see.
Find the original. A reverse image search with Google Lens or TinEye often turns up the first place an image was posted, and whether it was described as AI there.
Look at it properly. Zoom in on the details that generators still get wrong, listed below.

What no label means

Nothing, on its own. A missing label is the normal state of an image online, whether it came from a camera or a generator, because screenshots, messaging apps, many social networks and most editing tools remove metadata. Our detector reports such files as having no record, and that is the honest answer. It is not a verdict that the image is real.

The reverse is also true in a narrower way. Camera data proves little, because EXIF is plain text that any editor can copy from another photo. The strongest evidence a file can carry is a signed Content Credentials manifest that validates, and even that tells you what the signing software recorded, not what the scene looked like.

Visual clues that still help

Text: signs, labels, screens and jersey numbers are often garbled, especially in the background.
Hands and small parts: extra or merged fingers, earrings that do not match, glasses whose arms go nowhere.
Symmetry and repetition: faces in a crowd that look alike, patterns that repeat too perfectly.
Light: shadows pointing in different directions, reflections that do not match the scene.
Edges and backgrounds: objects that melt into each other, hair that merges with the background.

Treat these as reasons for suspicion, not proof. Each new generation of models makes fewer of these mistakes, and a real photo can have odd reflections too. The source of an image usually tells you more than any single detail in it.

If you publish AI images

Keep the label. Some platforms require it: Google Merchant Center tells sellers that product images created with generative AI must keep the IPTC digital source type set to trainedAlgorithmicMedia, and not to remove it. Since compressing or resizing through a canvas deletes that field, upload the original file from the generator, or re-add the label with a metadata editor after editing, and check the result in the AI Image Detector before publishing.

Check an image now

The AI Image Detector reads every provenance label a file can carry and shows the evidence, in your browser, with nothing uploaded. Three sample files are built in if you want to see it work first.

Open AI Image Detector EXIF Viewer

The bottom line

AI labels exist and the biggest image makers use them, but they are fragile. In our test a single screenshot or re-save removed the Content Credentials nine times out of ten, and two careful conversions left an AI image looking like a camera photo. Read the file when you can, verify what you find, and when the file is silent, fall back on its source and your own eyes.

Frequently asked questions

How can I tell if an image is AI generated?+

First read the file with a provenance checker to find Content Credentials, an IPTC AI label or stored prompts. If there is nothing, run a reverse image search to find the source, check Google images with the Gemini app, and look for visual clues such as garbled text and odd hands.

Do all AI images have a watermark or label?+

No. OpenAI, Adobe, Google and Meta add labels to many of their images, but other tools add none, and labels are usually lost when an image is screenshotted, re-saved or uploaded to many sites.

Can AI image labels be removed?+

Easily, and often by accident. In our test the Content Credentials survived only a plain copy of the file, 1 of 10 common operations, and every screenshot removed all labels.

What are Content Credentials?+

The public name for C2PA, an open standard for a signed record inside a media file that says which tools made it and how it was edited. OpenAI, Adobe and Google add them to images from some of their AI tools.

How do I check for Google's SynthID watermark?+

Upload the image to the Gemini app and ask whether it was made with Google AI. The check, available since November 2025, only detects content made or edited by Google's own AI.

Are online AI image detectors accurate?+

Tools that judge the pixels can be right often, but none is reliable enough to prove anything alone, and results vary between generators and after edits. Metadata checks are exact about what a file declares, but most files declare nothing.

Does converting or compressing an image remove its AI label?+

Usually. Browser based tools, including ours, redraw the image and save a new file without the old metadata. A converter told to keep all metadata kept the IPTC label in our test, but not the Content Credentials.

Sources and method

Two test files were made with c2patool 0.27.22 and Python: a JPEG photo with ordinary camera EXIF, signed with the C2PA test certificate and an action declaring trainedAlgorithmicMedia, and a JPEG carrying only the IPTC DigitalSourceType field in XMP. Each was copied, screenshotted and saved through a browser canvas as JPEG, PNG and WebP in headless Chrome, re-saved and resized with Pillow 10.2, re-saved with Pillow keeping EXIF, and converted with cwebp both by default and with all metadata kept. Every output was read with the FreeImageTools AI Image Detector, whose readings matched c2patool on the C2PA project's public test files apart from one file with credentials stored online and one pre-release draft.

CITE THIS ARTICLE

In a test of ten common operations on an AI labelled image, the C2PA Content Credentials survived only a plain file copy and the IPTC label survived two, while two conversions kept the camera data but dropped the AI label. FreeImageTools, "How to Tell If an Image Is AI Generated in 2026", September 27, 2026, https://freeimagetools.org/blog/is-this-image-ai-generated