Detect AI Images Online: 4 Steps to a Result
If all you want to know is "is this image AI-generated?", the fastest route is to upload it rather than work through a methodology. This page explains how the detection actually runs and, more importantly, how to read the result — including which conclusions are confirmed and which are inference.
The whole process is four steps. The tool runs two independent analyses simultaneously and shows you both side by side — which is the difference from a detector that hands you one score and calls it done.
Step 1: get the right file
PNG, JPG/JPEG, WEBP, BMP and TIF/TIFF are supported, up to 30MB per image. Use the unprocessed original where you can — that matters more than which tool you pick.
Why the original file matters
Screenshots, social platform transcoding and re-saving as JPEG all destroy the statistical features pixel-level detection relies on, and strip container metadata. The original preserves both lines of evidence. With only a compressed copy, any result is inference.
Step 2: upload
No signup, no account — just drop the file in. There are 3 free detections per day, and the quota is not tied to any profile.
Technically, the image passes through a third-party image host (the detection model runs on the server, so this hop cannot be skipped), and the host's copy auto-expires within 5 minutes. Our server does not write the image to disk; any temporary file is removed immediately after analysis, even if the analysis fails.
Step 3: two independent analyses run in parallel
Signal A: metadata
The tool parses embedded PNG and JPEG metadata and checks three things:
- Generator signatures — software identifiers and parameter strings left by Midjourney, Stable Diffusion, NovelAI, DALL·E, ComfyUI, Flux.1 and others.
- AIGC markers — implicit labels under China's GB 45438-2025 standard (the TC260 namespace), carrying generation status, service provider, distribution platform, unique ID and signature.
- Camera EXIF — make, model, lens, aperture, shutter, ISO, capture time, GPS. This one supplies counter-evidence.
Signal B: pixel analysis
A deep learning model analyses the image pixels directly and returns a 0-100 AI-probability score. This route does not depend on metadata surviving — which is exactly why it is the most affected by re-processing.
Step 4: read the result properly
This is the step most people skip. Read it in this order:
| Metadata finding | Pixel score | How to read it |
|---|---|---|
| Generator signature or AIGC marker found | Any | Confirmed: AI-generated, the score is irrelevant |
| Complete camera EXIF present | High | Strong evidence of a real photo; a high score may come from heavy retouching |
| No generator info, no camera info | High | Inference: could be AI, could be a heavily processed real photo |
| No generator info, no camera info | Low | Inference: modern high-quality generations can genuinely fool pixel analysis |
| Camera EXIF present | Low | The most consistent outcome: a real photograph |
Why a disagreement is not a bug
If a photo carries a complete phone EXIF but scores high on pixels, the usual cause is heavy post-processing — skin smoothing, noise reduction, sharpening — pushing a real photo's statistics toward the generated distribution. The camera record is more trustworthy than the score here. The tool surfaces the conflict rather than silently picking a winner, because the conflict is itself informative.
About the daily free quota
Three free detections a day. That limit is not a marketing device — it is a compute reality. The model runs on modest hardware and the service is funded out of pocket. If you need it at team or batch scale, get in touch through the GitHub repository linked in the footer.
What gets stored
Not the image. The result is kept — the conclusion, not the picture: filename, dimensions, file size, MIME type, AI score, verdict text, and any metadata findings. That is what lets you review your history with your key and delete it at any time.
The privacy policy covers the details, including what is collected automatically and how long it is retained. The short version: the picture is gone within minutes; the verdict about it stays.
When detection is the wrong tool
Some questions are not "is this image AI-generated?" — and no detector will help with them.
- The image is real but the caption is false — extremely common. Tracing the source (reverse image search plus finding the original context) is far more effective than detection.
- Local retouching — if only a small region was AI-repainted, a whole-image detector may see nothing at all.
- Video — this tool handles still images; it does no per-frame video analysis.
- You need a legal-grade conclusion — a detection result is technical evidence, not an expert determination.
That last point deserves emphasis. Any detector output should be treated as evidence, not a verdict — this one included. It is the design principle, not a disclaimer.
Frequently asked questions
Is this AI image detector free?
Yes — 3 free detections per day, no account required. The quota is tied to no profile, and the site sets no cookies and runs no third-party analytics or advertising scripts.
Do I need an account?
No. There is no account, no email address and no password. The service sets no cookies and loads no third-party tracking scripts.
Is my uploaded image stored?
The original is not stored. The image passes through an image host and auto-expires within 5 minutes; our server never writes it to disk and removes any temporary file immediately after analysis. What is kept is the result — score, verdict and metadata findings — with no image bytes, and it can be deleted at any time.
Which image formats are supported?
PNG, JPG/JPEG, WEBP, BMP and TIF/TIFF, up to 30MB per image.
How accurate is the detection?
Treat the result as evidence, not a verdict. When a generator signature or AIGC marker is found, the conclusion is certain. A score on its own is inference, and re-compression degrades it. This is a shared limitation of every AI image detector, not something specific to this tool.
Can you detect AI video or audio?
No. This tool handles still images — no per-frame video analysis and no audio detection. Provenance rules for audio and video differ and require dedicated tooling.