Free AI image detector
AI Image Detector — Is This Image AI?
Read source clues in the file, then an optional model check when that is not enough. Free to use, with no account. A result is guidance, not proof.
Some checks send a file fragment and a smaller copy to our server and detection provider. Privacy by design
Drop an image here
choose a file · paste with ⌘V / Ctrl V
The checker is general. The file clues are not.
The upload above does not identify a product from the pixels. Each page below says what that generator documents about its own files, including the vendor’s check when one exists. For the question as a checklist, read is this image AI?
Midjourney
Original downloads may still have the ID tag checked at midjourney.com/verify. Screenshots usually do not. A v7 look is not a version.
GPT Image
Supported ChatGPT and API images can include Content Credentials and SynthID. OpenAI’s checker is openai.com/verify. We decode neither. Current ChatGPT images are GPT Image, not DALL·E.
Gemini and Imagen
Gemini’s image docs say generated images include a SynthID watermark. Some Nano Banana models also document C2PA. Legacy Imagen files are on the same page.
Flux
Hosted Black Forest Labs files can carry Content Credentials. A local PNG may carry a workflow. Flux 1.1 Pro is covered here; Flux.2 has its own page.
Stable Diffusion
Local PNGs may contain Automatic1111 parameters or a ComfyUI workflow. Those fields are editable. A typical install writes no C2PA manifest and no SynthID watermark.
Grok Imagine
xAI documents a visible watermark on website and app images. The API pages we used do not document a trusted C2PA manifest. A crop removes the easy clue.
Work with the image file you have
Read stored fields, choose a removal policy, and inspect a downloaded copy. These local file tools work without an account.
Photo sharing privacy
Choose local tools for photo GPS and sharing privacy. Follow three reproducible, self-authored examples from stored GPS to a checked JPEG copy.
Photo metadata tasks: dates, orientation, thumbnails and batch handoffs
Choose a local workflow for photo dates, timezones, display orientation, embedded thumbnails, batch metadata reports and verified cleaning copies.
PNG metadata tasks: read text, locate chunks and edit selected fields
Choose a local PNG workflow for stored text, chunk ranges and CRCs, or verified text editing. Follow three defined original, empty and Unicode-edited scenes.
ComfyUI workflow inspection: get the JSON, read nodes and list references
Choose a local path from a PNG or JSON file to stored nodes, connections, literal model references and scoped compatibility questions. Follow UI, API and unrelated-JSON scenes.
Image file checks: format, dimensions, hashes and defined pixel differences
Choose a local file check for container type, display dimensions, exact duplicate bytes or native-size sRGB RGBA differences, with three reproducible synthetic scenes.
Content Credentials tasks: distinguish storage, read claims and keep verification separate
Choose a local credential-reading workflow for manifest_read, absent and present_unreadable results, export a private-by-default report, or compare an explicit removal copy.
Choose a local tool
Metadata, saved PNG parameters and embedded Content Credentials.
Follow a file walkthrough
Known examples, actual outputs and the next step when a field is missing.
Start with the file you actually have.
JPEG, PNG or WebP, up to 10 MB. A social post link is not the image. Free to use subject to availability; no account required.
- 01
Keep the earliest copy
Drop the original download or paste a direct image URL. A screenshot or a social re-save is a new file, and the embedded clues are often already gone.
- 02
Read source clues, then the model if needed
Source text and camera metadata are read in your browser. These fields can be edited, and signatures are not verified. If a model check is still needed and available, header data and a smaller copy may be sent, with embedded metadata removed.
- 03
Read the label you were given
A completed score, an inconclusive middle band, and an unavailable check are different outcomes. Missing metadata does not prove the image is real. The step-by-step version of this question is on the Is this image AI page.
Three labels, plus a check that did not finish.
The bands below apply only to a completed pixel-model score. They are display thresholds, not a calibrated probability and not proof. A metadata-only reading describes editable source text. It is not one of these bands.
≤ 0.35
Likely real
A completed score in this band suggests a camera source. It can be wrong, and it does not name a camera or a person.
0.35 < score < 0.70
Inconclusive
The model cannot clearly separate the two classes. Inconclusive is not a soft yes, and it is not a photograph.
≥ 0.70
Likely generated
A completed score in this band suggests AI generation. It can be wrong, and it does not name the model.
What metadata can and cannot prove
Metadata is text beside the pixels: a camera make, a software name, a prompt, a workflow. It can support a story about this file, and it can be edited, copied onto another picture, or dropped by a screenshot. Seeing “Midjourney” or “Canon” is an observation, not a validation.
Content Credentials (C2PA) are a signed history. A real check confirms the signature, that it still matches these pixels, and that the certificate is trusted. This site does not do that, and the letters “c2pa” are not that check. Which generators document a manifest is on their own pages. SynthID is a pixel watermark on Gemini images and on supported OpenAI images. We have no decoder. A missing watermark, manifest, or camera tag means the clue is absent. Absence is not a camera proof.
What this check will not decide
This is not a named-model classifier, not an identity check, and not a video tool. We do not publish an overall accuracy figure. Scores move after edits, compression, or a generator the model has not seen. Request limits can stop a check before it finishes. Only upload images you are allowed to check.
A screenshot is the weakest file. Our eight-photo experiment measured a local JPEG export and a browser screenshot, not Instagram or WhatsApp. If the model check does not finish, the label is unavailable. Unavailable is not “likely real”.
Is this image AI? is the same question as a checklist. The metadata experiment states what those eight files actually lost.
Check an image without leaving the page
The Chrome Web Store lists Check AI Free. Right-click an image or an image link and the check opens in the side panel, or upload a file there. It does not scan pages on its own. Source fields can be edited, signatures are not verified, and the result can stay inconclusive or unavailable. The store page is the current version.
Clear image handling, at every step.
Source clues are checked in your browser first. When needed, header data and a smaller image copy may be sent to our backend and detection provider. Image URLs are retrieved by our server. See the privacy policy for processing and provider retention.
- Model copy
- Metadata removed
- Server image storage
- None
- Extension history
- Clear anytime
How does the check work?
We inspect source clues first. When needed and available, a smaller copy can be sent for a model check. Every result explains the evidence available.
How accurate is it?
We do not publish a validated overall accuracy claim. Metadata can be changed or removed; model scores can be wrong and vary with edits, generators and compression.
What happens to my image?
Source clues are checked in your browser first. When needed, header data and a smaller image copy may be sent to our backend and detection provider. Image URLs are retrieved by our server. See the privacy policy for processing and provider retention.
Does an AI avatar prove an account is fake?
No. Image generation and account identity are different questions. Use independent sources and account context.
Is it free?
No account is required. Checks are free subject to request limits and service availability.
What does an inconclusive result mean?
The model could not clearly call the file generated or camera-origin, or it did not produce a usable score. That is not evidence of a photograph and not evidence of AI. An unavailable check means the model did not finish. That is not a camera verdict either.
Can metadata or C2PA prove the image?
Ordinary metadata is editable text. Content Credentials are a signed history, and only a validator that checks the signature, the pixel binding and the certificate can rely on them. This site does not. A missing manifest is common for cameras and generators.
Will the result name the generator?
No. The optional score is a general AI-image estimate. Naming Midjourney, GPT Image, Gemini, Flux, Stable Diffusion or Grok still depends on an original export or that vendor’s checker.