The review desk · Practical guides
How to check an image for AI
Start with the original file and its source. An AI-generation score is one part of that review, not a certificate of authenticity.
Keep three kinds of evidence separate
- Source history: who supplied the image, where it first appeared, and what they say about its creation or editing.
- File information: embedded camera or software fields, generation settings and any verifiable provenance. Ordinary metadata can be missing, edited or copied.
- Model assessment: a detector’s analysis of the submitted image. It can be wrong and does not establish who created it.
CouldItBe’s local inspector describes file and pixel observations. The optional Sightengine scan returns an AI-generation score. A high score means stronger model signals; it does not mean that the same fraction of pixels was generated. See Sightengine’s model documentation.
Try a source-first review
- Ask for the original export or camera file, not just a social-media screenshot. Keep a copy before making edits.
- Open the image tool and examine its local observations. Do not infer origin from a missing software field or a polished appearance.
- If provider screening is available, choose a JPEG, PNG or WebP up to 10 MB. Read the sharing notice and consent before submitting. The selected file, including its metadata, goes through CouldItBe to Sightengine.
- Compare the completed result with source history and any disclosed generation or editing. Save the detector report separately from the local report.
GIF is supported for local inspection only. A paused or failed scan is not a low AI score. Video frame screening uses the same shared image allowance and does not establish whether an entire video is a deepfake.
A worked review: the screenshot problem
Original illustrative scenario · No measured scores
A contributor sends a screenshot of a landscape. Its metadata contains no camera model and no generation settings. The contributor later supplies an original export with a software field.
Does missing camera data prove the screenshot is AI-generated?
No. The screenshot describes a new file, not a complete record of the original scene. Missing metadata is a lack of evidence. Ask for the source file and creation history; do not label the picture AI-generated merely because camera information is absent.
Does a software name prove how the whole image was made?
No. A software tag is a declaration that may reflect export or editing. It can be changed. Keep it separate from verified provenance and from the detector’s assessment.
This scenario demonstrates evidence handling. No images were scored to create it, and it makes no claim about detector accuracy.
Where Content Credentials fit
Content Credentials can provide signed, tamper-evident provenance information. They are different from an ordinary software tag or an AI detector score. Their absence is not evidence that an image is fake, and valid provenance does not automatically prove that a depicted event happened. See the C2PA FAQ and C2PA’s guidance on missing credentials.
CouldItBe’s local metadata inspection is not a full C2PA signature validator. Use a dedicated verifier for that task. Independently check the caption, date, location, publishing permission and relevant source records.
Keep the evidence in context
These guides support a review, not a finding of misconduct. Read our methodology and actual test results, known limitations, and privacy policy. Report a correction if a source or product detail changes.
Browse all resources →