Use the original image when possible
Original dimensions and lightly compressed files preserve color, luminance, adjacent-pixel, and file information better than social-platform screenshots.
Read the human verification guideAI image-generation signal screening Text · Image · Video
How can you tell whether an image is AI-generated?
Choose an original image and confirm the upload notice to check generator metadata, C2PA and model signals. Screenshots and compression can remove evidence; uncertain results remain inconclusive.
- Three content types
- No login or sign-up
- Evidence you can inspect
Content screening and verification
Choose the content to inspect
Results separate heuristic screening signals, file facts, and evidence limitations.
Text content
Paste the text you want to inspect
GenDetect checks repetition, sentence length, punctuation, and character structure.
Your text is not saved or uploaded
Image content
Choose a local image
Automatically checks generator metadata, provenance credentials and model signals, with consent before server analysis.
本地启发式不会上传图片。
Server analysis requires confirmation of the upload and processing notice
Video content
Choose a local video
Server analysis supports videos up to 120 seconds and 100MB, inspecting at most six sampled frames after upload consent. Audio and temporal continuity are not checked.
Analysis time depends on file size
Image detection guide
How to tell whether an image is AI-generated
Check the original file for generation claims and content credentials, then review pixel signals and provenance. Only an explicit AI claim that is verified, trusted and bound to the file supports an origin conclusion.
Review local and global features
Limited color, excessive smoothness, and unusual noise may come from generation, editing, or compression, so visual signals should be considered together.
Read methodology and limitationsCheck metadata and provenance
Metadata can be forged or removed. A C2PA signature chain also needs a trusted signer and a relevant claim. Missing credentials do not prove a photo is real; check the author and reverse-image search results.
Review accuracy and false positivesDetection methods
What signals does AI content detection examine?
Each content type needs a different analysis. GenDetect separates screening signals, file facts, and limitations.
AI text detection
Examines vocabulary repetition, recurring phrases, sentence-length distribution, punctuation frequency, and unusual characters. More complete text usually produces steadier statistical signals.
Open text detectorAI image detection
Checks original-file metadata, provenance credentials and model signals. Unverified claims and uncalibrated scores remain supporting clues.
Open image detectorVideo-frame and metadata screening
Server analysis checks at most six frames of videos up to 120 seconds. It does not inspect audio or temporal continuity, and a negative result does not establish human origin.
Open video detectorHow to read the score
The screening index is a rule anomaly score, not a source probability.
This demo locally aggregates heuristics for text, images, and sampled video frames; images also check known generator metadata. The index is not trained or probability-calibrated and cannot prove origin.
Frequently asked questions
AI image detector FAQ
Learn what screenshots and compression remove, which visual signals are checked, and why a specific model such as Midjourney cannot be attributed reliably.
Does a screenshot, crop, or compression affect AI image detection?
Yes. Screenshots, crops, filters, and platform compression change pixel color, luminance, and adjacent edges and may remove metadata. Prefer the original image whenever it is available.
Which visual signals does an AI image detector examine?
This demo calculates sampled color diversity, average brightness, luminance variance, and adjacent-pixel edge density. Camera processing, editing, or compression can produce similar visual signals.
Can this detector tell whether Midjourney or DALL-E made an image?
A file may retain an OpenAI, DALL-E, Google AI, Stable Diffusion, Midjourney or similar generator marker. The tool attempts to validate the C2PA signature chain, but ordinary metadata can be forged or removed. Only a trusted, file-bound claim supports an origin conclusion; it cannot reliably infer an undeclared specific model.