Guide

How to tell if an image is AI-generated

Start with the file you actually have. A likeness, a smooth face or a suspicious caption is a reason to look, and it is a weak place to stop.

17 October 2026 · Check AI Free

People ask how to tell if an image is AI-generated after a screenshot arrives in a chat, a profile photo looks too polished, or a picture is offered as proof. Those are different files. A screenshot of a post, a recompressed download and an original export can show the same scene while carrying different embedded text. This checklist separates those cases. It does not produce a certificate, and a single misshapen hand is not a method.

Work through the file in this order

Earlier steps are more specific. Later steps cover more images and are easier to overread. Stop when you have a claim you can actually support, and write down what you still do not know.

1. Step 1

   Keep the earliest file, and name what it is

   Save the attachment you were sent before you open it in an editor. Note whether it is an original export, a download, a screen capture or a photo of a screen. A screen capture is a new picture of pixels. Our metadata experiment found camera make, model, capture time, software and an XMP container in all eight photographic originals, and in none of the browser screenshots of those same photos. That was a local capture, not a test of any social app.

2. Step 2

   Read text the generator may have saved

   Some Stable Diffusion front ends and ComfyUI save prompts or a workflow inside a PNG. Use the Stable Diffusion PNG info reader on the original PNG. A parameters or workflow field is a stored claim. It can be edited, deleted or copied onto another file. Its absence means the field is not in this file. The Stable Diffusion page uses the same local reader beside the general checker.

3. Step 3

   Look for Content Credentials separately from ordinary text

   Content Credentials are the C2PA way of attaching a signed history to a file. A validator has to check the signature, the binding to these pixels and the trust list. This site does not do that. Follow how to check Content Credentials and use a validator on the original file. A signed manifest can describe an AI generation, an edit or a capture. A missing manifest is common for both cameras and generators.

4. Step 4

   Search for an earlier copy

   Reverse image search asks where this picture, or something close to it, has already been published. A match to an older post can show that the file is recycled. No match leaves both a private photo and a brand-new generation possible. Google’s own help describes results as similar images and pages that contain the image, which is a publication search.

5. Step 5

   Ask for the history the sender should have

   A camera photo has a raw file, a camera roll or a burst. A generated image has a prompt, a seed, a project or an account history. Ask for that earlier artifact and compare it with the file under dispute. For a portrait used as a profile, the profile-picture guide keeps identity questions separate from the question of how the pixels were made. For a still that is being called a deepfake, read the scope of the deepfake page: this checker is for still images, and it does not authenticate a person.

6. Step 6

   Use a model score as one more reading

   Run the image check when you want a general estimate. On this site, a completed pixel-model score at or below 0.35 is displayed as likely camera-origin, and at or above 0.70 as likely AI-generated. The band between them is inconclusive. Those are display thresholds, not a calibrated probability and not a published accuracy. Two products can disagree for reasons covered in why AI detectors disagree.

7. Step 7

   Look at the picture last

   Hands, text, reflections, shadows and repeated texture still reward a slow look. Current generators miss those less often than early ones did, and ordinary photos contain blur, compression and odd crops. Treat a visual oddity as a prompt to return to the file, the source and the sender. It does not identify Midjourney, Flux, DALL·E or any other product by style alone. The Midjourney, Flux and DALL·E pages say the same thing in their own context: resemblance is not attribution.

Write down the claim you can actually make

A useful note is narrow. “This PNG contains a parameters field that names a checkpoint” is a statement about a file. “The pixels look like a generator to one model, and another model was inconclusive” is a statement about two tools. “The person in the picture did not attend the event” is a statement about the world, and neither metadata nor a score establishes it.

If the file passed through Instagram or a messenger, read what is known about metadata after sharing before you treat a blank EXIF panel as evidence of generation. Our own measurements cover a specified JPEG export and a browser screenshot. They are not an upload test of those apps.

SynthID and similar watermarks are a further, separate signal. Google DeepMind documents SynthID as a watermarking system. This site does not include a SynthID decoder, so a check here cannot report that a watermark is present or absent.