How accurate is the detector?

It reports a probability, not a verdict. Scores near 50% are labelled Uncertain ; treat any single result as a signal and combine it with other evidence. Read more on how accurate AI image detectors are .

Does my image leave my computer?

On this public instance, yes: the picture is uploaded to the AiPicDetect server (a Google Cloud Run container run by the author), scored in memory and never written to disk. The scrubbed copy is held in memory only until 100 newer results replace it or the container restarts, and nothing is sent to a third-party API. If you want nothing to leave your machine, run AiPicDetect yourself with one Docker command. Details are on the privacy page .

Yes. The code, the Docker image, the CLI and the GitHub Action are all in the AiPicDetect repository under the MIT licence. The detector is an open Hugging Face model you can inspect or replace.

Can I remove C2PA and other metadata?

Yes. After checking a picture, use Download clean copy . AiPicDetect rebuilds the image from its pixels, so EXIF, XMP, IPTC, C2PA and the ICC profile are all left behind. How the scrubber works .

Which formats are supported?

JPEG, PNG, WebP, HEIC/HEIF and most formats Pillow can decode, up to 50 MB.

Why is metadata listed?

C2PA content credentials and editing-software tags are provenance hints. The panel shows which blocks (EXIF, XMP, IPTC, C2PA, ICC) the file carries so you can weigh them alongside the score. See C2PA content credentials and how to remove image metadata .

Can I use a different model?

Yes. Set AIPICDETECT_DETECTOR_MODEL to any Hugging Face image-classification model whose labels name AI/fake vs. human/real content.

Yes. AiPicDetect is open source under the MIT licence. The hosted instance is free to use with a limit of 10 analyses per IP address every 24 hours; a self-hosted copy has no limit.

Does it work on screenshots or heavily compressed images?

It runs, but re-encoding, resizing and screenshots remove some of the pixel-level traces the classifier relies on, so expect lower confidence and more Uncertain results.

Can it tell which generator made an image (Midjourney, DALL·E, Stable Diffusion)?

No. AiPicDetect scores generated-versus-real pixel statistics in general; it does not identify the generator, and it has no knowledge of generators released after its model's training data was collected.

Why did a real photo score as AI?

Heavy filters, HDR processing, upscaling, illustrations and 3D renders share statistical features with generated images. The score is a probability, not proof; false positives happen.

Can I run it offline?

Yes. After the first run downloads the model into the Hugging Face cache, a self-hosted AiPicDetect needs no network access.

Is there a rate limit on the hosted instance?

Yes: 10 analyses per client IP in any rolling 24-hour window. Responses carry X-RateLimit-Remaining , and a request over the limit returns 429 with a Retry-After header.

Does taking a screenshot remove image metadata?

Taking a screenshot of a photo does strip most metadata — but not completely, and not in the way most people assume. Here is exactly what a screenshot removes, what it preserves, and what it adds.

What a screenshot removes

When you screenshot an image displayed on screen, the screenshot tool captures raw pixels from the display buffer. It does not copy the source file. As a result, the screenshot contains no EXIF, XMP, IPTC, or C2PA data from the original image. GPS coordinates, camera model, shutter speed, copyright fields, AI-generation provenance — all gone.

This is why screenshots are frequently used to "launder" AI-generated images: a Midjourney image screenshotted on an iPhone loses the XMP prompt data and any detector fingerprint. The resulting PNG looks like a photo taken on that phone.

What a screenshot adds

The screenshot file itself is not metadata-free. Depending on the platform:

iOS (iPhone/iPad): adds EXIF Make: Apple, Model: iPhone XX, the current date and time, and (if Location Services are on for screenshots) GPS coordinates of where the screenshot was taken.
macOS: adds Make: Apple, Model: MacBook… (or equivalent), creation timestamp, and the screen's ICC colour profile.
Android: adds device make and model, timestamp, and sometimes GPS if the screenshots app has location access.
Windows: minimal metadata, typically just a creation timestamp in the file system.

Screenshots and AI detection

JPEG compression and colour-space conversion during screenshot capture change the pixel-level statistics that AI detectors rely on. Uploading a screenshot of an AI image typically lowers the AiPicDetect confidence score compared to uploading the original file. This does not mean the image is real — it means the detector has less signal to work with.

For accurate AI detection, always prefer the original file. If only a screenshot is available:

Check the screenshot's own metadata — the device and timestamp may still be informative.
Use visual inspection alongside the detector score.
Run a reverse image search — the original source may still be indexable.

Does removing metadata from the original prevent detection?

No. Stripping metadata removes file-level tags; it does not change the pixels. A metadata-scrubbed AI image still scores AI on a pixel-level detector. Only re-rendering (such as passing the image through a diffusion inpainting model) would change the pixel signature, and even that is not guaranteed to fool a modern detector.

Check a screenshot for AI origin →

Last updated 2026-09-13 · source on GitHub

Try the detector